ABSTRACT

Accurate forecasting of aging biomarkers enables proactive public health to predict, identify, prevent diseases and manage their progression. Commonly used longitudinal models (GLM, GEE, GLMM) struggle with nonlinear dynamics in sparse panel data from population surveys. This study uses Survey of Health, Ageing and Retirement in Europe (SHARE) Waves 5–6–8 (≈10,800 participants) to forecast key biomarkers at Wave 9, including BMI, grip strength, and binary outcomes such as hypertension and diabetes. Multiple deep learning models were benchmarked—CNN, DNN, LSTM, GRU, CNN-LSTM, CNN-GRU, and multi-task learning (MTL). Performance improved with more historical waves, while delta-t intervals added little. Hybrid CNN-RNN models excelled in single-task prediction (CNN-LSTM: BMI MAE=1.17; CNN-GRU: grip MAE=3.48), with MTL delivering comparable cross-task accuracy by leveraging shared representations. Binary outcomes reached AUC>0.96 across deep models. SHAP confirmed prior outcomes as dominant predictors, validating temporal learning. Overall, these multi-wave neural architectures outperformed statistical baselines, supporting their potential utility for longitudinal aging predictions in precision health.