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

Wenna Chen^1*Rongfu Lv²Xiaowei Du¹Xiangyu Chen¹Hao Wang¹Jincan Zhang²Ganqin Du^1*

• ¹The First Affiliated Hospital of Henan University of Science and Technology, Luoyang, China
• ²College of Information Engineering, Henan University of Science and Technology, Luoyang, China

Parkinson's disease (PD) is the world's second most common neurodegenerative disorder. Traditional diagnostics, which rely on clinical assessment and imaging, are often invasive, costly, and require specialized medical personnel, creating significant barriers to early detection. Given that approximately 90% of PD patients develop vocal impairments, vocal analysis has emerged as a promising non-invasive diagnostic tool. However, individual deep learning models are frequently limited by overfitting and poor generalizability, which constrain their diagnostic accuracy. To overcome these limitations, this study proposes a PD classification method utilizing spectrogram feature fusion with pre-trained convolutional neural networks (CNNs). The dataset, sourced from the First Affiliated Hospital of Henan University of Science and Technology, includes voice recordings from 61 PD patients and 70 healthy controls (HC). Preprocessing the raw speech signals yielded 2,476 spectrograms. Three pre-trained models, DenseNet121, MobileNetV3-Large, and ShuffleNetV2, were used for feature extraction. To ensure dimensional alignment for fusion, the output of MobileNetV3-Large was adjusted to a dimension of (1024, 7, 7) using a 1×1 convolutional layer before features were fused via summation. The results obtained from using 5-fold cross-validation indicate that models using feature fusion consistently surpass individual models across all evaluation metrics. Specifically, the fusion of MobileNetV3-Large and ShuffleNetV2 achieved the highest accuracy of 95.56% and an AUC of 0.99. Comparative experiments with existing state-of-the-art methods confirm that the proposed approach achieves competitive classification performance. The fusion of multi-model features more effectively captures subtle pathological signatures in PD speech, offering a reliable, low-cost, and non-invasive tool for auxiliary PD diagnosis. The code is available at https://github.com/lvrongfu/pjs.