Information science

A patent foramen ovale (PFO) is an interatrial shunt, with a prevalence of 20–34% in the general population. While most people do not have secondary manifestations of a PFO, some reported sequelae include ischaemic stroke, migraine, platypnoea–orthodeoxia syndrome and decompression illness. Furthermore, in some cases, PFO closure should be considered for patients before neurosurgery and for patients with concomitant carcinoid syndrome. Recent trials support PFO closure for ischaemic stroke patients with high risk PFOs and absence of other identified stroke mechanisms. While PFOs can be associated with migraine with auras, with some patients reporting symptomatic improvement after closure, the evidence from randomised controlled trials is less clear in supporting the use of PFO closure for migraine treatment. PFO closure for other indications such as platypnoea–orthodeoxia syndrome, decompression illness and paradoxical embolism are based largely on case series with good clinical outcomes. PFO closure can be performed as a day surgical intervention with high procedural success and low risk of complications.

Kenneth K Cho · Shaun Khanna · Phillip Lo · Daniel Cheng · David Roy

AI-based tools can help with image acquisition, reconstruction and quality; interpretation, diagnosis and decision support; and assisting with manual tasks

Meng Law · Jarrel Seah · George Shih

International use of acute kidney injury care bundles, including e-alerts, represents a potential pathway for significant improvement in acute kidney injury management in Australia

Anna C Bendall · Sven‐Jean Tan · Emily J See · Nigel D Toussaint

Objectives: To determine whether a multifaceted primary health care intervention better controlled cardiovascular disease (CVD) risk factors in patients with high risk of CVD than usual care. Design, setting: Parallel arm, cluster randomised trial in 71 Australian general practices, 5 December 2016 – 13 September 2019. Participants: General practices that predominantly used an electronic medical record system compatible with the HealthTracker electronic decision support tool, and willing to implement all components of the INTEGRATE intervention. Intervention: Electronic point‐of‐care decision support for general practices; combination cardiovascular medications (polypills); and a pharmacy‐based medication adherence program. Main outcome measures: Proportion of patients with high CVD risk not on an optimal preventive medication regimen at baseline who had achieved both blood pressure and low‐density lipoprotein (LDL) cholesterol goals at study end. Results: After a median 15 months’ follow‐up, primary outcome data were available for 4477 of 7165 patients in the primary outcome cohort (62%). The proportion of patients who achieved both treatment targets was similar in the intervention (423 of 2156; 19.6%) and control groups (466 of 2321; 20.1%; relative risk, 1.06; 95% CI, 0.85–1.32). Further, no statistically significant differences were found for a number of secondary outcomes, including risk factor screening, preventive medication prescribing, and risk factor levels. Use of intervention components was low; it was highest for HealthTracker, used at least once for 347 of 3236 undertreated patients with high CVD risk (10.7%). Conclusions: Despite evidence for the efficacy of its individual components, the INTEGRATE intervention was not broadly implemented and did not improve CVD risk management in participating Australian general practices. Trial registration: Australian New Zealand Clinical Trials Registry, ACTRN12616000233426 (prospective).

Ruth Webster · Tim Usherwood · Rohina Joshi · Bandana Saini · Carol Armour · Sue Critchley · Gian Luca Di Tanna · Shane Galgey · Charlotte M Hespe · Stephen Jan · Ajay Karia · Baldeep Kaur · Ines Krass · Tracey‐Lea Laba · Qiang Li · Serigne Lo · David P Peiris · Christopher Reid · Anthony Rodgers · Louise Shiel · Jessica Strathdee · Nuria Zamora · Anushka Patel

To the Editor: The coronavirus disease 2019 (COVID‐19) pandemic has created an unprecedented need for real‐time surveillance data to inform decisions and action by public health responders and primary health care practitioners. Early in the pandemic, many countries swiftly produced interactive national dashboards with mapping capabilities.1,2 A dashboard is an online tool for data management which optimises information access and data visualisation.3 Dashboards provide benefits compared with standard reporting, including sharing near real‐time data during rapidly evolving situations, and providing users with the opportunity to interact with the data. If designed appropriately, users can also interrogate data and ask questions based on their specific informational needs. Many dashboards also provide mapping capabilities, allowing for visualisation of spatial distribution of information, and monitoring trends geographically over time.2 Australia does not yet have an official and publicly accessible national interactive dashboard for COVID‐19. Some states and territories have developed publicly available COVID‐19 dashboards, but data are generally aggregated, making it difficult to answer specific questions that include time and location and source of infection. An interactive near real‐time dashboard could improve access to and comprehension of data for primary health care providers and public health responders. Researchers from the Australian National University, Menzies School of Health Research and the University of Queensland are developing a COVID‐19 Real‐time Information System for Preparedness and Epidemic Response (CRISPER) (https://crisper-graphc.hub.arcgis.com/) as a nationwide information and visualisation system for Australia. CRISPER aims to become the principal source of accurate, reliable and spatially explicit real‐time information for COVID‐19 (Box). The system currently uses publicly available postcode‐level data, primarily from state and territory health department websites. Gaining access to nationwide line‐listed data is underway, which will allow additional functionality, including a clinical dashboard detailing clinical outcomes (eg, hospital and intensive care unit admissions, deaths) stratified by demographics, comorbidities, time and place. Also under development is an automatic alert system providing registered users with daily or weekly email alerts on new cases, contract tracing alerts and/or testing rates based on user‐defined geographical areas of interest. We believe that CRISPER will improve accessibility of information for primary health care practitioners and public health responders and will enable them to make more timely and informed decisions. This system may serve as a prototype platform for rapid information sharing for other epidemic‐prone diseases. Box – Features of the Coronavirus Disease 2019 (COVID‐19) Real‐time Information System for Preparedness and Epidemic Response (CRISPER) CRISPER aims to optimise information access and visualisation for COVID‐19 through: a national summaries dashboard detailing cases, deaths and testing — information can be filtered or summarised by states and territories, time periods, and 7‐ or 14‐day rolling averages (https://graphc.maps.arcgis.com/apps/opsdashboard/index.html#/465d9e0cd44247b488b8431a56691417); and an interactive mapping tool of cases, testing and contact tracing alerts by location (postcode, local government areas, public health units) — information can be filtered by time periods and source of infection (currently available for New South Wales). A key feature distinguishing this tool from other dashboards is that the data in the different components are linked; for example, the epidemic curve is dynamic based on cases in the map window (https://graphc.maps.arcgis.com/apps/opsdashboard/index.html#/74e69c2ab40f41c892a652e95373622c)

Emma Field · Amalie Dyda · Colleen L Lau

In June 2015, 41% of Australian medical specialists were women,1 but only 28% of those in senior or leadership positions.2 Academic research is important for obtaining tenure and promotion in medicine. First authorship on publications is typically granted to junior authors and last authorship to directing senior authors. The proportion of women among first authors in six prominent American medical journals increased from 5.9% in 1970 to 29.3% in 2004, and for last authorship from 3.7% to 19.3%.3 However, a 2016 study found that the proportion of authors who were women in high impact medical journals had plateaued or declined since 2009.4 Examining Australian patterns of authorship could help identify barriers to the academic advancement of women in medicine. We identified in PubMed all journal articles published during 2005–2018 by the eight journals associated with peak bodies of Australian medical practitioners, and used the validated genderize. R tool to determine the probable gender of authors’ first names.5 We used Poisson regression to analyse first and last authorship (male = 0, female = 1) by year; we report the statistical significance of the deviation of the regression slope (B‐value) from zero. The relationship between number of authors and gender were assessed by linear regression, including an interaction term between gender and time. Formal ethics approval was not required for this analysis of publicly available data. Gender could be determined with at least 50% probability for the first authors of 26 621 of 27 804 articles (96%) and the last authors of 26 972 (97%). Between 2005 and 2018, the proportion of women among first authors in the eight journals increased from 522 of 1600 (32.6%) to 899 of 2391 (37.6%; P < 0.001); the proportion among last authors did not change (28.0%). The proportions of women among both first and last authors increased significantly in the Journal of Paediatrics and Child Health, the Australian and New Zealand Journal of Obstetrics and Gynaecology, and the Medical Journal of Australia, as did those of first authors (but not last authors) in the Australian and New Zealand Journal of Public Health, Emergency Medicine Australasia, the Australian and New Zealand Journal of Psychiatry, and the Australian and New Zealand Journal of Surgery; the proportions of neither changed significantly in Australian Family Physician (Box; Supporting Information, table 1). The mean number of authors on publications with women as first authors (3.8; standard deviation [SD], 2.4) was higher than for those with men as first authors (3.3; SD, 2.4; P < 0.001). The difference between author numbers was smaller, but statistically significant, with respect to last author gender (women: mean number of authors, 3.6; SD, 2.4; men: 3.5; SD, 2.4; P = 0.045) (Supporting Information, tables 2, 3). Our study did not distinguish between research, review, and other journal article types. While our findings may reflect overall involvement of women in research, they do not specifically define gender proportions among leaders of high impact academic research programs. The increase in the proportion of first authors of Australian medical journal articles who are women may reflect the rise in the proportion of female doctors from 33% to 43% between January 2006 and December 2018.1 It is also possible that women, under‐represented in their specialties, feel greater pressure than men to publish as first authors for purposes of career progression.2 Our data indicate that the proportion of women as first authors has increased, but that of last authorship has grown only in some specialities. Box – Proportions of women as first and last authors of articles in selected Australian medical journals, 2005–2018* * The raw data are included in the online Supporting Information, tables 4 and 5. † From 2018: the Australian Journal of General Practice.

Matthew J Lennon · Rose Kennedy · Hannah Ryan · Dennis R Neuen · Melissa Godwin

CHRIS supported the Victorian ICU response during the COVID‐19 pandemic The coronavirus disease 2019 (COVID‐19) pandemic put an unprecedented strain on intensive care resources throughout the world. Initially in Wuhan (China)1 and then in Lombardy (Italy),2 London (United Kingdom) and New York (United States),3 demand exceeded capacity, with 10–15% of the patients admitted to hospital developing critical illness. Australia has 191 adult and paediatric intensive care units (ICUs), with over 2300 ICU beds.4 This is equivalent to 8.9 ICU beds per 100 000 population, more than the UK but fewer than Italy and the US.5,6 In late March 2020, rising numbers of COVID‐19‐related admissions to ICUs were observed throughout Australia.7 The Australian and New Zealand Intensive Care Society (ANZICS) and the Australian Government Department of Health recognised that ICU demand was unlikely to be uniform, that capacity might be exceeded in one region but not in another, and that matching ICU resources to areas of greatest need might be required. A single sentence encapsulated the approach: “Why would we let a patient die in Western Australia if we can see a spare ventilator in Sydney?” A nationwide system to monitor ICU demand and capacity in Australia A nationwide dashboard of ICU activity, the Critical Health Resources Information System (CHRIS), was rapidly developed as a collaboration between Telstra Purple, Ambulance Victoria, ANZICS and the Australian Government Department of Health. All adult and paediatric ICUs (public and private) in Australia were instructed to enter data twice daily. This manual data entry typically took 5 minutes. Each ICU was immediately able to see patient numbers and resources available within every ICU in their region and also see an aggregate summary of all ICUs in Australia. CHRIS was available to all state and territory health departments, to all patient transport and retrieval agencies, and also to ICUs in New Zealand. The system went live on 1 May 2020, after 26 days of development. Three weeks later, 184 out of 188 eligible ICUs (98%) in Australia were contributing data. The ICU response to the second wave of COVID‐19 in Victoria After a decline in severe acute respiratory syndrome coronavirus 2 (SARS‐CoV‐2) infections throughout Australia, notifications rose again in Melbourne at the end of June 2020.8 In response, ICU directors from the lead hospitals of the nine designated Victorian health care clusters commenced a daily morning meeting with representatives from Ambulance Victoria, Safer Care Victoria and the Victorian Department of Health and Human Services. The group committed to maintaining standards of care expected under normal (non‐pandemic) conditions and to achieving this by proactively transferring patients (with or without COVID‐19) to another ICU if delivery of care was compromised by high local demand. Decisions to transfer patients were informed by data from CHRIS. Pre‐existing critical care transfer systems run by Ambulance Victoria were used. From the beginning of July to the end of September 2020, there were 237 ICU admissions with COVID‐19 pneumonitis, of which 210 (88%) occurred in July and August. Admissions were predominantly to public hospitals in north‐western Melbourne.9 The rapid and localised nature of presentations meant that it was faster to transfer patients to ICUs with vacant capacity than to open and staff additional beds, despite physical ICU bed spaces being available. Transfers from the emergency department or ICU at the four north‐western metropolitan hospitals alone accounted for 35% (46/133) of all critical care transfers in Victoria during July and August. Spare ventilators were available at all sites on all days. On six occasions in August, there were more than 140 ventilated patients (with or without COVID‐19) in Victoria. On each of these days, there were more than 500 spare ICU ventilators available (Box 1 and Supporting Information, graphic 1 in the video). Despite individual hospitals indicating transient increases in ICU bed numbers, there was no overall increase in open staffed ICU beds. As COVID‐19 cases rose, so too did numbers of critical care staff unavailable due to COVID‐19 exposure or illness, with 15 consecutive days when there were more than 60 staff unavailable (Box 2). Lessons learned CHRIS provided real‐time data on ICU activity and capacity. In addition to facilitating the transfer of critically ill patients, CHRIS also enabled early diversion of ambulance presentations to emergency departments at hospitals where ICUs had capacity. These approaches were integral to ensuring standards of care were maintained by clinicians, retrieval agencies and the Victorian health department. At the same time, there was visibility to the Australian Government Department of Health, which would, if required, coordinate a national response to overwhelmed ICU services. Although several individual ICUs came under strain, retrieval and critical care systems in metropolitan Melbourne were not overwhelmed. Strategies to redistribute critical care demand are likely to have contributed to high survival rates for ventilated patients with COVID‐19 in Victoria.9 Timely transfers to ICUs with open available beds could be facilitated. Availability of staff was more important in determining capacity to deliver care than availability of ventilators. The role for CHRIS in the future The local application of a national tool (CHRIS) for real‐time display of ICU activity and resources was a key component of the response to the COVID‐19 pandemic in Victoria. CHRIS has the potential to augment existing ICU monitoring systems. The tool may also assist in the response to local and national public health emergencies, such as mass casualty events, bushfires10 or thunderstorm asthma.11 Automated linkage of CHRIS to existing state‐based and national systems should be investigated. In addition, it may have potential use in monitoring health policy impacts more broadly. Box 1 – Snapshot of the Critical Health Resources Information System (CHRIS) summary page for Victoria during August 2020 ACT = Australian Capital Territory; COVID‐19 = coronavirus disease 2019; ECMO = extracorporeal membrane oxygenation; HDU = high dependency unit; ICU = intensive care unit; NSW = New South Wales; NT = Northern Territory; NZ = New Zealand; QLD = Queensland; SA = South Australia; TAS = Tasmania; VIC = Victoria; WA = Western Australia. Box 2 – Number of ventilated (dark blue) and non‐ventilated (light blue) patients in Victorian intensive care units and the number of critical care staff unavailable to work due to coronavirus disease 2019 (COVID‐19) exposure or illness (green dots), listed each morning in the Critical Health Resources Information System (CHRIS) LOWESS = locally weighted scatterplot smoothing.

David Pilcher · Nicholas R Coatsworth · Melissa Rosenow · Jason McClure

Technology is changing the way that people with type 1 diabetes are monitoring and managing their blood glucose levels Diabetes technologies have transformed management options in type 1 diabetes. The most notable innovations include the commercialisation of insulin pumps, advancements in glucose monitoring and the capacity for these technologies to interact. New technologies offer enhanced flexibility in insulin delivery and opportunities to improve glucose levels and enhance quality of life. Recognising these benefits, the uptake of advanced technologies in Australians with type 1 diabetes has increased. In 2018–2019, 41% of children and 26% of adults attending hospital diabetes clinics managed their type 1 diabetes with insulin pumps, and 55% of children and 13% of adults newly commenced continuous glucose monitoring (CGM).1 In this article, we provide a guide to current diabetes technologies available in Australia, describe their benefits and discuss important factors in assessing an individual’s suitability. Continuous glucose monitoring Accurate and accessible glucose monitoring is key to effective diabetes management. Finger‐prick testing of capillary blood for self‐monitoring of blood glucose (SMBG) became commercially available in the 1980s, and CGM since 1999. CGM is distinguished from SMBG by the measurement of glucose concentration within interstitial fluid by a small subcutaneous glucose‐sensing electrode. Data are transmitted to a receiving device (eg, insulin pump or smartphone) and converted into a continuous graphic display. Interstitial glucose concentrations correlate with plasma glucose, albeit with an average 7–8 minute time lag for equilibration of glucose between blood and the interstitial compartment. This delay is exaggerated at times of rapidly changing glucose. CGM systems come in different forms, with the main distinguishing features being the capacity to alert users and/or carers to high and/or low blood glucose levels set to individual preference. Various systems are available (Box 1).2,3 Product selection depends on the desirability of alarm functions, ease of sensor application, need for finger‐prick calibration, cost, and connectivity of the device to existing technologies (insulin pumps, Apple v Android systems). Modern CGM systems are reliable. Their performance is assessed by the mean absolute relative difference, an accuracy metric expressed as the percentage difference relative to a reference glucose concentration. CGM devices with a mean absolute relative difference < 10% are appropriate for treatment decisions.2,4 To optimise sensor performance, SMBG testing is still required to calibrate a number of real‐time CGM devices, yet devices are increasingly reliable such that newer factory calibrated devices no longer require user calibrations5 — a welcome feature for convenience and reduced finger‐prick burden. CGM offers several benefits to users, including on‐demand glucose testing, trend arrows, alarm functions and easy detection of out‐of‐range glucose levels. For clinicians, CGM offers additional data beyond glycated haemoglobin (HbA1c) measurements. HbA1c measurements are point estimates of haemoglobin glycation over 3 months, whereas CGM displays day‐to‐day glucose variability that often challenges people with type 1 diabetes. Consensus statements have attempted to harmonise the various commercially available CGM systems into a common reporting method to aid comparison between devices and also provide clinicians and users with more clinically meaningful data and targets (Box 2).2,6,7 Clinical trials provide evidence that CGM use may improve glycaemia in type 1 diabetes. A recent meta‐analysis of 15 randomised controlled trials comparing CGM with standard care (usually SMBG) in people with either type 1 or type 2 diabetes over 12–36 weeks found that CGM was associated with a slight reduction in HbA1c levels (weighted mean difference, − 0.17%), and increased time in range (TIR; 71 minutes/day).8 Added up over a year the benefit equates to an additional 18 days of TIR. The effect on TIR was independent of diabetes type, or method of insulin delivery (insulin pump v needle injections). Overall, studies favour CGM to improve glucose variability (optimal coefficient of variation in diabetes management, 34%; coefficient of variation reduced by 3.0–6.7%) and reduce hypoglycaemia (0.4–1.2 hours reduction in time spent with glucose levels < 3.9 mmol/L) compared with SMBG.2 The benefits of CGM and flash glucose monitoring for type 1 diabetes management have been recognised by the Australian Government, which first pledged $54 million in 2016 to fully subsidise CGM in people under 21 years of age. In 2019, a further $100 million in subsidies was added. Eligibility criteria were again expanded in March 2020 (Box 3).9 Insulin pumps An insulin pump delivers short‐acting insulin continuously via a cannula self‐inserted into subcutaneous tissue. In the 1970s, the first insulin pumps were large and bulky and delivered a single basal rate of insulin. Modern pumps are more discrete, the size of a pager. An insulin pump set‐up includes two major components (Box 4): Insulin pump — case with display, battery and an insulin reservoir connected to a plunger that controls the passage of insulin into the line tubing. The insulin pump is programmed to deliver continuous quick‐acting insulin in equal aliquots (0.01–0.025 mL) across an hour depending on the pre‐set rate to replicate basal insulin. Programmed rates can be customised to vary across a 24‐hour period, distinguishing delivery from long‐acting insulin delivered at an inflexible basal rate. The user must initiate bolus doses for meals or for correction of an elevated glucose reading, but pre‐programmed settings provide dose calculations (insulin‐to-carbohydrate ratio for meals, and insulin sensitivity factor for correction doses). Line tubing and infusion set — up to 60 cm of thin plastic tubing connects the insulin reservoir to a subcutaneous teflon cannula (tubeless insulin patch pumps with variable hourly rates are not currently available in Australia). Potential benefits of insulin pumps include: ► Flexibility in dosing — useful for extreme insulin sensitivity, erratic schedules, more convenient and frequent bolusing, to accommodate exercise, or to manage the dawn phenomenon (increased insulin requirements in the early morning period due to counter‐regulatory hormone secretion). ► Bolus calculation capacity. ► Less frequent insertion events (about every 3 days) — favourable for individuals with needle phobia. ► Insulin delivery and glucose data can be generated electronically and remotely for review. As only quick‐acting insulin is used in insulin pumps, insulin deficiency (leading to possible diabetic ketoacidosis) may occur within 2–3 hours of discontinuation of the insulin pump, or in the event of set occlusion. Set occlusion is one of the leading causes of ketoacidosis in insulin pump‐treated individuals but is rapidly corrected with recommencement of insulin in the absence of intercurrent infection (Box 5). However, insulin pump use has not resulted in the increased diabetic ketoacidosis events anticipated when first introduced, aided by appropriate education.10 There are out‐of‐pocket costs, especially for individuals without private health insurance, and running costs are higher than with insulin injection. Wearing an externally attached device to the body 24 hours a day is a deterrent to some, but a convenience for others who dislike carrying needle tips and insulin pens. Insulin pumps can be used either as a stand‐alone device or in conjunction with CGM sensors (Box 6). Sensor‐augmented insulin pumps have the added benefits of suspended insulin delivery for predicted low glucose (predictive low glucose suspend), or at the threshold of hypoglycaemia (low glucose suspend) to reduce the frequency and duration of hypoglycaemia. In a study of individuals with documented nocturnal hypoglycaemia, those randomised to insulin pumps with low glucose suspend function for 3 months had 32% less frequent hypoglycaemia than without suspend function.11 Other trials have also demonstrated reduced time in hypoglycaemia without increase in time in hyperglycaemia.12 The latest insulin pump systems (hybrid closed loop; HCL) can provide a further degree of automation of insulin delivery. HCL pumps provide real‐time adjustment of insulin delivery in response to ambient glucose levels detected by a CGM sensor, via an inbuilt control algorithm. The user is still required to manually deliver boluses for meals or adjust insulin for exercise. A recent study comparing HCL to sensor‐augmented insulin pump therapy reported improved TIR during daytime hours as well as overnight, and a small reduction in time in hypoglycaemia over 6 months.13 There is currently only one registered HCL insulin pump in Australia. Future technologies may provide further integration of CGM and insulin pump devices via phone‐based applications. Tailoring treatments to individual needs The optimal approach for the management of type 1 diabetes depends on individual and practical considerations. Initiation of insulin pump therapy requires extended consultation to discuss device selection and cannula insertion technique, and review carbohydrate counting and troubleshooting (including diabetic ketoacidosis risk mitigation). It also requires a multidisciplinary approach involving an endocrinologist, credentialled diabetes educator and dietitian.14 In concert with the individual with diabetes, factors to discuss include: the need for alerts and alarms: presence of hypoglycaemia unawareness and susceptibility to alarm fatigue; affordability and eligibility for CGM supplied under the National Diabetes Services Scheme (Box 3); access to training and education; customisation of glucose targets for pregnancy, age and comorbidities; ability to use software to upload data and share reports with health professionals; and allergies to cannula or CGM site adhesives. Conclusion Diabetes technologies are being increasingly adopted by people with type 1 diabetes, and clinicians should familiarise themselves with the spectrum of devices. These advancements offer potential benefits for people with diabetes, although prescribing these devices requires evaluation of cost and benefit for the individual. Human factors are the main determinant of success and satisfaction, highlighting the importance of consideration of the needs of the individual. Box 1 – Types of continuous glucose monitoring (CGM) systems2,3 Professional (retrospective): professional CGMs were the first CGM systems approved by the United States Food and Drug Administration in 1999. They provide blinded glucose data for review by a health care provider. The iPro (Medtronic) and Freestyle Libre Pro (Abbott) are currently available systems in Australia. Real‐time CGM: patient‐inserted systems include Guardian (Medtronic), Guardian Connect (Medtronic) and G6 (Dexcom). The Eversense (Senseonics) CGM implantable system is inserted subcutaneously by a physician and worn for 90–180 days with a transmitter adherent to the overlying skin with alert capacity (not currently available in Australia). Intermittently viewed CGM or flash glucose monitoring: Freestyle Libre for continuous glucose measurements shown retrospectively at the time of physical scanning of the sensor does not have alert capacity. Freestyle Libre 2 will have optional alerts but is not yet available in Australia. Box 2 – Internationally accepted continuous glucose monitoring (CGM) metrics for clinical use and comparison between devices (adapted from guidelines)2,6,7 Percentage sensor wear and data captured — to gauge completeness of data capture (optimal wear time assessed as > 70% capture across a 14‐day time period) Mean glucose — the sum of all glucose levels, divided by number of measurements; a surrogate of overall glucose control, with reasonable correlation with glycated haemoglobin Glucose variability — standard deviation of glucose/mean glucose × 100 = coefficient of variation (CV); goal is CV < 36% in type 1 diabetes Time in range (3.9–10.0 mmol/L) — aim > 70%; the ranges can be tailored to the individual depending on their age and comorbidities (eg, older individuals or pregnancy) and provide an estimate of level of current glycaemic control otherwise not reflected in a glycated haemoglobin measurement Time in hypoglycaemia: ► < 3.9 mmol/L — goal < 4% (includes proportion values < 3.0 mmol/L) ► < 3.0 mmol/L — goal < 1% ► number of CGM events < 3.0 mmol/L for 15 minutes or more in previous 2 weeks — focuses on the importance of moderate hypoglycaemia Time in hyperglycaemia: ► 10 mmol/L — goal < 25% (including time > 13.9 mmol/L) ► 13.9 mmol/L — goal < 5%. Other reportable data (from insulin pump downloads): ► total daily insulin, % basal — a summary of current total insulin delivery, split into dose delivered as basal and bolus insulin; this allows for comparison between visits Standardisation of CGM reporting improves comparisons between devices and treatments and enhances decisions in diabetes care for both clinicians and people with diabetes Box 3 – Access to subsidised continuous glucose monitoring (CGM) through the National Diabetes Services Scheme (NDSS)*9 The following groups can access CGM or flash glucose monitoring through the NDSS: Children and young people under 21 years of age with type 1 diabetes Children and young people with conditions very similar to type 1 diabetes, such as cystic fibrosis‐related diabetes or forms of genetic diabetes (including maturity onset diabetes of the young), who require insulin Women with type 1 diabetes who are actively planning pregnancy (up to 12 months before conception), pregnant or immediately post‐pregnancy (pregnancy plus 3 months from expected date of confinement) People with type 1 diabetes aged 21 years or older who have concessional status Applications can be made through the patient’s credentialled diabetes educator or endocrinologist *Criteria valid from 1 March 2020. Box 4 – Major components of insulin pump and continuous glucose monitoring (CGM) set‐up Box 5 – Steps in managing ketosis caused by insulin pump line occlusion (in the absence of vomiting) Insulin pen injection using pump‐advised correction dose for high blood glucose Replace insulin pump cannula set Run increased basal insulin rates (200%) temporarily for 2 hours to restore subcutaneous insulin reservoir and missed insulin At 2 hours, deliver correction insulin dose with insulin pump Monitor blood ketones every 3–4 hours using a ketone meter to ensure ketone levels < 1.5 mmol/L Box 6 – Options for insulin delivery and glucose monitoring CGM = continuous glucose monitoring; HCL = hybrid closed loop; MDI = multiple daily injections; PLGS = predictive low glucose suspend; SMBG = self‐monitoring of blood glucose.

Jennifer R Snaith · D Jane Holmes‐Walker

Posts can provide valuable feedback during public consultation for health guidelines

Benjamin C Riordan · Daniel T Winter · Paul S Haber · Carolyn A Day · Kirsten C Morley

Rapid advances in AI in ophthalmology are a harbinger of things to come for other fields of medicine The autonomous detection and triage of eye disease, or even accurate estimations of gender, age, and blood pressure from a simple retinal photo, may sound like the realms of science fiction, but advances in artificial intelligence (AI) have already made this a reality.1 Ophthalmology is at the vanguard of the development and clinical application of AI. Advances in the field may provide useful insights into the application of this technology in health care more broadly. Artificial intelligence Once described as the capacity of intelligent machines to imitate human intelligence and behaviour, AI now describes many theories and practices used to achieve computer intelligence (Box 1).2 Machine learning is an application of AI that uses algorithms or statistical models to make decisions or predictions. Complex patterns and relationships are learned from data to generate an outcome.2 Machine learning traditionally relies on the extraction of features from the data by human operators which then serve as input variables to optimise algorithm performance. The performance of these systems is constrained by the features that are recognised as important by humans. In contrast, artificial neural networks are an advanced method of machine learning able to extract features without explicit programming.2 Deep learning is the construction of multiple layers of artificial neural networks which can identify features in data that are not recognisable by humans. Although deep learning systems may be powerful, they lack human‐crafted inputs, meaning that large quantities of data are typically required to train algorithms. Artificial intelligence in ophthalmology As a discipline, ophthalmology is at the forefront of AI system development and translation in clinical practice. Leading uses of the technology include detecting, classifying and triaging a range of diseases, such as diabetic retinopathy, age‐related macular degeneration (AMD), glaucoma, retinopathy of prematurity, and retinal vein occlusion, from clinical images.3 The increasing global burden of eye diseases, coupled with the development of new therapies for previously untreatable conditions, has served as a major driver for AI innovation in ophthalmology. As a case in point, there are presently over 430 million people living with diabetes, most of whom require annual or biennial screening for retinopathy using retinal photography. This vast demand for diabetic eye screening services has stimulated the development of AI algorithms to identify sight‐threatening disease. Several algorithms have achieved performance that meets or exceeds that of human experts.4,5 Accordingly, in 2018, the United States Food and Drug Administration approved an AI system to detect referable diabetic retinopathy from retinal photographs, the first autonomous diagnostic system to be approved in any field of medicine.6 Advances in deep learning have extended to other imaging modalities that are commonly used in ophthalmology. Ocular coherence tomography is an imaging technology that produces highly detailed, depth‐resolved images of the retina. A recent collaboration between researchers and clinicians at Google DeepMind, Moorfields Eye Hospital and University College London culminated in the development of a deep learning system capable of detecting and triaging more than 50 different retinal conditions at levels equivalent to a panel of experienced ophthalmologists.7 AI systems with the capacity to detect a wide range of diseases, such as this, are likely to be most useful in clinical practice. A highly anticipated innovation is the development of AI systems capable of accurate disease prediction. Such tools could assist in managing patient expectations, improve the quality of care and reduce treatment costs.3 In ophthalmology, prediction models have been trained to personalise re‐treatment intervals for patients with neovascular AMD,8 predict progression from early to late AMD,9 estimate the extent of future visual field defects in patients with glaucoma,10 and predict diabetic retinopathy progression.11 Although these models presently achieve only moderate levels of accuracy, their performance has been shown to be superior to humans in several studies.3,8 Future advances in the accuracy of prediction models will likely come from the use of large longitudinal datasets drawing on multiple data sources, together with the development of more advanced AI systems.3 Despite these significant advances, AI systems are not in widespread clinical use and in some cases real‐world performance has been inferior compared with in silico validation.2,3 Training and validation of deep learning algorithms with large, representative data (eg, data from people of different ethnicities) acquired using multiple devices (eg, different retinal camera models) and data collection protocols (eg, retinal photographs acquired with and without pupil dilation) are key to achieving clinical applicability.4,5 This approach was used in the development of deep learning systems for retinal photographic screening for diabetic retinopathy, AMD and glaucoma which are now being used in large scale screening programs in Singapore and China.4,5 In these programs, AI is used to identify images without evidence of disease, so that human graders can focus their efforts on the images of those with disease, enabling improved efficiency and cost savings.12 Challenges to the clinical adoption of artificial intelligence Several obstacles to the adoption of AI in health care remain. The training of deep learning systems requires access to large amounts of medical data which has significant implications relating to privacy and data protection. In the context of ophthalmology, this is particularly pertinent, as the retinal vasculature may be considered biometric data, making it impossible to completely anonymise retinal photographs.3 Furthermore, characteristics that are not visible to human examiners, such as age and sex, can now be accurately predicted from a single retinal photograph using deep learning.1 Several recent major breaches of data protection laws relating to AI system development have already come to light.13 While individual patient data used to train an algorithm do not remain within the system, incorrect handling and sharing of data may lead to patients withdrawing consent to the use of their data under General Data Protection Regulation laws. It is not certain how data withdrawal requests will be dealt with when an individual’s data have been used in the process of training a deep learning system. Accordingly, developments in AI need to be accompanied by advanced data protection and security measures. Another challenge to the acceptance of deep learning algorithms in medicine is the difficulty in determining the basis for clinical decisions made by these systems, informally described as the “black box” problem. Visualisation tools have been developed to assist clinicians by highlighting the salient image features that contribute to the AI system classification (Box 2).12 This has the potential to create trust in system‐generated decisions, particularly if the features correspond with those used by experienced clinicians for clinical decision making.14 Interpretability is particularly important when considering legal liability in the event of patient harm arising from the use of AI in medicine. In traditional malpractice cases, a physician may be asked to justify the basis for a particular clinical decision and this is then considered in light of conventional medical practice.15 In comparison, challenges in identifying the basis for a given decision made by AI might pose problems for clinicians whose actions were based on that decision. The extent to which the clinician, as opposed to the technology manufacturer, should be held accountable for harm arising from AI use is a subject of intense debate.15 Factors such as the manner in which these AI systems are used and their classification as either products or software are likely to have important bearings on how cases are litigated.15 Further challenges for existing regulatory frameworks come from algorithms that continue to learn and evolve over time.15 Understanding how a given system is trained, its accuracy, and its operational limits is of great importance. Oversampling of a particular population or disease severity during training has the potential to introduce bias.4 Therefore, consideration of performance thresholds will help to inform appropriate use of AI systems. The Australian Government, through the CSIRO and Data61;16 the Australian Council of Learned Academies;17 the Australian Academy of Health and Medical Sciences;18 and specialty groups, such as the Royal Australian and New Zealand College of Radiologists,19 have made significant efforts to develop frameworks and policies for the effective and ethical development of AI. These consultative works have highlighted key priorities, including building a specialist AI workforce, ensuring effective data governance and enabling trust in AI through transparency and appropriate safety standards. Through targeted investment in research and development, Australia is aiming to advance its AI competitiveness. These framework documents provide guidance for developers, clinicians and health care consumers to navigate this rapidly evolving field. Broad dissemination of these documents should form part of a wider public engagement and education campaign to ensure that AI is developed and used in a considered and careful manner in health care. Rapid advances in AI in ophthalmology are a harbinger of things to come for other fields of medicine. While these technologies may eventually lead to more efficient, cost‐effective and safer health care, they are not a panacea in isolation. The successful integration of AI into health systems will need to first consider patient needs, ethical challenges and the performance limits of individual systems. Box 1 – Relationship between artificial intelligence and its subtypes Box 2 – Original retinal photograph of right eye with macular degeneration (A). Heat map of image A showing visualisation of traditional features associated with macular degeneration, such as central scarring (B). Original retinal photograph of left eye with referable diabetic retinopathy (C). Heat map of image C showing visualisation of traditional features, such as micro‐aneurysms and haemorrhages (D)

Jane Scheetz · Mingguang He · Peter Wijngaarden

To the Editor: We congratulate Hill and colleagues1 for their timely research on the performance of symptom assessment smartphone applications (apps) in Australia. The apps in the study were selected using structured criteria2 to identify those featuring most prominently in internet search engines and app stores. However, we note that this strategy is biased against an important class of symptom checkers. Because the app store search included “medical diagnosis” and “health symptom diagnosis”, the authors’ approach was less likely to identify many symptom checkers regulated in Europe under the CE (Conformité Européene) Marking system. Specifically, these apps must not describe themselves as “diagnostic tools”, as diagnosis is a function carried out by a doctor. We believe this to be the reason why the CE‐marked Ada health assessment app was not identified or selected by the authors.1 This represents a missed opportunity for analysis, as Ada has been freely available in Australia since 2016,3 and was downloaded at least 200 times more frequently in Australia between November 2018 and January 2019 than either Symptomate or Symcat, which were included in the study (App Annie [www.appannie.com] download data; viewed June 2020). Other studies have found that the Ada app performs well when compared with the other apps assessed, as recently published.4

Stephen Gilbert · Paul Wicks · Claire Novorol

The MJA aims to be an outstanding general medical journal, broadly relevant to all specialties in medicine and health, with a national and global focus, a journal that influences policy and practice

Nicholas J Talley

At the end of 2019 and into 2020, catastrophic fires in Australia consumed homes, lives, wildlife and land. Just as the fires subsided, Australia, like the rest of the world, faced another emergency — the COVID‐19 pandemic.1 It is instructive to reflect on lessons from the health disasters of the past year. Following publication of The Lancet Countdown on health and climate change,2 the Medical Journal of Australia (MJA)–Lancet Australian Countdown on health and climate change was published in December 2020.3 This annual report on health and climate change in Australia is in its third year and comprises the efforts of five Australian institutions, in collaboration with University College London, facilitated by a partnership between The Lancet and the MJA.3 All three reports make sobering reading.3,4,5 2019 was Australia’s hottest and driest year on record, with average temperatures 1.52°C above normal and mean rainfall 40% below the 30‐year average before 1991.3 Australia’s 2019–20 bushfires burned 10 million hectares, directly killed 33 people and destroyed more than 3000 homes.6 Smoke engulfed major capital cities, including Sydney and Melbourne, and smoke exposure caused an estimated 417 excess deaths and over 3000 hospital admissions.3,7,8 The catastrophe laid bare how extreme heat is a severe health risk.9 The ecological damage of the bushfires was enormous;6 almost 3 billion animals were killed or displaced, and natural systems of biodiversity and species were harmed, perhaps irreparably.6 Severe storms and floods followed the fires, bringing further damage. Insured losses from disaster events totalled AU$3.7 billion in 2019, with bushfires accounting for $2.2 billion, although the total costs of the so‐called Black Summer fires could be much higher.3 The devastation of the bushfires led the Australian Government to establish the Royal Commission into National Natural Disaster Arrangements. The final report of the Royal Commission in October 2020 identified climate change as a major driver and acknowledged the risk of increasing extreme weather events.6 However, the Royal Commission’s scope was limited to disaster management (mitigation, preparedness, response and recovery) and did not discuss root causes of climate change such as the fossil fuel industry’s grip on Australia’s energy infrastructure, economy, political will and public discourse. Australia has no decisive national plan to address climate change and its health consequences.3 The Australian Government is a signatory to the Paris Agreement, but has declined to affirm net zero carbon emissions by 2050 — or by any date — unlike the UK and the EU; China has also committed to this goal by 2060. Unlike this inadequate approach to the climate crisis, Australia’s response to COVID‐19 was rapid and effective, despite facing the pandemic while the last bushfires still burned.10,11 Strong community engagement with public health measures enabled effective management of the first and second waves, making Australia’s, together with those of New Zealand and parts of Asia, among the more successful responses to COVID‐19.12 Key to this success was the valuing by governments of science and data to guide decision making. The pandemic forced politicians from across the Australian political divide to prioritise the evidence and expertise of the medical, scientific and public health communities over the voices of conservative commentators, business leaders and politicians. Tough political decisions were made for the sake of the nation’s health. This bipartisan, science‐based approach is a model for the future management of climate change, if implemented alongside an appropriate national plan. Australia’s First Nations people, who are at increased risk of poorer health outcomes than the general population in a pandemic, have provided exceptional leadership in their response to COVID‐19, resulting in low rates of virus transmission thus far.13,14 The country’s Indigenous populations are also disproportionately vulnerable to future climate change natural disasters.6 Since the traditional owners of Australia’s land are effective, resilient caretakers of the country and experts in land and fire management, the Royal Commission recommended federal, state and territory governments learn from and engage with their expertise.6 As another initiative capitalising on local expertise, The Lancet Countdown’s regional report in partnership with the MJA has led to improved performance indicators for climate change. For example, the Australian Countdown reports3,4,5 developed the wildfire (bushfire) indicator, which the Countdown is now adopting globally. These data have encouraged more direct engagement with Australian policy makers and health professionals, and provided direct funding guidance to Australia’s National Health and Medical Research Council (NHMRC), which is expected to translate into funding changes in 2021.15 Australia’s leading medical and nursing bodies have recognised climate change as a health emergency.8 Governments of states and territories have committed to zero net carbon emissions by 2050, with climate change adaptation plans incorporating the health sector and investment in renewable energy.3,16 With the unprecedented disasters of 2020, public sentiment in Australia has shifted, as more people realise climate change is here now, with impacts for all. In November, the Climate Change (National Framework for Adaptation and Mitigation) Bill 2020 was introduced into the Australian federal Parliament by the independent Member of Parliament Zali Steggall, with wide public support, including from the Australian Medical Association and more than 100 major businesses.17 The outcome of the 2020 US election and the environmental platform of the incoming administration of Joe Biden coincides with a more positive stance in Australian politics towards addressing climate change.18 In the MJA–Lancet Countdowns,3,4,5 Australia embraced the importance of local data, local experts and local stories. A regional China Countdown report is also being published in parallel this year,19 and in 2021, there will potentially be Countdown collaborations for the EU, South America, the US, and Small Island Developing States. Looking forwards, Australia should as a priority establish a National Health and Climate Change Centre within the Australian Government Department of Health to develop a National Plan for Health and Climate Change with real‐time monitoring. As well as preparing to manage climate‐related health sequelae, Australia’s health sector should commit itself nationally to zero net carbon emissions by 2040 in line with the National Health Service in the UK, preferably with the states and territories responsible for implementing evidence‐based interventions.20 Reducing unnecessary medical tests and procedures will serve to reduce carbon emissions, health care costs and harmful outcomes.21 Research funded by the NHMRC and the Medical Research Futures Fund should guide better ways to efficiently reduce the carbon footprint of Australia’s health care services. Australia has an obligation under the Paris Agreement to submit enhanced nationally determined contributions by the end of 2020. We recommend that the Australian Government agree to a target of a 50% reduction in carbon emissions by 2030, which is what is likely required to limit global warming below 1.5°C.3,4,5 The Australian Government needs to recognise that fossil fuels are no longer a sound investment and join with other jurisdictions that are committed to shifting completely to renewable energies to make that sector the most cost‐effective for jobs and energy security. In Australia the crises of 2020 were unprecedented, shocking and predictable. We remain hopeful all Australian governments will aspire to the leadership shown nationally with the COVID‐19 pandemic and effectively deal with climate change now, understanding the major health risks of neglecting this issue. We anticipate health and corporate leaders, as well as leaders across other sectors, will continue to drive change. Interrogating successes and failures nationally in the MJA–Lancet annual Australian Countdown provides a robust model for monitoring and positive change. The flow on benefits to health and wellbeing, the economy and society from such change will be enormous. This article is co-published in The Lancet.22

Nicholas J Talley · Fiona J Stanley · Tamara Lucas · Richard C Horton

The dangers of modern football are often scrutinised, but has safety actually evolved over time?

Jacob L Jewson · Peter Brukner · Thomas J Gara · Lauren V Fortington

In health care, we now need to be curators and disseminators of accurate and timely information, not solely producers, this includes digital sources

Rebecca A Szabo

Rapid dissemination of information should not come at the expense of quality, ethical standards or oversight The coronavirus disease 2019 (COVID‐19) pandemic has resulted in wide‐ranging health, social and economic impacts. By October 2020, global cases exceeded 41 million, with 1.1 million deaths.1 Urgent requirements for information were met with data on epidemiology, clinical features and recommended management being circulated on social media and pre‐publication servers. While this has allowed timely sharing of data, it has also brought risk of misinformation, with consequent changes to medical practice and misdirection of scarce resources based on flawed evidence. Medical publishing uses peer review to provide independent and critical assessment to verify data integrity, validity of interpretations, and confidence in conclusions. This process can take many weeks; however, in a rapidly spreading pandemic, speed is a competing priority. We hypothesised that these considerations may have altered the nature of medical publication. Accordingly, we characterised various aspects of COVID‐19‐related articles published in the five leading general medical journals with the highest impact factors (Web of Science) compared with an equivalent period in the preceding year. Procedures for identifying, classifying and comparing publications were specified a priori. Research ethics approval was not required. Publications were identified in the United States National Library of Medicine PubMed database. All articles published between 1 January and 31 May (inclusive) in 2019 and 2020 in The New England Journal of Medicine, The Lancet, JAMA, The BMJ and Annals of Internal Medicine were included. The sampling timeframe was defined by the first public health notification of COVID‐19 in China on 31 December 2019, ending at the time of the conduct of the literature search (Box 1). Within the 2019 search results, 60 articles were randomly selected using a random number generator in Stata 15.1. Publications without abstracts were excluded. Journal websites for each study period were searched for retracted articles. Three reviewers independently abstracted the variables contained in Box 2 and Box 3. The h‐index (a measure of publication productivity and citation impact) of the first and last author was taken from Web of Science. A fourth investigator reviewed all data, harmonising interpretations and resolving any errors. Data were analysed using Stata 15.1. Skewed continuous data were described using medians with interquartile ranges (IQRs) and compared using the Wilcoxon–Mann–Whitney test. Categorical data were compared using the Fisher exact test or χ2 test as appropriate. Exact P values are reported and those less than 0.05 deemed significant. During January to May 2020, PubMed listed 4001 articles, of which 1120 (28%) were related to COVID‐19. There were 134 articles with PubMed‐coded abstracts which were included for full review (Box 4). One additional COVID‐19 article was identified in the search for retracted articles but excluded from quantitative comparisons because it lacked an abstract. During the same period in 2019, 54 articles were ultimately identified as eligible for comparison (Box 4). Compared with 2019, among the COVID‐19‐related publications in 2020, there were more case reports or case series, cohort studies, editorials and commentaries and fewer randomised controlled trials (7/134 [5.2%] v 19/54 [35.2%]) (Box 2). A similar proportion (37/52 [68.5%] non‐COVID‐19‐related articles v 74/134 [55.2%] COVID‐19‐related articles; P = 0.09) reported primary data. Of the 2019 articles, only two of 54 (3.7%) originated in China, whereas 32 of 134 (23.9%) of the COVID‐19 articles published in 2020 were from China. The proportion of COVID‐19 articles in 2020 for which a correction was published was higher than for non‐COVID‐19 articles published in 2019 (28/124 [20.9%] v 4/54 [7.4%] respectively; P = 0.03). Time to the first publication of a correction was no different (median, 6 days [IQR, 4–14] v 7.5 days [IQR, 5–18] respectively; P = 0.53). Three 2020 COVID‐19 articles,2,3,4 but none of the 2019 articles, were retracted after publication. Only one journal, JAMA, routinely reported when a manuscript was submitted. In this journal, the median time from submission to publication fell from 139 days (IQR, 130–144) in 2019 to 23 days (IQR, 12–30) in 2020 (P < 0.001). The median number of authors and their publication productivity and impact, as quantified by their median h‐indices, were similar. There was no statistically significant difference in the number of studies willing to share data under appropriate circumstances (P = 0.19), or those receiving commercial funding (P = 0.97). The measured characteristics of randomised trials related to COVID‐19 were not statistically different to studies of an equivalent type published in the preceding year; however, numerically fewer subjects (median, 199 [IQR, 127–397] v 424 [IQR, 225–1076]; P = 0.07) and centres (median, 10 [IQR, 1–55] v 30 [IQR, 4–168; P = 0.15) participated (Box 3). Similarly, the observational study sample size was significantly smaller (median, 152.5 [IQR, 15–3481] v 191 972.5 [IQR, 1407.5–756 444]; P < 0.001), and the number of participating centres was numerically lower in the 2020 COVID‐19 group (median, 1 [IQR, 1–7] v 26 [IQR, 1–49]; P = 0.07). While not significantly different between groups due to the low numbers, 11 (16.7%) observational studies among the COVID‐19 publications did not report oversight by an ethics committee or institutional review board, and only nine (56.3%) case reports and case series with ten patients or fewer stated that patient consent had been obtained or that an exemption from this requirement had been granted. In the first 5 months of the COVID‐19 pandemic, the five leading medical journals published a substantial number of articles that differed in many respects from their usual material. The journals examined were the clinically focused general medical journals with the top five Web of Science 2019 impact factors, ranging from 21.3 to 74.6, representing the medical literature with the greatest international influence on health policy and clinical practice. As reasonably expected, there was a seven‐fold reduction in the proportion of articles reporting randomised controlled trials, and a compensatory increase in small case series, opinions and editorials. While there were few (n = 2) articles in the random selection of 2019 papers that were published from China, nearly one‐quarter of the COVID‐19 publications came from this country, as anticipated given the location of the earliest cases. There was no difference in the median h‐indices of authors, suggesting experienced academics pivoted rapidly to COVID‐19 research. In circumstances which usually require consent, just under half of the COVID‐19 studies did not explicitly state consent was obtained, despite clear recommendations by the International Committee of Medical Journal Editors.5 The proportion of articles that referenced appropriate ethics committee or institutional review oversight was statistically unchanged; however, it is still a concern that 11 (16.7%) observational COVID‐19 studies lacked any statement to this effect. In addition, several other articles stated that they had been exempted from the requirement for ethical review due to the nature of the pandemic. Respect for personal autonomy and the value of independent oversight have always imposed additional workload on those seeking broader public health benefits. If COVID‐19 has created challenges in adhering to the usual practices of obtaining ethics approval and consent, consideration should be given to whether these processes could be amended to improve speed and accessibility, particularly during global health emergencies. There was a near three‐fold increase in the proportion of studies that published corrections, perhaps reflecting the observed reduction in time from submission to publication observed in the one journal for which these data were available. It is likely this figure is an underestimation, given that corrections and retractions would be expected to continue over time. Three COVID‐19 studies were retracted. The publication of one of these articles4 had important implications, resulting in the temporary cessation of the World Health Organization's trial of hydroxychloroquine.6 While the corrections and retractions may be an artefact of increased speed to publication, it is also possible that their higher number might be the effect of enhanced focus on research related to COVID‐19. Nonetheless, journals must retain the integrity of review processes if they are to offer value beyond alternative online means of information dissemination. This review has found similar results to bibliometric studies relating to the COVID‐19 pandemic, which have identified higher numbers of case series and reviews and fewer randomised clinical trials.7,8,9 We did not examine other articles from 2020 to understand the effect of COVID‐19 on contemporaneous publications, or to be able to comment on whether observed changes were specific to COVID‐19 or true of all 2020 articles. We note the convenience sampling of two similar periods may overestimate the magnitude of our findings. The cohort of 2019 studies for comparison was selected at random, rather than being matched by study type or size. When identifying h‐indices, we had difficulty identifying some Chinese authors, highlighting a bias against researchers without a name that can be distinctively rendered in the English language alphabet. Further implementation of unique author identifiers, such as the Open Research and Contributor ID (ORCID; www.orcid.org) or ResearcherID (Clarivate Analytics) would address this problem. We did not assess the quality of published studies or adherence to reporting guidelines. As part of their early response to the worldwide problem presented by the COVID‐19 pandemic, there was a significant change in the characteristics of articles published by leading medical journals, with some evidence of a tendency towards publishing articles prematurely and those with lower internal validity. While these unique circumstances no doubt warranted such a change, rapid dissemination of information should not need to come at the expense of quality, ethical standards or oversight. Others have suggested several solutions to this challenge, including a two‐track review process for pandemic and non‐pandemic research, rapid preliminary assessment of research methodology by skilled in‐house reviewers before deciding whether to send for peer review, sharing of peer‐reviews between reviewers and journals, and mentored peer reviewing by research trainees.10 As part of pandemic preparedness, planning to facilitate augmentation of resources available to medical publishers, allowing maintenance of standards of review, should occur. Box 1 – Search strategy ((“JAMA”[Journal]) or (“The New England Journal of Medicine”[Journal]) or (“Annals of Internal Medicine”[Journal]) or (“BMJ”[Journal]) or (“Lancet”[Journal])) and (2020/1/1:2020/5/31[Date — Entry]) or and (2019/1/1:2019/5/31[Date — Entry]) Articles related to COVID‐19 were identified by adding and ((“covid”[All fields]) or (“coronavirus”[MeSH Terms]) or (“coronavirus”[All fields]) or (“coronaviruses”[All fields])) Box 2 – Characteristics of publications 2019 non‐COVID‐19 2020 COVID‐19 P Total number of articles 54 134 Article type Systematic review/meta‐analysis/narrative review 8 (14.8%) 16 (11.9%) < 0.001 Randomised controlled trial 19 (35.2%) 7 (5.2%) Cohort study 11 (20.4%) 25 (18.7%) Cross‐sectional study 5 (9.3%) 8 (6.0%) Case–control study 1 (1.9%) 2 (1.5%) Case series 2 (3.7%) 30 (22.4%) Case report 0 (0.0%) 4 (3.0%) Diagnostic evaluation 0 (0.0%) 1 (0.7%) Opinion 7 (13.0%) 33 (24.6%) Other 1 (1.9%) 8 (6.0%) Reported primary data 37 (68.5%) 74 (55.2%) 0.09 Correction published 4 (7.4%) 28 (20.9%) 0.03 Days from publication to correction, median (IQR) 6 (4–14) 7.5 (5–18) 0.53 Retracted 0 (0.0%) 3 (2.2%) 0.56 h‐index of first author, median (IQR) 13.5 (3–36) 11.5 (6–30) 0.54 h‐index of last author, median (IQR) 26 (14–38) 21 (10–38) 0.14 Associated editorial of eligible articles 21 (38.9%) 44 (32.9%) 0.43 Number of masthead authors, median (IQR) 8 (5–19) 7 (4–18) 0.52 Number of total authors, median (IQR) 8 (5–23) 7 (4–19) 0.23 Region of origin China 2 (3.7%) 32 (23.9%) < 0.001 United States 24 (44.4%) 67 (50.0%) Europe 20 (37.0%) 24 (17.9%) Rest of world (high income countries) 3 (5.6%) 11 (8.2%) Rest of world (low income countries) 5 (9.3%) 0 (0.0%) COVID-19 = coronavirus disease 2019; IQR = interquartile range. Box 3 – Characteristics of studies reported table#t3 tbody td:nth-child(n+2) P. Pleft { text-align: center; } 2019 non‐COVID‐19 2020 COVID‐19 P Randomised controlled trials 19 7 Number of subjects, median (IQR) 424 (225–1076) 199 (127–397) 0.07 Participating centres, median (IQR) 30 (4–168) 10 (1–55) 0.15 Studies that received funding of any type from a commercial source 8 (42.1%) 3 (42.9%) 0.97 Studies in which a commercial entity had influence over any aspect of study conduct or reporting 7 (36.8%) 2 (28.6%) 0.69 Studies stating willingness to share data under appropriate circumstances 15 (78.9%) 7 (100.0%) 0.19 Studies stating individual patient consent or waiver was granted 19 (100.0%) 7 (100.0%) 1.0 Studies noting review by ethics committee 19 (100.0%) 7 (100.0%) 1.0 Observational studies* 19 66 Number of subjects, median (IQR) 191 972.5 (1407.5–756 444) 152.5 (15–3481) < 0.001 Participating centres, median (IQR) 26 (1–49) 1 (1–7) 0.07 Studies that received funding of any type from a commercial source 0 (0.0%) 4 (6.1%) 0.27 Studies in which a commercial entity had influence over any aspect of study conduct or reporting 0 (0.0%) 3 (4.5%) 0.34 Studies stating willingness to share data under appropriate circumstances 8 (42.1%) 15 (22.7%) 0.09 Studies not stating individual patient consent was obtained or a waiver was granted 3 (15.8%) 17 (25.8%) 0.37 Studies not noting review by ethics committee 0 (0.0%) 11 (16.7%) 0.06 Case reports/case series (≤ 10 patients) 1 16 Studies stating individual patient consent was obtained 1 (100.0%) 9 (56.3%) 0.40 COVID-19 = coronavirus disease 2019; IQR = interquartile range. * Observational studies included cross-sectional studies, case–control studies, cohort studies and case series reporting data from one patient or more. Box 4 – Publication identification flow diagram COVID‐19 = coronavirus disease 2019.

Kirsty A Whitmore · Kevin B Laupland · Clare M Vincent · Felicity A Edwards · Michael C Reade

Artificial intelligence is being used in several different ways to curb the current pandemic while demonstrating its potential to be even more effective for the next one

Ian A Scott · Enrico W Coiera

There are legal considerations for both clinicians and patients when recording health care consultations Studies show that patients often have inaccurate recall of health care events and diagnoses.1 Concentration during a medical consultation may be “hampered by unspoken anxieties or pain, making it difficult to recall detail”.2 Audio recordings of consultations can be useful for patients and clinicians to assist memory and understanding. They have mainly been evaluated in oncology and paediatrics.3,4 Patients report that listening to their consultation recording increases knowledge and understanding of their illness, and recordings can assist with treatment decision making, increasing a sense of empowerment.5 Sharing recordings with family can facilitate support and understanding. Clinicians likewise recognise recordings’ benefits for patients and for improving the quality and efficiency of their care.6 Research in the United Kingdom found that 69% of patients wish to record consultations.7 Increasingly, patients are using smartphones to record consultations, either with permission or covertly.7,8 Recording systems have been developed by health services themselves, transformed by the ubiquitous use of smartphones and other flexible technologies.9,10,11 Examples include the Open Recording Automated Logging System (ORALS) software in the United States9 and telephone‐based digital recording in Denmark.11 In Australia, the Second Ears smartphone app, developed at the Victorian Comprehensive Cancer Centre in 2018, is designed to make recordings available to both the patient and the hospital health information management service.6,10 Patients can choose whether to download and use the app (either before their appointment or in the clinic), access the recordings on their smartphone, and share them with family and friends.6,10 Common design features of such health service‐led recordings address data security, file storage and patient consent. Whether the clinician or the patient controls the recording process may differ across technology platforms; for instance, in the Danish example above, the clinician initiated the recordings, whereas with Second Ears the patient would do so. The use of consultation recordings often raises legal questions.5,7,10,12 In this article, we compare the legal implications of overt and covert recordings of health care consultations and address key concerns identified by clinicians, notably the requirement for consent to record and share the recording, and the use of recordings in negligence claims.8,13,14,15 We distinguish between three recording types: Overt patient‐led recordings: for example, a patient recording a consultation with the clinician's consent. These recordings are akin to a patient's handwritten notes. Overt health service‐led recordings: for example, the Second Ears app, where both clinician and patient consent (actively or impliedly) to the recording; the app is facilitated by the health service and the primary version of the recording stored on their system. Covert patient‐led recordings: for example, a patient recording without the clinician's knowledge or consent. As each legal question is identified, we consider the law in the context of the Second Ears app. This article is general in nature and does not constitute legal advice. References to legislation are current at 13 October 2020. References to state or territory laws relate to the location of the recording or the place at which the sharing of the recording originated. We do not address the issue of intentional recording of private conversations by third parties, either overtly or covertly. Consent to record a consultation Clinician consent to patient‐led recordings Clinicians consider that their consent to be recorded is a key issue. Perhaps surprisingly, at law in many Australian jurisdictions, the patient need not obtain explicit consent from the clinician. In Victoria, Queensland and the Northern Territory, the law does not consider a recording of a conversation that is made by one of the parties (as opposed to a third party). In New South Wales, Tasmania and the Australian Capital Territory, patients can record their consultation without the clinician's consent (or, by extension, their knowledge) if the recording is only for the patient's own use (ie, to listen back to the recording later), or to protect their lawful interests (such as in a negligence claim). In South Australia and Western Australia, clinician consent is required (ie, two‐party consent) for recording a consultation for later listening‐back by the patient (Box 1). Patient consent to health service‐led recordings Where the recording is made on an app like Second Ears with data stored by the health service, this is an act of health information collection about an individual that requires the patient's express or implied consent. The patient's decision to download and install the app can act as implied consent; the app's terms and conditions could also include a clear statement about patient consent. Consent of other people captured incidentally in any overt recording A consultation recording — whether patient‐led or health service‐led — might accidentally capture another conversation, for instance from the clinic's reception desk. No consent of the third party is needed in this case, because they are not a party to the recorded conversation. Typically, Australian surveillance device laws do not regulate recordings of conversations occurring in circumstances in which the parties ought reasonably to expect to be overheard, such as in public or an open hospital ward. This means that if a patient is overtly recording their own consultation while in a curtained cubicle, their inadvertent capture of another clearly heard conversation in the next cubicle would not require the consent of those having that conversation. Consent when someone else joins any overt recording If another person, such as the patient's relative or another clinician, enters a room where a consultation is being recorded, but does not join in the conversation, the new person is not a party to it and that person's consent is therefore not needed. However, if the new person does join the conversation, they become a party to it. Box 1 indicates when that new party's consent to be recorded is required. In SA and WA it is usually required. In NSW, the ACT and Tasmania it is required if the patient makes the recording intending to share it with anyone else, but not if the recording is intended only for the patient to listen to. Consent, when required, can be either express or implied. An example of how this situation might be addressed could be a health service policy to have a door sign stating prominently that a recording is in progress and that by entering the room the new participant consents to be recorded. A person entering the room could then signal their non‐consent by verbally requesting the recording be stopped. This applies to health service‐led and patient‐led recordings. Covert recordings by patients Covert recording by patients is not uncommon; a survey conducted in the UK found that 15% of respondents self‐reported recording clinical encounters without permission. A further 35% of respondents would consider covert recordings in the future.7 In the US, a similar survey found that far fewer respondents recorded covertly (2.7%);8 possibly because some health services routinely provided permission for recording. Currently, the proportion of Australian patients who record covertly is unknown; anecdotally, however, clinicians report that it is occurring.16 Covert recording has been described as a topic of “significant legal ambiguity”.17 In Australia, as noted above, the law varies significantly by jurisdiction. Only SA and WA require two‐party consent and thus prohibit patients covertly recording for their own use (Box 1). Covert recordings: legal penalties Not all consultation recordings require consent. In SA and WA, where two‐party consent is required, a person making a covert recording for their own use is subject to legal penalties; for example, in SA, fines of up to $15 000 or imprisonment for up to 3 years. In Toth v DPP (NSW) [2014] NSWCA 133, a case concerning a patient's illegal covert recording, the magistrate imposed an 18‐month good behaviour bond. Dealing with unwanted recording If their consent is legally required but the clinician does not want to be recorded, they can simply ask the patient to discontinue the recording. Regardless of whether the act of recording legally requires their consent, a clinician's refusal to be recorded, or the exposure of covert recording by a patient, may lead to breakdown of the therapeutic relationship,14 necessitating transfer of care to another clinician as per the Medical Board of Australia's code of conduct (https://www.medicalboard.gov.au/codes-guidelines-policies/code-of-conduct.aspx). While discontinuing a relationship may be appropriate in the context of misuse of an audio recording or its use with malicious intent, it would be a drastic response to a simple request by the patient to record, given the benefits of doing so. Health service‐led systems such as Second Ears may overcome this problem by incorporating clear frameworks around participation, consent and sharing. Sharing recordings with others Health care organisations sharing recordings Recordings made by the health service with the patient's consent (eg, via the Second Ears app) form part of the medical record and the organisation can lawfully share the recording in various ways, which are broadly similar across Australian states and territories. These include: with the person's consent; without the person's consent for a directly related purpose as long as the person would “reasonably expect” the disclosure (eg, in transferring care to another provider at the same service: F v Medical Specialist [2009] PrivCmrA 8); to defend a legal claim; for research in the public interest (if certain privacy guidelines are met, such as those set out by the National Health and Medical Research Council18); and with an immediate family member of the patient for compassionate reasons or to provide the patient with care when the patient is incapable of providing consent. This mirrors other parts of the medical record such as written notes and scans. If the recording is de‐identified (which may be difficult because voice patterns are distinctive and health information discussed during consultations is often reasonably identifiable), it can usually be used without patient consent for communication training within the health service. Consent may provide a more appropriate legal basis for such use. Patients sharing recordings Apps such as Second Ears facilitate patients’ sharing of recordings with family and others for treatment decision making and care. The law relating to such sharing of recordings with third parties varies between jurisdictions and also turns upon the question of whether the original recording was overt or covert. Separate legislative provisions address the act of recording compared with the recordings’ subsequent use. Two‐party consent is generally, but not always, required for patients to lawfully share recordings with third parties (Box 2). In Queensland, Tasmania and the ACT, there is a distinction between patients sharing a recording with immediate family (which can be done without the clinician's consent to share) and sharing with the wider world (which requires the clinician's consent). In NSW, unusually, a recording that is originally lawfully made with only one party's consent but with no intention to share can be subsequently shared without restriction (eg, on social media) (Surveillance Devices Act 2007 (NSW), section 11). Clear communication and consent remain the most desirable mechanisms to frame patients’ expectations and choices around the sharing of recordings with others, even where consent is not legally required. For the avoidance of doubt, an agreement to create a recording — whether a clinician's oral agreement for a patient to record on their smartphone, or the terms and conditions built into an app — should explicitly address the extent to which a patient can share the recording with others. Such an agreement might, for instance, permit the patient to share the recording with family but not publish it at large, for example, on public social media. This could override any legislative entitlement to share a recording openly. If a patient distributed the recording in violation of the terms and conditions, the health service could pursue a legal claim for breach of contract. We are not aware of previous such claims. Health services would need to weigh up the financial and reputational costs of pursuing such a claim. The use of recordings in legal proceedings Recording the consultation does not change clinicians’ medico‐legal obligations to patients. Such recordings provide transparency of the discussion and could be used as evidence of appropriate information sharing with patients, thus meeting the clinician's required standard of care. Clinicians have a duty to provide sufficient information on inherent risks of treatment and alternative treatments, to enable patients to exercise a meaningful choice. A claim may lie in negligence if the patient can demonstrate a “failure to warn”, where the clinician did not meet the appropriate standard of care and the patient consequently made an uninformed choice about treatment which resulted in harm. The importance of patient‐centred communication was highlighted in the UK decision of Montgomery v Lanarkshire [2015] UKSC 11 and the Australian case Rogers v Whitaker [1992] HCA 58. In a claim for negligent non‐disclosure, where the patient states that the clinician did not provide information concerning material risks about the proposed procedure, the recording could be used to provide evidence of the consultation. In most states and territories, whether the recording itself was taken with both parties’ consent or by one party covertly does not affect its admissibility in court. In jurisdictions where covert recording is not lawful (Box 1), an exception typically exists permitting a person to covertly record a private conversation to protect their lawful interests. An example is where there is a serious dispute between two parties regarding different versions of an arrangement (Georgiou Building v Perrinepod [2012] WASC 72). The relevant lawful interest must exist at the time of the recording (Marsden v Amalgamated Television Services [2000] NSWSC 465). The recording's lawfulness is a separate issue to its admissibility. It has been established that tape recordings are admissible to provide primary evidence of the conversation or sounds recorded on the tape. In the case of Butera v Director of Public Prosecutions (Vic) [1987] HCA 58, it was held that the tape is “a part of the machinery by which the evidence is produced”. It would follow that the recording on an app such as Second Ears provides evidence of the conversation that took place between the clinician and patient. Such a recording is admissible in court if the content is relevant and otherwise admissible, the voices are properly identified, and the recording has provenance — it is authentic, accurate and has not been tampered with. In this instance, the voices recorded would fall within the category of hearsay evidence — that is, representations made out of court that are led as evidence of the truth of the fact. As audio recordings fall within the definition of “document” in the Evidence Act 1995 (Cth) (which is uniform with most state and territory Acts), they may be admissible if they conform to the statutory requirements. As an example, in Victoria courts have the discretion to admit recordings as evidence if the evidence is relevant (Evidence Act 2008 (Vic), sections 55 and 56) and if the desirability of admitting the evidence outweighs the undesirability of doing so (Evidence Act, section 138). The recording will form only part of the record of information flow between clinician and patient. Contemporaneous notes and other non‐recorded conversations will also be relevant to determine if the standard of care has been met. There is no evidence that audio or video recordings of consultations increase litigation.19,20 A study evaluating the provision of consultation video recordings to patients found that in the high risk specialty of neurosurgery, none of the 2807 patients recorded used the video in a legal action.19 Recordings might actually reduce conflict and litigation because they overcome differences in recollection between two parties.21 Ownership of recordings Traditionally, the law has not conceived of information as property (Boardman v Phipps [1967] 2 AC 46). In Australia, patients have no proprietary interest in a doctor's medical notes (Breen v Williams [1996] HCA 57) (although legislation provides a right to access them). Nor do doctors have any proprietary interest in a patient's handwritten notes, or by extension, an overt patient‐led recording. However, a health service‐led recording such as one made using the Second Ears app could be said to be jointly created. As there are two copies of it, one held by the patient and one by the health service, it could be argued that each has some proprietary interest. A recent exploration of this position posited that there may be multiple rights holders of health data.22 This view has yet to be tested in the courts. It is appropriate to focus instead on the obligations of the different parties to protect and store the recording data. Data security and storage of overt recordings A recording made on a system such as Second Ears forms part of the medical record and the organisation must take reasonable steps to protect it from misuse, loss and unauthorised access or disclosure. Any contract with a third‐party organisation (eg, a cloud storage provider) should also reflect these requirements and address issues of security and access. Health records must be retained for a specified period; in Victoria, NSW and the ACT, this is 7 years after the patient last received care from the organisation, after which the records should be destroyed if they are no longer needed. By comparison, patients need neither keep nor protect their own copy of a recording. If the recording is made using a third‐party app, the terms and conditions of that app are relevant, adding further complexity in relation to custodianship and data protection. Conclusion Health service‐led recording technologies, of which Second Ears is an example, can draw on a framework that makes explicit all parties’ rights and responsibilities, and ensure that an authenticated version of the recording is maintained securely. Such an approach promotes shared expectations between patients and clinicians and is likely to reduce miscommunication. Our analysis found surprising diversity in Australian legislation pertaining to consultation recording, leading us to conclude that, to avoid confusion, expressly articulated permissions around the act of recording and the extent of sharing recordings are desirable. While covert recording is not uniformly unlawful in Australia, transparency promotes trust and enhances the clinician–patient relationship. There is some evidence that concerns about a heightened litigation risk as a consequence of recording are unfounded; rather, the existence of a recording should minimise conflicting recollections and enhance a sense of collaboration. While the act of recording does not alter a clinician's duty to disclose relevant information to a patient, communication skills training may be a way to alleviate concerns about being recorded.10 Box 1 – Patient‐led recordings: when is consent from the other party required for the act of recording? Jurisdiction Patient makes recording for unspecified purpose Patient makes recording intending it for personal use only Patient makes recording that is reasonably necessary for the protection of their own lawful interests Legislation Victoria, Queensland, Northern Territory Consent not required Consent not required Consent not required Surveillance Devices Act 1999 (Vic): no relevant provision Invasion of Privacy Act 1971 (Qld), s 43(2)(a) Surveillance Devices Act 2007 (NT): no relevant provision New South Wales, Australian Capital Territory, Tasmania Consent required Consent not required Consent not required Surveillance Devices Act 2007 (NSW), s 7(3) Listening Devices Act 1992 (ACT), s 4(1)(b), (3) Listening Devices Act 1991 (Tas), s 5(1)(b), (3)(b) South Australia, Western Australia Consent required Consent required Consent not required Surveillance Devices Act 2016 (SA), s 4 Surveillance Devices Act 1998 (WA), s 5 Box 2 – Can a patient share their lawfully made recording with third parties for general purposes* without the clinician's consent for the sharing? Jurisdiction Sharing with immediate family and friends† Sharing with public at large Legislation Victoria, Northern Territory No (clinician consent for sharing required) No (clinician consent for sharing required) Surveillance Devices Act 1999 (Vic), s 11(2)(a) Surveillance Devices Act 2007 (NT), s 15(2)(a) Western Australia No (clinician consent for sharing required) No (not even with clinician consent) Surveillance Devices Act 1998 (WA), s 9(2)(a)(ii), (3) Queensland, Tasmania, Australian Capital Territory Yes‡ No (clinician consent for sharing required) Invasion of Privacy Act 1971 (Qld), s 45(2)(a), (d) Listening Devices Act 1991 (Tas), s 10(2)(a), (d) Listening Devices Act 1992 (ACT), s 5(2)(b), (e) New South Wales, South Australia Yes§ Yes§ Surveillance Devices Act 2007 (NSW), ss 7(3)(b), 11(1). Surveillance Devices Act 2016 (SA), ss 4(2)(a)(i), 12(1). * Legislation usually deals separately with the sharing of recordings for different purposes, such as “in the public interest”, for protecting the “lawful interests” of the person who is sharing the recording, “in the course of legal proceedings”, “in the performance of a duty”, or as authorised by law. This table solely addresses when clinician consent is required for the sharing of a recording with a family member or with the public at large when the purpose of the sharing is not specified. This may include for the patient's health and wellbeing. It does not address sharing for other purposes. † This is typically expressed in legislation as: persons who have, or are believed on reasonable grounds by the person who is communicating or publishing the recording to have, such an interest in the private conversation (ie, the health care consultation) as to make the sharing reasonable under the circumstances. ‡ In these jurisdictions, the original recording may be lawfully made covertly by the patient for their own use, and then shared with family, without the clinician's consent. § Section 11 of the Surveillance Devices Act 2007 (NSW) is silent about the sharing (publication or communication) of recordings that were made lawfully. A recording that is made by one party without an original intention that the recording be published or otherwise disseminated is lawful in NSW: section 7(3)(b)(ii). Section 12 of the Surveillance Devices Act 2016 (SA) is silent about the sharing of recordings that were made lawfully, such as a recording made with the consent of both parties under section 4(2)(a)(i).

Megan Prictor · Carolyn Johnston · Amelia Hyatt

As we face a second wave of the pandemic, mobility data may assist government public health action

Lucinda Adams · Robert J Adams · Tarun Bastiampillai

To the Editor: Torda and colleagues1 highlight the impact of the coronavirus disease 2019 (COVID‐19) pandemic on medical education, which has prompted the rapid shift to online teaching for medical students. We need to ensure that these recent changes in medical education are thoughtfully blended with the reintroduction of face‐to‐face teaching when it occurs. Before integrating these changes, it is critical we reflect and review three key elements: Preparing students: blended learning, where online learning is combined with traditional face‐to-face teaching, is likely to capture more students’ learning styles but is also often associated with increased need for self‐directed learning, which may mainly benefit high achieving students.2,3 It is critical we equip all our students to engage effectively in adult learning to maximise the benefits of blended learning and develop engaged independent learners.4 This is an opportunity to develop these skills by ensuring that staged and increasing self‐directedness is built into new material and forms of delivery.5 Preparing educators: the attitude and preparedness of educators running or engaging in online education is crucial. As vital stakeholders, lecturers should be seen as educators and be supported and developed as such, including the training in both design and delivery of online learning.6 Preparing delivery and its content: facing the option of moving material back from online learning to face‐to-face learning, each move must be critically analysed to determine what is the most effective form of delivery. Historical modes of delivery need not be the default. Indeed, we have been given a once in a lifetime opportunity for a major, if incidental, review of each part of the curriculum and the best way it can be delivered — online, face‐to-face, or maybe a mix of both. As the mode of delivery is reviewed, the content can be refined and tailored for the students’ needs. Many of us know the deafening and discouraging silence when students are quiet in response to a question, both face‐to‐face and online. However, we are at a turning point in medical education where we must take the time to reflect and move forward with excitement regarding what has worked, and have the courage to leave behind what has not.

Lucy E Kirk · Imogen Mitchell

The advent of CAR T‐cell therapy has seen significant improvements in survival and is a potential cure for patients with advanced haematological malignancies Cancer immunotherapy is a burgeoning field which, in the last decade, has produced unprecedented improvements in outcomes across a variety of advanced malignancies. The eventual translation of decades of research into clinically available immunotherapies stems from the expanded knowledge of the role that the immune system plays in preventing tumour initiation and progression as well as the mechanisms by which tumours learn to evade this immune surveillance.1 Immunotherapies that have reached the clinic include monoclonal antibodies and, more recently, their augmented counterparts including antibody–drug conjugates and bi‐specific T‐cell engagers. Other treatments are immunomodulatory, meaning that they augment endogenous anti‐tumour immune activity. These include immune checkpoint inhibitors such as pembrolizumab, which are prolonging survival in melanoma and several solid organ malignancies as well as relapsed or refractory Hodgkin lymphoma. Cellular immunotherapies offer the potential to overcome immune tolerance and generate immune memory.1 Allogeneic stem cell transplantation (ASCT), a largely unmanipulated form of cellular immunotherapy, acts by completely replacing the recipient’s entire haematopoeitic and immune systems, leveraging differences between the recipient and donor to produce a graft‐versus‐tumour effect, with the potential negative consequence of immune attack on recipient’s normal tissues, known as graft‐versus‐host disease, as well as other serious toxicities. ASCT has been the only curative option for many patients with haematological malignancies. However, it is generally considered a consolidative therapy; that is, the patient’s malignancy must be in or near complete remission in order to be effective. This is not always possible in refractory cases. For others, ASCT may be contraindicated because of age or comorbidities. With advances in genetic manipulation technology, the notion of combining the specificity of a monoclonal antibody with the cytotoxicity and memory of a T‐cell came to fruition in the chimeric antigen receptor (CAR) T‐cell. “Chimeric” here means that the DNA comes from two or more sources; the antigen‐binding domain of the CAR construct is an antibody fragment, tethered to the intracellular signalling domain of the T‐cell receptor, with an additional co‐stimulatory domain acting to improve their expansion and persistence in vivo. The fundamental steps in generating and delivering CAR T‐cell therapy are summarised in Box 1.2 Specific toxicities are characteristic of CAR T‐cell therapy, the two most important being cytokine release syndrome and neurotoxicity. Cytokine release syndrome is an inflammatory state induced by the rapid proliferation of CAR T‐cells and tumour cell death, releasing an array of inflammatory cytokines. The hallmark is a fever, with the potential for hypotension, hypoxia and organ dysfunction. As one of the key cytokines driving the syndrome is interleukin‐6, its blockade using the interleukin‐6 receptor antagonist tocilizumab is now routinely used for more severe grades of cytokine release syndrome. The pathogenesis of neurotoxicity has not been fully elucidated; however, it most often manifests with speech disturbance or aphasia, dysgraphia and attention deficits, with more severe manifestations including altered level of consciousness, seizures and, rarely, cerebral oedema. Fortunately, even patients with severe neurotoxicity who are adequately supported in intensive care settings most often have complete neurological recovery. By far the most successful antigen target of all CAR T‐cell therapies developed to date is the pan‐B‐cell antigen CD19, as it arguably comes closest to the characteristics of an ideal target. CD19 is widely expressed across the full maturation spectrum of B‐cell malignancies, from B‐cell lymphoblastic leukaemia cells to mature B‐cell lymphomas, giving broad applicability. Second, CD19 is only expressed on B‐cells (normal and malignant) and not other tissues. Third, the toxicity resulting from the on‐target, off‐tumour effects, in this case normal B‐cell aplasia, is manageable by immunoglobulin replacement in patients who experience recurrent or severe infections. The decision for health authorities to fund a personalised, genetically engineered treatment is a complex one, taking into account considerations such as cost, efficacy, safety, the level of evidence and the maturity of outcome data, alternative therapies, equity of access, and resource utilisation. The cost of a single product is measured in hundreds of thousands of dollars and the mechanism by which such therapies will be funded is certainly not self‐evident. In the Australian context, the new therapy is evaluated by the Medical Services Advisory Committee, an independent committee that appraises new medical services proposed for public funding, taking into account all the above‐mentioned considerations, and providing advice to government. Moreover, given the high cost and limited, immature data, regulatory bodies worldwide have come to unprecedented outcomes‐based reimbursement agreements with pharmaceutical companies — such as rebates and staged payments according to defined response criteria — in order to mitigate risk. Two CAR T‐cell products targeting CD19 were approved by the United States Food and Drug Administration in 2017 and 2018: tisagenlecleucel and axicabtagene ciloleucel. The landmark studies which led to their approval, and a summary of their key outcomes, are shown in Box 2.3,4,5,6 In Australia, tisagenlecleucel is approved by the Therapeutic Goods Administration for paediatric and young adult patients up to 25 years of age with B‐cell lymphoblastic leukaemia that is refractory, in relapse after transplant or in second or later relapse, as well as adult patients with relapsed or refractory diffuse large B‐cell lymphoma after two or more lines of systemic therapy. In April 2019, a joint state and federal government funding initiative commenced for tisagenlecleucel for the B‐cell lymphoblastic leukaemia indication, and in January 2020, the government announced its funding for diffuse large B‐cell lymphoma. Soon after, the Therapeutic Goods Administration approved axicabtagene ciloleucel in February 2020 and the Medical Services Advisory Committee made a positive recommendation for its public funding for the lymphoma indication. Further, Novartis announced an agreement with an Australian cell and gene therapy manufacturing company for manufacture of tisagenlecleucel for the region to commence in late 2020 (https://www.celltherapies.com.au/kymriah-to-be-manufactured-at-cell-therapies-pty-ltd-marking-australias-first-on-shore-commercial-production-of-car-t-therapy/). In the case of tisagenlecleucel for relapsed or refractory B‐cell lymphoblastic leukaemia, evaluation began with a comparison with best available therapy. In the ELIANA trial4 outcomes compared very favourably with other chemo‐ or immunotherapeutic salvage options such as clofarabine7 and blinatumumab,8 respectively. For example, blinatumomab, a bispecific T‐cell engager, in the paediatric relapsed or refractory setting produced a complete remission rate of 39% within the first two cycles, with a relapse‐free survival at 6 months of 42%, and this therapy is considered to be a bridging therapy to ASCT. Tisagenlecleucel on the other hand can be used as a stand‐alone therapy; however, it is notable that a substantial proportion of responders do relapse, particularly between 6 and 12 months, which raises the question of whether this treatment should also serve as a bridge to ASCT. The available evidence is currently insufficient to confidently provide an answer, and practice therefore varies among treating centres worldwide. However, a major concern is the financial implications of CAR T‐cell therapy as a bridge to ASCT, which itself is a highly resource‐intensive therapy, with some suggestion that the cost‐effectiveness may be unbalanced if this practice were routine. In the case of high grade B‐cell lymphomas, patient outcomes also appear to be superior to other available treatments in the third line setting. In the ZUMA‐1 trial, the recently updated 3‐year overall survival rate of 47% does likely reflect a significant cure fraction.6 In comparison, the SCHOLAR‐1 retrospective study of the outcomes of patients with refractory diffuse large B‐cell lymphoma showed that this pooled patient population only achieved complete remission rates of 7% with conventional therapies and had a median overall survival of 6.3 months.9 One concern is that the outcomes of the CAR T‐cell trials may not be generalisable to the real‐world population where patient selection may not be as strict as in clinical trials. Interestingly, the real‐world data seems to be conflicted on this, with the US experience from the Center for International Blood and Marrow Transplant Research registry being comparable to trial data for both tisagenlecleucel and axicabtagene ciloleucel, while preliminary United Kingdom experience appears to be considerably worse.10,11,12 The cause of this discrepancy is unclear. In terms of future directions, many clinical trials are assessing CAR T‐cells in earlier lines of therapy. For example, two trials are randomising patients in first relapse of large B‐cell lymphoma to receive either CAR T‐cell therapy or standard salvage plus autologous stem cell transplant: ZUMA‐7 (NCT03391466) and BELINDA (NCT03570892). The results of these trials, if favourable, could greatly alter treatment paradigms. Other trials are assessing CAR T‐cells in other B‐cell lymphomas, such as follicular non‐Hodgkin lymphoma (ELARA [NCT03568461]) and mantle cell lymphoma.13 The response to KTE‐X19, an anti‐CD19 CAR T‐cell therapy with a manufacturing process that removes circulating tumour cells, seen in the ZUMA‐2 trial in relapsed or refractory mantle cell lymphoma (overall response rate of 93%) is the highest reported response rate in patients with mantle cell lymphoma who failed previous BTK inhibitor treatment, with a high proportion of durable responses in this very challenging malignancy.13 Strategies to improve the availability and timeliness of CAR T‐cell therapy include the development of third party allogeneic CAR T‐cells, which could produce off‐the‐shelf treatments for many patients. Other alterations to the CAR construct aim to improve characteristics such as persistence and safety, as well as addressing the problem of antigen escape, where the malignancy loses the targeted antigen, potentially through multi‐antigen targeting. Others are combining CAR T‐cells with immunomodulatory therapies such as immune checkpoint inhibition to improve efficacy. Finally, there is great interest in CAR T‐cell therapies for malignancies such as multiple myeloma, acute myeloid leukaemia and T‐cell lymphomas and leukaemias, many of which are at various phases of clinical trials. The furthest advanced are CAR T‐cell therapies targeting B‐cell maturation antigen in multiple myeloma. JNJ‐4528, an investigational B‐cell maturation antigen CAR T‐cell therapy, has recently demonstrated very high response rates in the phase 1b/2 CARTITUDE‐1 study in relapsed or refractory myeloma.14 In the 29‐patient cohort, the overall response rate was 100%, with 69% complete remission, the median time to complete remission being 1 month, and measurable residual disease negativity in all 17 evaluable patients. These are very promising times in cancer immunotherapy and the task ahead for regulatory authorities will be immense as evidence rapidly accumulates for these high cost therapies. In the meantime, we are pleased to add CD19 CAR T‐cell therapy to our armamentarium and await the results of trials across a wide range of haematological and solid organ malignancies. Box 1 – Overview of the processes for manufacture and delivery of a chimeric antigen receptor (CAR) T‐cell product Procurement of T-cells, usually via leukapheresis (1); transduction of the CAR genes via viral vector or non-viral methods (2); ex vivo expansion of the CAR T-cells (3); preconditioning with lymphodepleting chemotherapy (4); and infusion into a patient (5). ◆ Box 2 – Summary of data from pivotal CD19 chimeric antigen receptor (CAR) T‐cell trials Trial name CAR T‐cell product Disease Complete response rate Other response parameters Safety ELIANA4 Tisagenlecleucel Relapsed or refractory paediatric B‐ALL 81% 12‐month OS, 76%; 12‐month EFS, 50% Grade ≥ 3 CRS, 47% Grade ≥ 3 NT, 13% JULIET5 Tisagenlecleucel Relapsed or refractory DLBCL 38% Median OS, 12 months Grade ≥ 3 CRS, 23% Grade ≥ 3 NT, 11% ZUMA‐13,6 Axicabtagene ciloleucel Relapsed or refractory DLBCL 58% 3‐year OS, 47% Grade ≥ 3 CRS, 13% Grade ≥ 3 NT, 28% B‐ALL = B‐cell acute lymphoblastic leukaemia; CRS = cytokine release syndrome; DLBCL = diffuse large B‐cell lymphoma; EFS = event‐free survival; NT = neurotoxicity; OS = overall survival.

Adrian G Selim · Constantine S Tam

AI has arrived, with the potential for enormous change in the delivery of health care, but are we ready? Artificial intelligence (AI) is the trigger for the next great transformation of society: the fifth Industrial Revolution. AI has already arrived in health care, but are we ready for the kind of changes that it will introduce? In this article, we map out the current areas where AI has begun to permeate and make predictions about the kind of changes it will make to health care. Definition of AI AI comprises any digital system “that mimics human reasoning capabilities, including pattern recognition, abstract reasoning and planning”.1 It includes the concept of machine learning, where machines are able to learn from experience in ways that mimic human behaviour, but with the ability to assimilate much more data and with potential for greater accuracy and speed. Machine learning is a research field that has seen recent advances due to exponential increases in computing power (a phenomenon known as Moore's law), algorithmic coding that mimics the human cognitive process (deep learning), and access to large, linked sources of big data. The scope of AI can be specific, performing narrowly defined tasks (narrow AI) such as image interpretation, or more general, applying knowledge and skills in different contexts (general AI) such as making a diagnosis and predicting disease outcome. On the other hand, machine learning can also be designated “supervised”, in which a dataset is provided for the algorithm to evaluate its performance, or “unsupervised”, in which the machine is allowed to extract unknown potential features in developing an algorithm. The arrival of AI into current practice AI, machine learning, and deep neural network tools can assist medical decision making and management, and have already permeated into at least three different levels: AI‐assisted image interpretation; AI‐assisted diagnosis; and AI‐assisted prediction and prognostication. From diagnosing retinopathy to cardiac arrhythmias, from screening for skin cancer to breast cancer, from predicting outcome of stroke to self‐management of chronic diseases, AI and machine learning devices can replace many time‐consuming, labour‐intensive, repetitive and mundane tasks of clinicians and give possible suggestions of management plans (Box 1).2,3,4,5,6,7 While the advancement and new capabilities and opportunities are exciting, the responsibility and liability issues of AI‐assisted clinical diagnosis and management need much deliberation. AI‐assisted image interpretation One of the major advances in AI is pattern recognition enhancing image‐based diagnosis in radiology, pathology and endoscopy. AI‐assisted image analysis aids the detection of adenoma and polyps during colonoscopy. It can even provide optical biopsy to determine the nature of lesions with implications of treatment.8 Wireless capsule endoscopy is a groundbreaking advance in medical technology, allowing painless examination of the gut, reaching areas where conventional endoscopes cannot reach. However, reading thousands of images produced by the capsule is extremely time‐consuming. Deep neural network systems trained to read images of capsule endoscopy can scan thousands of pictures within minutes to reduce the burden of time and energy for gastroenterologists and also minimise the chance of missing significant lesions.9 Similarly, systems have been trained to read echocardiographic images to provide physiological measurements within seconds, and to read coronary computed tomography angiography images to determine coronary calcification, coronary stenosis severity, and functional haemodynamic effects of the stenosis. AI‐assisted diagnosis The diagnosis of many conditions (eg, acute or old myocardial infarction) and arrhythmias (eg, atrial fibrillation and ventricular tachycardia) can be made by experts reading electrocardiograms (ECGs) according to well established rules. Application of such rules in algorithms have allowed computers to make these diagnoses automatically in ECG machines for many years, but the diagnoses are subject to verification by physicians using the same rules. AI using machine learning and deep neural network can do the same from raw ECG data, but does not rely on the same rules, and thus can do much more than conventional ECG analysis. In the most basic AI formulation, diagnoses of important cardiac arrhythmias from a single lead rhythm strip or continuous single lead ECG recordings were made by machine learning using deep neural network algorithms with greater accuracy than an individual cardiologist and similar to a consensus panel of cardiologists.10 Where AI excels, however, is in discerning patterns not apparent to the experts, such that current or future paroxysmal atrial fibrillation can be diagnosed from an ECG in sinus rhythm,11 and asymptomatic left ventricular dysfunction can be diagnosed by a 12‐lead ECG.12 AI‐assisted prediction and prognostication AI may predict the occurrence of certain diagnoses and prognosticate clinical outcomes of patients based on clinical datasets, genomic information and medical images. Cardiologists have developed algorithms to assess the risk of cardiovascular disease and claimed that their prediction is superior to existing scoring systems. Gastroenterologists have also developed AI models to predict recurrence of bleeding and requirement of surgery in patients with gastrointestinal bleeding.13 Combining genomic, epigenetic and metagenomic data with biochemical and lifestyle information using machine learning will be a very powerful tool in medicine. However, mechanisms or reasons for reaching the machine decision may not be comprehensible to clinicians. The integration of various datasets in multilayer informatics could take prediction, prognostication and prevention of diseases to new levels that cannot be achieved by conventional statistical models. This capability, if validated in properly designed studies, will offer new dimensions to personalised medicine. Preparing for the future of AI Health disparities, excluded populations and data biases The quality of AI in health care is dependent on the quality of the data on which it is based. Algorithms are being developed and validated on data generated by health care systems where current practices may already be inequitable. A system built on poor quality, biased data will reflect those problems (“garbage in, garbage out”). If a health care system has excluded populations of patients, the structural inequalities of health care will be repeatedly reinforced by the AI. This is a not a new problem and we must do better science and be awake to the limits of data quality and evidence‐based medicine. Data sovereignty and stewardship AI is built on access to big data. Big data in health care is primarily generated by public health systems, funded by the public for the public. Increasingly, claims over the health data generated by these public systems are being contested.14 There was enormous public outcry over the use of British National Health System data by Google‐owned DeepMind, a company creating an AI‐based smartphone application for kidney disease. Many were angry about the private use of public data, when there was little public control over what would happen to the data or what benefit was being provided back to the National Health System.14 Issues of data sovereignty therefore threaten the existence of effective AI. Patient data should not be provided to technology giants without a good governance structure to protect data sovereignty. Changing standards of care An immediate issue for the use of AI is the question of how it will transform standards of care. Common law jurisdictions judge health professions, by and large, by measuring performance against competent professional practice as set by the professions themselves. If AI keeps its promise of benefit and it is integrated more into practice, standards of care must require AI use, and traditional forms of therapeutics will be forced to change. We will see a time when all medicine and allied health work as a team with AI. Those who refuse to partner with AI might be replaced by it. Legal responsibility for AI‐caused injury AI promises to massively reduce the occurrence of iatrogenic harms via increasing the quality of decision making, but the continued existence of AI‐related injury is easy to foresee. As machine algorithms improve themselves without human intervention, making the “black box” more opaque, regulatory agencies such as the Australian Therapeutic Goods Administration and the United States Food and Drug Administration need to refine their regulations. To the extent that AI continues to play a role in assisting clinical management, questions of responsibility for harm should be determined by ordinary rules of product liability. It is likely that courts will determine some of these liability questions by using analogies with vicarious liability — an employer is responsible for the negligence of the staff when the injury occurs in the course of the staff's employment. A doctor using AI should be responsible for AI decisions made in the course of treatment, especially if the doctor retains the power to make the final decision regarding treatment. But as AI takes on more autonomous decision making, it might be argued by some doctors that they should not be responsible for that which they cannot control. Similarly, it seems unfair for doctors to be held responsible for an AI decision when they are unable to deduce how and why that decision was made. Such matters are outside the scope of clinicians’ expertise and best dealt with legally as a product liability claim. A stepwise gradation model of shared responsibility between the human doctor and the machine in diagnosis and clinical management has been proposed15 (Box 2). Conclusions Before AI tools can be put into daily use in medicine, data quality and ownership, transparency in governance, trust‐building in black box medicine, and legal responsibility for mishaps are some of the hurdles that need to be resolved. Much effort is needed to translate algorithms into problem solving tools in clinical settings and demonstrate improvement in clinical outcomes with saving of resources. Box 1 – Examples of artificial intelligence (AI) permeation into clinical practices of different specialties Clinical management AI capability Diabetic retinopathy2 Detection of early changes in fundi of patients with diabetes Reading the retina and blood vessels to identify patients at risk of developing complicated diabetic retinal disease Breast cancer3 Diagnosis of early breast cancer based on mammography Reading mammographic pictures to detect early malignant transformation in breast cancer screening Skin cancer4 Diagnosis of skin cancer by its clinical morphology Identification of skin cancer by pictures and classification of types of skin neoplasia Cerebrovascular disease5 Predicting outcome after a cerebrovascular accident Predicting the outcome (mobility, morbidity and mortality) of stroke 90 days after the event Non‐communicable chronic diseases6 Monitoring of diabetes and heart failure in primary care setting Assisting patients monitoring of blood pressure and blood glucose at home and transmitting information to family medicine clinics Heart failure7 Predicting the clinical outcome of patients with heart failure Predicting in‐hospital mortality among patients with heart disease based on echocardiography Box 2 – Levels of artificial intelligence (AI)‐assisted decision in diagnosis and clinical management and possible share of responsibility between human doctor and machine

Joseph JY Sung · Cameron L Stewart · Ben Freedman