Federated learning for violence incident prediction in a simulated cross-institutional psychiatric setting

Inpatient violence is a common and severe problem within psychiatry. Knowing who might become violent can influence staffing levels mitigate severity. Predictive machine learning models assess each patient's likelihood of becoming based on clinical notes. Yet, while benefit from having more data, data availability limited as hospitals typically do not share their for privacy preservation. Federated Learning (FL) overcome the limitation by training in decentralised manner, without disclosing between collaborators. However, although several FL approaches exist, none these train Natural Language Processing In this work, we investigate application to Processing, applied task Violence Risk Assessment simulating cross-institutional psychiatric setting. We compare four models: two local models, federated model data-centralised model. Our results indicate that outperforms has similar performance These findings suggest be used successfully setting step towards new applications notes