Parkinson's disease detection using spectrogram-based multi-model feature fusion networks.

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Parkinson's disease (PD) diagnosis relies on costly and invasive methods, hindering early detection.
• Vocal impairments affect ~90% of PD patients, presenting an opportunity for non-invasive diagnosis.
• Individual deep learning models often suffer from overfitting and poor generalizability in PD detection.

Purpose of the Study:

• To develop and evaluate a PD classification method using spectrogram feature fusion with pre-trained convolutional neural networks (CNNs).
• To overcome the limitations of single deep learning models for improved PD diagnosis.

Main Methods:

• Collected voice recordings from 61 PD patients and 70 healthy controls.
• Extracted features using three pre-trained CNNs: DenseNet121, MobileNetV3-Large, and ShuffleNetV2.
• Fused extracted features via summation after dimensional alignment for classification.

Main Results:

• Feature fusion models consistently outperformed individual models.
• The fusion of MobileNetV3-Large and ShuffleNetV2 achieved 95.56% accuracy and 0.99 AUC.
• The proposed method demonstrated competitive performance against state-of-the-art approaches.

Conclusions:

• Multi-model feature fusion effectively captures pathological speech signatures in PD.
• This method offers a reliable, low-cost, non-invasive tool for auxiliary PD diagnosis.
• The approach holds significant potential for clinical application in PD management.