Explainable machine learning-based forecasting of COVID-19 cases: a comparative analysis of maharashtra and Kerala, India

Background The Coronavirus Disease 2019 (COVID-19) pandemic in India exhibited substantial spatial and temporal variation in transmission dynamics, influenced by human mobility, meteorological conditions, vaccination rollout, and non-pharmaceutical interventions. Methods This study forecasted district-level COVID-19 cases one-day-ahead in the districts of 2 Indian states of Maharashtra (33 districts) and Kerala (14 districts) between April 26, 2020 and June 20, 2021, using four Machine Learning (ML) models: Multiple Linear Regression (MLR), Random Forest (RF), Extreme Gradient Boosting (XGB), and Long Short-Term Memory (LSTM) networks. 17 explanatory variables spanning mobility, meteorology, vaccination, social gatherings, lockdown restrictions, and population density were incorporated, with separate models developed for the 2020 (pre-vaccination) and 2021 (vaccination) periods in each state. Models were evaluated using 10-fold walk-forward cross-validation, and SHapley Additive exPlanations (SHAP) analysis was applied to the best-performing model to quantify the magnitude and direction of each variable’s contribution. Results LSTM was the overall best performing model, with R 2 of nearly 0.8 (NRMSE = 0.07) in Maharashtra and 0.6 (NRMSE = 0.15) in Kerala. SHAP analysis of LSTM identified mobility-related variables, particularly workplace, retail, and residential mobility, as the dominant explanatory variables in most models, though their importance and direction of association shifted between states and pandemic phases: reduced workplace mobility was associated with higher predicted cases in Maharashtra in 2020, while higher workplace mobility was associated with higher predicted cases in 2021. Increased grocery and retail activity was associated with higher case counts in 2020, while higher residential mobility was associated with higher cases in 2021. Population density remained an important and comparatively stable variable in Kerala across both years but was influential only in 2020 in Maharashtra. Higher vaccination instances were consistently associated with higher predicted case counts in 2021 for both states. Lockdown restrictions showed opposite associations between the two states in 2021, coinciding with higher predicted cases in Maharashtra but lower predicted cases in Kerala. Meteorological variables showed comparatively modest but more consistent effects across states and years. Conclusion These findings demonstrate that mobility patterns and population density were the variables most consistently associated with district-level COVID-19 case trajectories in Maharashtra and Kerala, while vaccination and lockdown measures carried associations that varied with timing and state context. These results highlight the importance of context-specific models that account for changes in transmission patterns over time when understanding and forecasting epidemic spread. Clinical trial registration Not applicable