Title : A cyber-physical-social lens for malaria elimination in Africa: Comparative forecasting and uncertainty quantification using LSTM, GRU, and ensemble methods
Abstract:
Eliminating malaria in sub Saharan Africa demands an orchestrated strategy that integrates computational forecasts, physical interventions, and community level behaviors. Guided by the Cyber Physical Social Systems (CPSS) paradigm, this study models and predicts annual malaria case counts in Ethiopia from 2000 to 2020, while explicitly quantifying forecast uncertainty. We benchmark a comprehensive suite of methods: conventional time series models (ARIMA, SARIMA, naïve, drift, exponential smoothing, moving average), machine learning algorithms (Random Forest, XGBoost, SVM, multilayer perceptron – MLP), deep learning architectures (LSTM, GRU), and a simple unweighted ensemble combining RF, XGB, LSTM, and GRU. Ablation experiments assess the effect of input sequence length (2,3,5), number of hidden layers (1,2,3), and the inclusion of rolling window features (mean and standard deviation over 2 and 3 year windows). Uncertainty is reported as 95% prediction intervals computed as ±1.96 × RMSE. Strikingly, the multilayer perceptron (MLP) delivers the highest accuracy by a large margin, achieving an MAE of 14.54 and an RMSE of 17.17, far surpassing all other models. The unweighted ensemble (MAE = 79.21, RMSE = 83.59) and LSTM (MAE = 72.38, RMSE = 82.47) perform reasonably, whereas ARIMA (MAE = 69.93, RMSE = 72.53) and SARIMA (MAE = 54.64, RMSE = 58.94) are considerably less accurate. Classification performance using the median training threshold (120.5 cases) mirrors regression results: the MLP achieves perfect accuracy (1.0) and AUC ROC (1.0), while other classifiers reach at most 0.75 accuracy. Feature importance analysis from Random Forest and XGBoost identifies the first lag (yₜ₋₁) and the 2 year rolling mean as the dominant predictors, linking transmission persistence to social determinants such as delayed health seeking behavior, incomplete treatment, and inconsistent use of insecticide treated nets. Residual diagnostics for the MLP show approximately normal errors, no heteroscedasticity, and no significant autocorrelation. Diebold Mariano tests confirm that MLP forecast errors are significantly smaller than those of all other models (p < 0.05), and McNemar’s test confirms the MLP’s classification superiority. Ablation results demonstrate that for LSTM/GRU, a sequence length of three years with two hidden layers is optimal, and that including rolling features improves Random Forest and XGBoost MAE by 8–12%. Training time for the MLP is under one second, enabling rapid retraining. The proposed CPSS informed forecasting framework can enable early warning systems, guide rational allocation of vector control resources (insecticide treated nets, indoor residual spraying, artemisinin based therapies), and incorporate community feedback loops, thereby accelerating progress toward malaria elimination in Ethiopia and potentially across Africa.
