Predicting level of care in home care

Abstract

This thesis aims to develop prediction models that help Dutch residential care homes anticipate when a client will require more care. Supervised machine learning is used to support timely reindication applications and reduce the financial burden of delayed applications using historical structured and unstructured data from care homes. Currently, there is no such predictive system in place, in part because of the uniqueness of the Dutch long-term care system and the scarcity of standardized, labeled data that integrates administrative and clinical data of residential healthcare clients. To address this gap, multiple classification and regression models were trained and evaluated across different experimental setups, including client-month aggregates, text-only features, and a custom BERT-based model. The findings show how data-driven models can be used to uncover reindication needs and serve as a basis for proactive decision support in residential care environments.