Federated Transferring Multi-Channel CNN for Diagnosis of Parkinson's Disease under Unseen and Small Data

Gait signals have been extensively investigated to diagnose Parkinson's disease (PD). However, the limited data and non-negligible individual differences hinder the practical application of PD diagnosis. To address this issue, a federated transferring multi-channel convolutional neural network is proposed to diagnose PD and assess its severity level under unseen and small samples. Moreover, the inflationary performance has been explored and avoided to improve the generality of the methods. A multi-channel convolutional neural network is designed as a backbone model and is dynamically federated and pre-trained using public datasets. Subsequently, the obtained global network is personally fine-tuned based on the "XJTU dataset" we collected using a wearable system from 20 PD patients and 32 healthy controls. Results indicate that federated transfer learning helps diagnose PD under unseen and small data. Integrating this method into a wearable system has the potential to offer daily-life services for elderly individuals.