An Integrated AI-Driven Public Health Analytics Framework for Predicting Chronic Disease Burden Using Social Determinants of Health, Medical Imaging, and Electronic Health Records

Abstract

This study created an integrated framework for artificial intelligence prediction of both individual and population burden of chronic diseases among adolescents from electronic health records, measures of social determinants and liver ultrasound images. A cohort of 24,000 adolescents (10 to 19 years) was simulated and assessed through 24 months. Each modality-specific feature was obtained using a Temporal Transformer, a multilayer perceptron, and an EfficientNet-B0, and then fused together by using attention. The integrated framework showed 90.6% accuracy, 87.8% sensitivity, 91.2% specificity, 0.944 ROC–AUC and 84.9% F1-score. Those most important factors were BMI-for-age z score, fasting glucose, liver steatosis, family history and neighbourhood deprivation. Integrating multiple modes improved risk stratification and disease burden estimation over conventional and unimodal risk stratification models based on simulations. These computational findings would need to be tested in “real world” multicentre adolescent cohorts before being clinically or public healthly implemented, however. The concept of fairness, calibration, privacy, safety and potential clinical usability is also to be evaluated separately.