Assessment of patients with Parkinson's disease based on federated learning

This paper presents federated Learning (FL), which is based on wearable devices, and applies the actual leg agility data that has been collected from people living with Parkinson's disease (PD) to the model. Studies have shown that the implementation of FL can effectively protect the data privacy of PD patients. The classification accuracy of leg agility data is reduced by 2.72% when compared to the conventional method of summarizing all the data. However, it is higher than the model accuracy of each data owner, having increased by 22.68%. Secondly, during the communication process, the upload or download of the model parameters of each terminal node is interrupted for N times at the same time, and it is found that interrupting the upload of parameters reduces the accuracy of the central model. The impact of interrupting the download parameters on the central model is negligible. Then, the communication process of the terminal nodes with different data amounts was interrupted respectively, and it was found that the accuracy of the central model was basically not affected. Finally, noise is introduced to the various parameters in the communication process. The accuracy of the central model begins to gradually deteriorate as soon as the noise intensity reaches 0.012 or higher.