A key lesson that emerged during the COVID-19 pandemic was that leaders, such as policy makers and clinicians, did not have the information that they needed to guide complex decisions around management of the pandemic. One of the best approaches for addressing complex problems is the use of large datasets to make inferences. Major advantages of using artificial intelligence (AI) and machine learning (ML) in this scenario are the opportunity to leverage all the available data to make estimations and predictions that have a higher likelihood of being correct than when such decisions are made without data analysis.
In many industries, these techniques are widely used, because they are powerful and offer substantial business advantages. For example, when you buy something online from many companies, they are usually predicting what else you might like to purchase. However, the healthcare sector has lagged in using AI and ML for several reasons. Healthcare organizations have invested far less in information technologies than other industries. Much of the data were not available digitally until recently, and in general, these data are siloed, fragmented, and unstructured. Furthermore, privacy concerns are especially important, creating additional barriers to development of large, linked datasets.
Thus, when a cluster of pneumonia cases was identified in Wuhan, China and were eventually found to be caused by a novel coronavirus, not all public health leaders had data at their fingertips to manage the pathogen effectively. AI was, however, used to detect the outbreak early.1 When the virus began to spread rapidly, decision makers had to make choices about closing borders and shutting down businesses, and how to manage scarce resources like personal protective equipment and ventilators, using first principles of public health rather than extensive data modeling. In this pandemic, most such decisions were made using first principles.
To better understand how AI and ML could be used for pandemic preparedness and response, we performed a scoping review, which was recently published in npj Digital Medicine.2 We identified algorithms and tools that were developed in response to prior pandemics, the severe acute respiratory syndrome (SARS) global outbreak, and early in COVID-19, with implications for how we can prepare for the next pandemic. We searched the peer-reviewed, preprint, and grey literature to capture a wide range of AI applications. A key challenge of our work was synthesizing the vast body of literature. We addressed this by separating the analyses into a comprehensive review of applications that used ML techniques and a limited review of traditional modeling approaches.
Overall, we found relatively little use of ML tools in prior pandemics, and that many of the COVID-19 studies that did use AI were performed post-hoc with most using traditional approaches. Figure 1 shows an example of projections based on traditional modeling approaches that were used to inform public health decision making during COVID-19.
Figure 1. Example of projected hospital resource use for COVID-19 patients using traditional modeling approaches from the Institute for Health Metrics and Evaluation (Image courtesy of the University of Washington, available under Public License and used with permission).
However, highly complex ML models are necessary to solve difficult tasks like image interpretation, such as determining whether people are social distancing in public spaces (Figure 2).
Figure 2. Aura Vision solution for real-time monitoring of adherence to social distancing (Image courtesy of Aura Vision, used with permission).
Through our review, we identified multiple key use cases in which ML could be very useful in future pandemics (Figure 3).
Figure 3. Six key use cases for leveraging machine learning for pandemic preparedness and response.
Many innovative tools and algorithms were developed early in the COVID-19 pandemic, which if implemented could have helped to control the spread of the virus and improve clinical management of infections. However, the major rate-limiting step was data access. Some very valuable data sources were developed relatively rapidly, but they might have been more impactful if available as the pandemic was emerging.
Furthermore, Emergency Use Authorizations for safe and effective solutions were granted by the U.S. Food and Drug Administration (FDA) months after the COVID-19 pandemic started, limiting utility and uptake, which does not come as a surprise. Implementation of AI in healthcare has faced many challenges including regulatory approval; only 64 AI or ML-based medical devices and algorithms were FDA-approved before the start of the pandemic.3 However, steps are being taken to improve this process, including FDA’s release of the Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan in January 2021,4 which should help with deployment of these types of solutions more rapidly during the next pandemic.
We found that many of the use cases listed above were explored using historical data from past pandemics or the global SARS outbreak and in a small number of cases prompted continued research and development. Notably, MITRE developed a prototype for mining web content to identify outbreaks such as SARS,5 and BlueDot was created in response to the SARS global outbreak.1 Continued investment in AI for infectious disease outbreak detection by BlueDot led to the company being among the first to identify and report on the emergence of COVID-19.
Following the 2009 H1N1 pandemic influenza, investigators affiliated with the National Institutes of Health explored the use of ML to differentiate between H1N1 pneumonia and other infectious lung conditions using computed tomography.6 Further research and investment into these types of solutions could have led to more rapid implementation of tools to help with early detection of COVID-19. Likewise, with almost two decades between the global SARS outbreak and COVID-19, we could have learned a lot from exploring historical data using novel ML techniques to develop more accurate forecasting models.
Many sectors came together to enable data to be gathered to improve pandemic response, including both the technology and healthcare sectors. The Data Coalition was one such effort. IBM, which sponsored our work, was one of the members. Other efforts such as the Johns Hopkins Coronavirus Resource Center proved instrumental, as did Worldometer among many others. To effectively deal with the next pandemic, we need large databases which include many types of information, not simply case counts and mortality rates.
AI has great potential, but even a year into this pandemic, AI tools have not been widely employed to guide policy or clinical decision making. After COVID-19, we will have vast amounts of public health and clinical data. It will be critical to ensure that these data are curated and readily available to help inform management of the next pandemic. We also need to retain support for maintaining these databases and linkages going forward—as no one knows when the next pandemic will be, and it is easy to lose focus given the multiple competing priorities policymakers face. In the interim, the relevance and performance of leading-edge ML techniques should be explored for addressing the six key use cases identified through our review, and investments should be made into development of supporting systems, associated governance, and pre-trained models that could be quickly modified or fine-tuned for novel pathogens—since development, approval, and implementation all take time. If these critical steps are done, and leaders use these data to make evidence-based choices, our response to the next threat should be much more effective.
References:
- Stieg, C. How this Canadian start-up spotted coronavirus before everyone else knew about it. CNBC. https://www.cnbc.com/2020/03/03/bluedot-used-artificial-intelligence-to-predict-coronavirus-spread.html (2020).
- Syrowatka, A. et al. Leveraging artificial intelligence for pandemic preparedness and response: a scoping review to identify key use cases. npj Digit. Med. 4, 96 (2021).
- Benjamens, S., Dhunnoo, P. & Meskó, B. The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database. npj Digit. Med. 3, 118 (2020).
- Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. U.S. Food and Drug Administration. https://www.fda.gov/media/145022/download (2021).
- Damianos, L. et al. MiTAP for SARS detection. In Demonstration Papers at HLT-NAACL 2004, 13–16 (2004).
- Yao, J., Dwyer, A., Summers, R. M. & Mollura, D. J. Computer-aided diagnosis of pulmonary infections using texture analysis and support vector machine classification. Acad. Radiol. 18, 306–314 (2011).
Please sign in or register for FREE
If you are a registered user on Research Communities by Springer Nature, please sign in