Federated Learning for Privacy-Preserving Chronic Pain Rehabilitation Outcome Prediction Across Multi-Centre Physiotherapy Clinics

Abstract

Rehabilitation of chronic pain in various physiotherapy centres generates a significant amount of detailed long-term patient information that can be used to forecast. But laws on data privacy, such as GDPR Article 9 and HIPAA, prevent us from collecting this data in a single location, which is necessary in conventional machine learning processes. This paper presents a Federated Learning architecture that enables the prediction of the results of rehabilitation in five physiotherapy clinics, and patient data remains confidential. It addresses two key problems: providing privacy and addressing disparities in data distribution among clinics that cannot be addressed by standard federated approaches. To prevent inter-clinic drift in the system, a combination of FedProx and proximal regularisation ( μ =0.01) are used. It additionally employs three privacy protection layers: noise addition to data (differential privacy, ε = 3.2, δ = 10 -5) with the help of Rényi accounting, safe cryptographic aggregation, and encrypted communication with the help of TLS. Such procedures comply with GDPR Article 89 and HIPAA conditions. The system consists of a local model with two components that process both static and time-based data with a Dense network and an LSTM network, respectively, enabling the system to learn about initial health risks and the time-based variation of recovery. It was tested on 5,000 synthetic patient records with 43 validated variables in five clinics in five-fold cross-validation and leave-one-clinic-out testing. It achieved a weighted F1 score of 0.774, AUC-ROC of 0.861, and Cohen's κ of 0.661. This is 7.1 percentage points better than local training and only 4.7 percentage points worse than a centralised model that does not safeguard privacy. The privacy measures only reduced the F1 score by 1.5 percentage points. The four-class prediction target is clinically meaningful, as indicated by clinical results of 30.2% reduction in pain (VAS), 27.5% decrease in disability (ODI), and a 14.3% improvement in physical health scores (SF-36). These findings demonstrate that privacy-preserving federated learning can deliver useful predictions of rehabilitation outcomes in various clinics without causing damage to patient privacy or violating the law.