Clustering-based demand forecasting with an application to immunoglobulin products

Efficient healthcare supply chain management can benefit greatly from accurate demand forecasting, which can help reduce costs and prevent patient treatment delays. This is particularly challenging for demand forecasting for the human immunoglobulin blood product stored in hospital blood banks due to the diverse patient population and therapeutic applications. Our study proposes an iterative clustering-based demand forecasting framework to address this issue. We cluster patients based on domain knowledge and demand pattern characteristics using the robust and sparse K-means algorithm. We then employ time-series analysis techniques to forecast demand for each cluster, aggregate the forecasts, and evaluate the performance. The potential variables affecting the clustering and forecasting results are identified to make this process iterative and to find the best clustering scheme based on forecast performance. For example, the optimal number of clusters, K , in a K-means algorithm is unknown. Therefore, we choose K to optimize the forecast performance. Clustering algorithms can also be sensitive to feature selection, so using an extension of K-means with weighted features, the bound on feature weights is included as an unknown input variable in the iterative process. We further enhance the forecasting model by incorporating individual patient-level predictions from the cluster identified with extended treatment plans, which contains patients with more data points and better individual predictability. The proposed framework outperforms baseline ARIMA and LSTM network models trained on aggregate demand data. Moreover, the results show improved performance as data size increases.