Area-Level Medication Demand Prediction Via Multitask Spatial-Temporal Machine Learning

Understanding and predicting area-level medication demand has become increasingly critical to advancing population health, given its direct implications for disease prevalence and resource optimization. Traditional approaches to forecasting medication demand primarily rely on statistical models or single-task machine learning methods. These methods often focus on isolated medication demands, thereby overlooking the complex interactions of spatial–temporal patterns and multi-morbidity correlations, failing to accurately simulate the real world. This article takes a significant leap forward by introducing an innovative multi-task learning architecture designed to address the problem of area-level prescription demand forecasting. Utilizing the mixed model of experts’ structure allows for the extraction of inter-medication correlations, while the employment of graph attention networks and gated recurrent units captures the essential spatial–temporal patterns within prescription data. Moreover, this study employs a specialized loss function to balance the intricate tradeoffs among various multi-task learning objectives, thereby ensuring optimal performance across all tasks. Experimental results show that our multi-task learning framework improves predictive performance by up to 10.2% compared to single-task and traditional methods. Furthermore, our model outperforms the strongest multi-task baseline, ASTGCN, with an additional 2.7% gain in mean absolute error and root mean square error. The results demonstrate the effectiveness of multi-task learning for area-level medication demand forecasting offering a solution for accurate and scalable area-level medication demand forecasting.