Deep learning architectures for modeling and forecasting stroke cases in Ghana.

INTRODUCTION: Stroke remains a leading cause of global morbidity and mortality, ranking second in deaths and third in disability-adjusted life years (DALYs). Its burden is particularly severe in low- and middle-income countries such as Ghana, where stroke is currently the leading cause of death. However, local data and predictive modeling remain limited, hindering effective health planning and intervention. This study aimed to model and forecast stroke incidence in Ghana using advanced deep learning techniques to support data-driven public health strategies. MATERIAL AND METHODS: Monthly stroke case data from 2018 to 2023 were obtained from Ghana Health Service. Four deep learning models; Long Short-Term Memory (LSTM), Bayesian LSTM (BLSTM), Convolutional LSTM (ConvLSTM), and Bayesian ConvLSTM (BConvLSTM), were employed to capture spatiotemporal patterns in stroke incidence. Diabetes prevalence was included as a covariate. Model performance was evaluated using mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE). Data analyses were carried out using python version 3.13.7 and R version software. RESULTS: LSTM and BLSTM models showed strong forecasting performance, with LSTM yielding the lowest errors. ConvLSTM and BConvLSTM models underperformed significantly. Forecasts from 2024 to 2028 reveal initial variability in 2024, with monthly cases between 1,694 and 2,007, followed by gradual stabilization through 2028, where values converge between 1,774 and 1,781. CONCLUSION: The study highlights a persistently high but stabilizing stroke burden in Ghana. It underscores the urgent need for targeted interventions addressing modifiable risk factors, particularly diabetes, and supports LSTM as the most effective model for forecasting in this context.