Optimized Deep Learning Model for Early Parkinson’s Disease Detection Using Speech Spectrograms

Parkinson’s disease (PD) is a progressive neurodegenerative disorder, and early detection is crucial for timely intervention and management. This study proposes an innovative approach for non-invasive PD detection using deep learning techniques applied to speech analysis. Speech recordings of Parkinson’s patients and healthy individuals were converted into spectrograms using the Short-Time Fourier Transform (STFT), which were then processed using Principal Component Analysis (PCA) for dimensionality reduction. We employed several convolutional neural network (CNN) architectures, including CNN, MobileNet, ResNet50, and Inception, to classify the data. A 5-fold cross-validation was applied. Among the models, CNN demonstrated superior performance, achieving 99.19% accuracy, 100% precision, and 99.21% F1 score without cross-validation, which improved to perfect scores across all metrics when cross-validation was applied. The results indicate that deep learning, coupled with speech signal processing and cross-validation, provides a powerful tool for the early, non-invasive detection of Parkinson’s disease. This approach holds significant potential for improving diagnosis accuracy and enabling timely medical intervention.