Exploring Spectrogram-Based Audio Classification for Parkinson's Disease: A Study on Speech Classification and...

Area of Science:

• Computational Linguistics
• Biomedical Engineering
• Artificial Intelligence

Background:

• Parkinson's disease (PD) significantly impacts patients' vocal abilities, leading to speech impairments.
• Accurate and early detection of PD is crucial for effective patient management and treatment.

Purpose of the Study:

• To compare the efficacy of an Audio Spectrogram Transformer (AST) and a Convolutional Neural Network (CNN)-based Pretraining, Sampling, Labeling, and Aggregation (PSLA) model in classifying speech from normal individuals and those with Parkinson's disease.
• To quantitatively and qualitatively analyze the performance of these models in identifying speech patterns associated with Parkinson's disease.

Main Methods:

• Utilized two advanced speech classification models: AST, a transformer-based approach, and PSLA, a high-performance CNN-based model.
• Performed quantitative analysis comparing accuracy and Area Under the Curve (AUC) metrics.
• Employed Class Activation Map (CAM)-based Explainable AI (XAI) techniques, including GradCAM and EigenCAM, for qualitative assessment of model focus on acoustic features.

Main Results:

• PSLA achieved higher accuracy (over 4% improvement) and a superior AUC (97.43%) compared to AST (94.16%).
• XAI analysis revealed that the PSLA model effectively focused on the muffled frequency bands characteristic of Parkinson's speech.
• Heatmap analysis visually confirmed the model's attention to relevant speech features, even in cases of misclassification.

Conclusions:

• The PSLA model is a more suitable tool for diagnosing Parkinson's disease through speech analysis compared to the AST model.
• The study validates the practical applicability of AI models in identifying Parkinson's disease via speech, supported by explainable AI insights.