Federated Learning for Integrating Diverse Speech Biomarkers for Parkinson's Disease Classification

The diagnostic paradigm of Parkinson's disease (PD) involves analyzing multiple biomarkers including speech, EEG, gait, tremor, and fMRI at different stages of diagnosis along with the inspection of the clinical history of patients. Multiple research works have aimed to automate and expedite the diagnostic process utilizing numerous Machine learning methods, however, mostly including single or two biomarkers. Hence, there is a diagnostic potential of integrating different biomarkers relevant to PD to enhance the accuracy and robustness. These biomarkers encompass different data forms, such as 1D signals, 2D images, and 3D scans. However, a significant challenge arises from the limited availability of comprehensive datasets. Many datasets do not contain all relevant biomarkers for the same set of subjects, limiting the ability to study them together. To overcome this limitation, we adopt a fusion-based federated learning approach by training separate models on different types of speech biomarkers sourced from different databases. The central server, which acts like a diagnostic centre, unifies insights from different speech biomarkers across multiple sites evaluating the relevance of other diverse biomarkers from various sources. Through this fusion federated framework, we achieve an 85% classification accuracy in PD diagnosis by effectively integrating different speech data types from multiple sources. Our findings highlight the potential of federated learning to combine disparate biomarker data for enhanced disease classification, paving the way for more holistic and data-driven approaches in medical machine learning applications.