Objective: To improve privacy and personalisation for artificial intelligence (AI), this work compared federated learning (FL) to a conventional, centralised system for predicting Movement Disorder Society-Unified Parkinson’s Disease Rating (MDS-UPDRS) scores from real-world digital measures of gait.

Background: Cloud-based AI with smart-health technology presents a powerful tool to passively monitor disease progression. However, current methods require a centralised system, for which privacy concerns arise from transmitting patient data. A potential solution is FL, a system which shares model parameters with the cloud instead of identifiable patient data, thus protecting patients’ confidentiality.

Method: As a part of the CiC-PD study [1], 55 people with Parkinson’s were assessed using the MDS-UPDRS and, to measure gait, wore an inertial measurement unit on the lower-back for 7 consecutive days. Daily gait measures and patient demographics (age, sex, and body mass index) were used to estimate the MDS-UPDRS Part III score with a fully connected neural network (NN) in a conventional, centralised system (C-NN) and a federated system (FL-NN). For FL-NN, a local model was trained for each individual participant, and those model weights were sent to and aggregated by the central server, giving the global NN model which was then redistributed to the local models. This global model was tuned to participants’ training data and then evaluated with their testing data, repeating this process until accuracy stopped improving. For both C-NN and FL-NN, 10-fold cross validation was used to evaluate the model’s (global model for FL-NN) accuracy when predicating MDS-UPDRS Part III score in new patients, as a preliminary investigation for the TORUS project [2].

Results: Data was available for 53 participants. C-NN performed reasonably well with a mean absolute error (MAE) of 10.01 (Figure 1a). By comparison, the global FL-NN model had a slightly reduced accuracy with an MAE of 12.56 (Figure 1b). The personalised local models for the FL-NN were more accurate, with a mean MAE of 7.80.

Conclusion: This work demonstrates that FL can protect privacy and provide personalisation for participants with minor costs to the accuracy of the NNs. Future work will improve these NNs, leading to an accurate AI system that passively tracks Parkinson’s progression, whilst protecting patients’ private data.

Scatter plots of true vs predicted MDS-UPDRS score

References: [1] Packer et al BMJ Open 2023 [2] https://torus.ac.uk/

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/privacy-and-personalisation-predicting-parkinsons-severity-with-federated-learning/