Multi-Modal Analysis of Parkinson Disease data Using Advanced Deep Learning Techniques

Abstract

Early diagnosis is essential for effective treatment of Parkinson’s disease (PD), a progressive neurological disorder that affects movement and cognitive functions. This study presents a multi-modal analysis for PD classification using deep learning algorithms applied to clinical, audio, and handwriting image data. Recurrent neural networks (RNNs), including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional LSTM, are employed for clinical and audio data analysis, while convolutional neural networks (CNNs) are utilized for handwriting image classification. The results demonstrate varying model performance across different data modalities. Among the clinical data models, GRU achieved the highest accuracy of 82.42%, indicating its effectiveness in capturing sequential dependencies in medical records. For audio-based classification, RNN outperformed all other models with 94.87% accuracy, while LSTM and GRU showed comparable performance, each reaching 92.31% accuracy. In the image modality, CNN without Batch Normalization attained 82.93% accuracy, whereas Batch Normalization improved performance to 85.37%, highlighting its role in stabilizing training and enhancing feature extraction. These findings emphasize the importance of modality-specific deep learning models and their potential to enhance early and accurate PD detection. The study emphasise the significance of multi-modal approaches in medical diagnostics, paving the way for improved, non-invasive, AI-driven assessments.