TremorFusion: AI-driven feature extraction for multi-class Parkinson's tremor classification using CSVM and DeepK-CNN

Parkinson's disease (PD) is a common neurological disorder that, if left untreated, can significantly impair a patient's social and physical functioning. This study evaluates the feasibility of a wrist-worn accelerometer sensor device for detecting and classifying five distinct types of PD tremors. A novel feature extraction technique is proposed for the Cubic Support Vector Machine (CSVM) model, along with a pre-processed tremor spectrum analysis utilizing Non-negative Matrix Factorization (NMF) for a hybrid model that integrates k-Nearest Neighbor (KNN) with a Convolutional Neural Network (CNN), referred to as DeepK-CNN. Poincaré-based analysis was employed to extract time-domain features from tremor data, which were subsequently selected for the CSVM model based on the area under the receiver operating characteristic (ROC) curve. The dataset comprises 90 patients, with 30 exhibiting tremor types I-V, contributing 151 labeled tremor episodes over 15 h of recorded data. Using leave-one-patient-out fivefold cross-validation, the CSVM model demonstrated superior performance, achieving 95.28% mean sensitivity, 95.51% mean specificity, and a mean false alarm rate (FAR) of 0.15/24 h. The DeepK-CNN model achieved an 87.51% mean sensitivity, 91.65% mean specificity, and a mean FAR of 0.64/24 h. Performance evaluation yielded p-values of <0.001 for the CSVM model and <0.05 for the DeepK-CNN model. The proposed method enhances real-time tremor classification, potentially improving personalized treatment strategies and facilitating remote patient assessment for clinical adoption.