Integrating Artificial Intelligence and Engineering Technologies for Advanced Public Health Risk Prediction and Intervention

Tariq Saeed Mian
Department of Information Systems, College of Computer Science and Engineering, Taibah University, Madinah Almunwarah, Saudi Arabia

Hisham Farooq Saeed
Department of GP-Emergency, Raza Hospital, Shaher Sultan, Muzaffargarh, Pakistan

Hibatullah Saeed
Faculty of Medicine, Medical University-Sofia, Sofia, Bulgaria

Evena Shallonia Fung
Faculty of Engineering and Quantity Surveying, INTI International University, Nilai, Malaysia

Abstract

Background: Public health agencies increasingly combine artificial intelligence (AI) with engineered data infrastructures automated laboratory reporting, hospital information systems, smartphone-derived mobility sensing and policy trackers—yet evidence on whether such integration improves early warning, and whether earlier warning could plausibly change outcomes, remains fragmented and is frequently undermined by temporal leakage and optimistic validation. This study aimed to quantify the incremental predictive value of integrated multi-stream surveillance for anticipating hospital surges and to estimate, in silico, the potential consequences of AI-triggered intervention timing.

Methods: We assembled a country-day panel for 26 countries (March 2020–June 2022; 20,115 country-days) from the Our World in Data COVID-19 database and Google Community Mobility Reports. The outcome was a 14-day-ahead hospital surge (?50% relative and ?10 per million absolute rise in 7-day mean COVID-19 hospital occupancy). Twenty-five features from seven data-stream groups were engineered using only information available at the forecast date, including a causal (past-only) effective reproduction number. Logistic regression, random forest, gradient boosting and a multilayer perceptron were trained on data before 1 July 2021 and evaluated prospectively on Delta and Omicron waves (July 2021–June 2022) against a trend-extrapolation rule. We assessed discrimination, calibration, net benefit, incremental value of each data stream, episode-level lead time and robustness. A stochastic SEIR–hospital model (2,000 Monte Carlo draws per scenario) then translated the empirical alert characteristics into intervention-timing outcomes.

Results: Test-period surge prevalence was 14.3%. The random forest achieved an AUROC of 0.924 (95% cluster-bootstrap CI 0.899–0.945) and AUPRC of 0.682 (0.596–0.760), versus 0.873 (0.846–0.897) and 0.561 for the trend rule; calibration was good after country-grouped recalibration (Brier 0.074; expected calibration error 0.031). Linear and neural models that performed well in cross-validation degraded markedly under temporal shift (AUROC 0.646 and 0.668). Adding case-surveillance signals to hospital-load features improved AUROC by 0.037 (0.016–0.058), whereas mobility, policy, immunity and structural streams added no detectable discrimination. At equal specificity (90.7%), the model detected 71.7% of surge-days versus 61.1% for the rule and flagged 86% versus 81% of 57 surge episodes. Substituting a retrospectively smoothed reproduction number inflated AUROC to 0.950, illustrating leakage risk. In simulation, acting on model alerts reduced median peak occupancy by 59.9% versus 39.1% for capacity-threshold (reactive) activation, at the cost of 18.9 additional intervention-days per wave and false activation in 14.5% of non-surge scenarios.

Conclusions: Integrating clinical-load and case-surveillance streams with robust machine learning produced well-calibrated, prospectively valid surge warnings, but engineered behavioural and policy streams contributed little once epidemic-growth signals were included. Simulated benefits arise mainly from earlier activation per se and are hypotheses requiring prospective, implementation-level evaluation rather than demonstrated effects.

Keywords: Artificial Intelligence ,Early Warning System ,Machine Learning ,Digital Epidemiology ,Mobility Data ,Data Leakage ,Decision Curve Analysis ,Seir Model; Public Health Intervention.