Vocal Features-Driven Parkinson’s Identification through Attention-Based LSTM-BiGRU with LOPO CV and EBM

Parkinson’s Disease (PD) is a progressive neurodegenerative disorder that causes loss of motor functions, emphasizing the need for prompt and accurate detection. This study aimed to design an automated system for early PD detection focusing on vocal features, as verbal disturbances often appear before more noticeable physical symptoms. A multisource dataset comprising speech recordings, biomedical voice metrics, and replicated voice recordings was utilized for this task. Several data augmentation techniques were explored to address data imbalances and size limitations, with Grouped Synthetic Minority Over-Sampling Technique (SMOTE) and Conditional Generative Adversarial Networks (CGAN) with Structured Conditioning were identified as the most effective. Among validated feature reduction methods, Boruta emerged as the optimal approach. This study focused on hybrid models that integrated Deep Learning (DL) and Machine Learning (ML), including attention mechanisms. Performance analysis identified the Attention-based Long Short-Term Memory - Bidirectional Gated Recurrent Unit (LSTM-BiGRU) as the superior model, attaining accuracies of 97.50%, 96.92%, and 96.25% across the datasets, respectively. Validation with K-fold and Leave-One-Participant-Out (LOPO) Cross-Validation (CV) ensured the model’s generalizability. Additionally, Explainable Boosting Machines (EBM) were employed to interpret the model’s decision-making process, enabling accurate and interpretable diagnostic tools for early PD detection.