Table 6 Commonly used data sources and types of machine learning suitable for each use case based on studies from the in-depth review about past pandemics, SARS, and COVID-19.
Key use case | Commonly used data sources | Suitable types of ML |
---|---|---|
Forecasting infectious disease dynamics and effects of interventions | Publicly available counts (e.g., Johns Hopkins COVID-19 map, Worldometer, World Health Organization), media reports, commercial publications, web searches (e.g., Google Trends), social media (e.g., Twitter), census data, population-level comorbidity statistics, data on outbreaks of similar pathogens | Augmenting traditional models: neural networks |
Data-driven ML: recurrent neural networks | ||
Surveillance and outbreak detection | Social media (e.g., Twitter), web searches, news reports, medical record data (structured and unstructured fields) | Text mining: natural language processing |
Classification: support vector machines, transformer neural networks | ||
Real-time monitoring of adherence to public health recommendations | Cameras in public spaces | Compliance with mandated quarantine: proprietary facial recognition |
Adherence to mask wearing, social distancing, and sanitation: proprietary computer vision | ||
Real-time detection of influenza-like illness | Cameras and sensors in public spaces | Interpretation of thermal imaging: neural networks, computer vision |
Triage and timely diagnosis of infections | Exposure history, medical record data (structured and unstructured fields, laboratory results, chest imaging) | Interpretation of chest imaging: convolutional neural networks |
Triage based on routinely collected medical record data: no standout ML | ||
Interpretation of unstructured clinical notes: transformer neural networks | ||
Prognosis of illness and response to treatment | Medical record data (structured and unstructured fields, laboratory results, chest imaging) | Based on chest imaging: combination of convolutional and recurrent neural networks |
Based on routinely collected medical record data: no standout ML | ||
Interpretation of unstructured clinical notes: natural language processing, neural networks |