Can <scp>AI</scp> help in the fight against <scp>COVID</scp> ‐19?

Abstract

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 The corolnavirus disease 2019 (COVID-19) pandemic has accelerated efforts to incorporate artificial intelligence (AI) into clinical care at a time when, in many countries, health care systems are facing unprecedented strain on their resources. Before COVID-19, AI was already permeating into health care,1 and reviews are emerging of how it may assist in efforts to combat the current pandemic.2-5 We describe several applications of AI relevant to COVID-19, some having had immediate clinical application, others awaiting further refinement and evaluation. The AI-automated HealthMap system at Boston Children's Hospital first alerted the world about the novel coronavirus on 30 December 2019, with a Canadian-based AI model, BlueDot, issuing a similar alert a day later. Researchers warned of the top 20 destination cities for passenger arrivals from Wuhan to which the disease could spread.6 These AI-enabled early warning systems use natural language processing to scan social media, online news articles and government reports for signs of emerging pandemics to help inform governments and agencies such as the World Health Organization. AI-assisted analysis and modelling have also helped reconstruct the progression of an outbreak, elucidate transmission pathways, identify and trace contacts, and determine real or expected impacts of various public health control measures (Box 1).7-11 How data are collected, and how these algorithms are deployed, raise difficult issues of consent, privacy, ethics and trade-offs between public and private good. Some countries, like Taiwan, have mandated a top-down approach to data harvesting. Others, including Australia, encourage individuals to voluntarily download apps to input symptoms and COVID-19 status and permit health authorities to access this information in identifying potential contacts. However, the efficacy of app-mediated contact tracing depends on the level of population uptake, its ability to accurately detect infectious contacts, and the extent of adherence to self-isolation by notified contacts.12 Expert position statements regarding design, scope, security and usage of such apps aim to prevent mission creep towards unauthorised surveillance of society at large.13 Detecting COVID-19 in most health systems currently involves testing symptomatic patients presenting to stand-alone fever clinics, general practices or emergency departments. This takes time, consumes personal protective equipment and testing reagents, and poses transmission risk to staff. Digital symptom checkers soliciting information about symptoms and risk factors may screen out individuals with very low likelihood of COVID-19 who do not require testing. In a pre-clinical study using hypothetical cases, an AI-powered chatbot identified patients with COVID-19 with sensitivity, specificity and overall diagnostic accuracy of 97%, 96% and 96%, respectively.1 However, a side-by-side test of eight different chat boxes on the same set of symptoms produced conflicting results,15 suggesting the need to identify all the data elements necessary for highest accuracy. Data from phone hotlines used to pre-screen individuals based on travel history and symptoms,16 and from sensors (cameras, microphones, temperature and inertial sensors) embedded within smartphones, can all be used to detect COVID-19.17 Neural networks embedded in cameras can distinguish patterns of tachypnoea caused by COVID-19 from those caused by influenza or the common cold.18 AI-powered thermal-scanning face cameras, capable of screening up to 200 people per minute, are being used by some Australian private hospitals to remotely detect people with fevers, sweating and discolouration, and prevent them entering public spaces.19 Diagnosing COVID-19 in sick patients presenting to hospital is currently performed by reverse transcription polymerase chain reaction testing of nasopharyngeal and throat swabs. However, initial tests may only be 70% sensitive and turnaround times can be 24 hours or more.20 Machine learning (ML) models combined with virus detection systems using CRISPR (clustered regularly interspaced short palindromic repeat; a tool which uses an enzyme to edit genomes by cleaving specific strands of genetic code) can rapidly design severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) assays which have high sensitivity and speed.21 AI has also been used to design, within a few weeks, point-of-care immunoassays for detecting viral antigens within 20 minutes, and such testing kits are now in use.22 For hospitalisations where reverse transcription polymerase chain reaction testing is unavailable, untimely, or yields negative results among patients highly suspected of suffering COVID-19 pneumonitis, deep learning algorithms applied to imaging data captured by chest x-rays or computed tomography (CT) chest scans may help early diagnosis. Studies suggest that such cases show particular image patterns which may, combined with polymerase chain reaction testing, improve sensitivity to more than 90%.23, 24 However, radiological appearances of COVID-19 can overlap with other forms of lung inflammation, imaging can be insensitive or misleading in mild or asymptomatic cases,25 and CT scanning creates a transmission risk. Hence, the Royal Australian and New Zealand College of Radiology and peer societies overseas do not recommend CT scanning, including AI applications, to screen for COVID-19, or to diagnose it as a first choice test.26 For patients with lower respiratory tract illness who pose diagnostic uncertainty, deep learning algorithms applied to chest x-rays may be more feasible and impose less risk (Box 2).27-30 However, current algorithms may perform poorly on the 80% of COVID-19 cases which are mild and under-represented in the data used to train and test these algorithms.28-30 Predictive models able to identify, on admission, patients likely to deteriorate and require respiratory support can assist triage and resource allocation decisions. While older age, being male, and having certain comorbidities (hypertension, cardiovascular disease, diabetes) portend worse outcomes,31 these factors do not necessarily predict outcomes at an individual level, especially in younger patients. Some ML algorithms can more accurately estimate risk of death, development of acute respiratory distress syndrome, and duration of hospitalisation (Box 3).32-36 Patients diagnosed with COVID-19 but not requiring hospitalisation can be monitored remotely at home using wearable devices measuring temperature, blood pressure and arterial oxygen levels, and transmitting these data to central virtual care units,37 as exist in some Australian hospitals. AI-assisted analysis alerts staff to worsening status with activation of outreach care or patient recall for admission. In developing effective treatments for COVID-19, ML-based repurposing frameworks have used algorithms to identify baricitinib (for rheumatoid arthritis),38 atazanavir (an anti-human immunodeficiency virus drug)39 and afatinib (for lung cancer)40 as potential treatments. Deep learning-based algorithms have helped design six new molecules that could halt SARS-CoV-2 replication41 and identify ten promising agents from among 4895 drugs.42 Algorithms using natural language processing applied to the PubMed database have identified a poly-ADP-ribose polymerase 1 inhibitor (CVL218) as a potential candidate, currently undergoing clinical testing.43 In expediting vaccine development, a deep learning system predicted targetable protein structures of SARS-CoV-2 within weeks compared with the months normally taken using traditional experimental approaches.44 AI has identified viral protein epitopes most likely to be immunogenic but not cross-reacting with human proteins,45 while a reverse vaccinology tool integrated with ML has identified genes that code for potential epitopes.46 AI-powered knowledge graphs can interrogate thousands of research articles and public documents to link genetic and biological properties of virus-caused diseases with composition and actions of existing drugs. The COVID-19 Open Research Dataset contains over 29 000 articles about SARS-CoV-2 and other coronaviruses,47 and is linked with several ongoing ML projects that daily attract many questions from research teams worldwide. AI mapping tools can track hospital bed capacity, and location, number and utilisation of intensive care unit (ICU) and hospital beds across the United States.48 Another tool tracks numbers of ventilated patients in that country and uses modelling software to predict breaking points for health care networks,49 estimating a shortage of 9100 ICU beds and 115 000 non-ICU beds for routine care at the peak of the pandemic. At the frontline, autonomous AI robots can transport drugs around the hospital and disinfect patients and hospital areas by emitting ulraviolet light, reducing interpersonal contact and saving time for medical and ancillary staff.5 While AI and ML can support COVID-19 responses across various domains, most applications have not reached operational maturity. The speed of research means that many reports are preprints awaiting peer review, while still attracting media coverage and clinician adoption before proper evaluation. Most ML models have relied on Chinese data, limiting generalisability to other populations. Those trained on limited and unrepresentative data are susceptible to overfitting and can perform poorly on real-world datasets. Many diagnostic and prognostic ML models published to date are poorly reported, lack external validation, and have high risk of bias.50 Overcoming these constraints requires scalable approaches to data sharing. An international consortium is assembling electronic health data from over 96 countries for rapid visualisation of regional differences and global commonalities.51 Such endeavours require balancing data privacy with public health concerns, and collaboration between clinicians, data scientists and policy makers across international borders and between private and public sectors. It is still too early to know the extent to which AI will have an impact on the COVID-19 outbreak. Although its role may be limited during the present pandemic, it may certainly help with the next one. No relevant disclosures. Not commissioned; externally peer reviewed.

Topics & Concepts