# Privacy and personalisation: predicting Parkinson’s disease severity from real-world gait with federated learning

## TL;DR

This study explores using federated learning to monitor Parkinson's disease severity from home without sharing patient data, balancing privacy and accuracy.

## Contribution

The paper introduces a federated learning approach for predicting Parkinson’s disease severity using real-world gait data while preserving patient privacy.

## Key findings

- Federated learning achieved similar accuracy to traditional machine learning without sharing patient data.
- Local personalized models outperformed global models in predicting Parkinson’s severity.
- Including a subset of participants in all local models improved global model performance but reduced local accuracy.

## Abstract

Cloud-based artificial intelligence (AI) combined with smart-health technology presents a powerful tool to passively monitor disease severity. However, current methods raise privacy concerns as they require transmitting patient data to the cloud. A potential solution is Federated Learning (FL), which only shares the weights of locally trained neural networks (NNs) instead of user data. Here, we simulated an FL system to demonstrate its application for evaluating Parkinson’s disease (PD) severity in a smart-home scenario.

Retrospective data including 89 people with PD wore an accelerometer on the lower-back at home for 7 days at 18-month intervals over 6 years. Patient characteristics (age, sex, and body mass index) and clinical measures of PD were additionally collected, including the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS)-Part III. Real-world daily gait measures along with these patient characteristics were used to predict the MDS-UPDRS-III score. For FL, a local model was trained for each participant, and a global model (an aggregation of these local models) was tested on unseen participants.

The performance of a simulated FL system was compared with that of a traditional Machine Learning (ML) approach in which patient data were shared. The traditional ML approach had a mean absolute error (MAE) of 10.43. The global FL model had a similar MAE of 10.22 but was underfitted, and the mean MAE of the local, personalised models was 4.83. Shapley Additive exPlanations (SHAP) analysis showed that while the participants’ age and sex were very important in traditional ML, this was not the case for the local FL models, leading to a decrease in global model performance. Here, we show that reserving a small number of participants from the system and including them in training data for all local models restored the importance of these features and improved global FL performance (MAE = 9.26) but reduced local performance (MAE = 6.83).

This exploratory study shows that our proposed approach enables FL to achieve similar accuracy to traditional Machine Learning without sharing any patient data but with costs to the local performance, leading towards a smart-home system that prioritises personalisation and patient privacy.