Machine Learning Applications in Public Health: Improving Risk Prediction and Resource Allocation

Abstract

Abstract

As the amount of health data in the Pakistan has grown in size and diversity, from electronic health records (EHRs) and insurance claims to public health surveillance feeds and social and behavioral data, a key method being adopted to transform data into actionable predictions and operational decisions is machine learning (ML). This review brings together recently published literature that addresses two closely related applications: (i) risk prediction, such as chronic disease, cardiovascular risk, opioid overdose, and infectious disease outbreak risk prediction, and (ii) resource allocation, including hospital bed and intensive care unit (ICU) capacity planning, emergency department staffing, and pandemic-scale forecasting illustrated by the Centers for Disease Control and Prevention (CDC) COVID-19 Forecast Hub. The use of ensemble and deep learning models (random forests, gradient boosted trees, long short term memory networks) is demonstrated to exceed traditional statistical baseline models on these tasks with areas under the curve (AUC) often greater than 0.85. The review also identifies continued challenges regarding safe and equitable deployment of the algorithms, including that of algorithmic bias from proxy labels, poor model interpretation, disjointed data infrastructure, and privacy limitations in multi-institutional data sharing. Potential solutions to make systems more transparent and equitable are discussed, including explainable AI (XAI), federated learning, and social-determinants-of-health (SDOH)-aware modeling. The review finds that achieving the full public health benefits of ML will require as much, if not more, attention to fair testing of models for accuracy, interoperability in data management, and operationalizing models into clinical practice for public health decision-making as to improvements in predictive accuracy.