Explainable Machine Learning for Drug Shortage Prediction in a Pandemic Setting

The COVID-19 pandemic poses new challenges on pharmaceutical supply chain including the delays and shortages of resources which lead to product backorders. Backorder is a common supply chain problem for pharmaceutical companies which affects inventory management and customer demand. A product is on backorder when the received quantity from the suppliers is less than the quantity ordered. Knowing when a product will be on backorder can help companies effectively plan their inventories, propose alternative solutions, schedule deliveries, and build trust with their customers. In this paper, we investigate two problems. One is how to use machine learning classifiers to predict product backorders for a pharmaceutical distributor. The second problem we focused on is what are the particular challenges and solutions for such task under a pandemic setting. This backorder dataset is different from existing benchmark backorder datasets with very limited features. We discuss our challenges for this task including empirical data pre-processing, feature engineering and understanding the special definitions of backorder under the pandemic situation. We discuss experimental design for predicting algorithm and comparison metrics, and demonstrate through experiments that decision tree algorithm outperforms other algorithms for our particular task. We show through explainable machine learning approaches how to interpret the prediction results.