An Interpretable Predictive Model of Vaccine Utilization for Tanzania

Providing accurate utilization forecasts is key to maintaining optimal vaccine stocks in any health facility. Current approaches to vaccine utilization forecasting are based on often outdated population census data, and rely on weak, low-dimensional demand forecasting models. Further, these models provide very little insights into factors that influence vaccine utilization. Here, we built a state-of-the-art, machine learning model using novel, temporally and regionally relevant vaccine utilization data. This highly multidimensional machine learning approach accurately predicted bi-weekly vaccine utilization at the individual health facility level. Specifically, we achieved a forecasting fraction error of less than two for about 45% of regional health facilities in both the Tanzania regions analyzed. Our "random forest regressor" had an average forecasting fraction error that was almost 18 times less compared to the existing system. Importantly, using our model, we gleaned several key insights into factors underlying utilization forecasts. This work serves as an important starting point to reimagining predictive health systems in the developing world by leveraging the power of Artificial Intelligence and big data.