Wave-aware mortality prediction in COVID-19: a multi-stage feature selection and explainable machine-learning framework

BACKGROUND: Mortality among hospitalized COVID-19 patients changed markedly across successive epidemic waves as circulating variants shifted, population immunity increased, and health-system pressures fluctuated. These increasingly heterogeneous conditions raise concerns about whether prediction models developed during early phases remain reliable in later epidemic contexts. OBJECTIVE: To develop a wave-aware machine-learning framework that differentiates between stable physiological predictors of mortality and context-dependent predictors whose importance changed across waves. METHODS: We analyzed 732,654 adult hospitalizations from Iran's national COVID-19 registry across Waves 2-5 and the post-Wave-5 period. All preprocessing, feature selection, and model training were conducted independently within each wave to preserve temporal structure and prevent leakage. A three-stage feature-selection approach Elastic Net shrinkage, Random Forest importance ranking, and Variance Inflation Factor filtering identified both stable and time-varying predictors. Model performance for Logistic Regression, Random Forest, and Deep Neural Networks was evaluated on wave-specific held-out test sets, and temporal robustness was assessed through cross-wave validation. RESULTS: A consistent physiological vulnerability core age, hypoxemia, and major chronic comorbidities was present across all waves. Nonlinear models outperformed Logistic Regression, with Random Forest achieving AUCs up to 0.94 and F1 scores up to 0.68 in within-wave testing. However, early-wave models showed marked degradation when applied to later waves: cross-wave AUC declined moderately (e.g., RF: 0.87 within-wave → 0.80 early→Wave-5), whereas F1 collapsed due to pronounced miscalibration (e.g., RF F1: 0.68 → 0.13 → 0.07). SHAP analyses revealed increasing importance of vaccination-related variables and shifting comorbidity profiles in later waves. Threshold-sweep analyses suggested that the F1-optimal probability threshold varied widely across models and waves, underscoring the impact of temporal drift on threshold-dependent performance. CONCLUSIONS: COVID-19 mortality risk was determined by a stable physiological core overlaid by dynamic, wave-specific factors shaped by changing variants, immunity, and hospital strain. The divergence between preserved discrimination and degraded F1 demonstrates that static prediction models are highly susceptible to temporal drift. Drift-aware approaches such as recalibration, periodic refitting, or adaptive thresholding may be essential to maintain clinical utility in evolving epidemic environments. The wave-aware, feature-selection-guided, explainable ML framework presented here offers a generalizable basis for developing temporally robust prediction tools for COVID-19 and other rapidly evolving infectious diseases.