Aggression Prediction across Multiple Institutions using Federated Learning
by Pablo Mosteiro (Utrecht University)
Electronic Health Records contain extremely valuable information about patients that could be exploited to improve clinical treatment. Due to ethical and legal considerations, sharing data and models across institutions is extremely challenging, limiting the possibilities for collaboration and cross-institutional learning. Federated Learning facilitates the decentralised training of machine learning models without disclosing data between collaborators for privacy preservation. In this work, we apply Federated Learning in a simulated cross-institutional psychiatric setting to predict patient aggression based on psychiatric clinical notes. The dataset is split up to simulate two psychiatric institutions by distributing patients based on psychiatric sub-wards. Two independent local models, a federated model, and a data-centralised model have been trained. The Federated Learning model outperformed the local models on our dataset and performed on par with the data-centralised model. Further work with more samples is needed to evaluate the statistical significance of our findings. The results pave the way for research on Federated Learning in a real cross-institutional psychiatric clinical setting.
-