Voice analysis-based machine learning models to diagnose Alzheimer’s disease(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2025年第5期
- Page:
- 685-692
- Research Field:
- 医学人工智能
- Publishing date:
Info
- Title:
- Voice analysis-based machine learning models to diagnose Alzheimer’s disease
- Author(s):
- ZHANG Yuxi1; 2; SUN Wei1; ZHU Guodong2; 3; REN Zhiyao2; 3; ZHANG Ruiqiu1; 2
- 1. School of Design, South China University of Technology, Guangzhou 510000, China; 2. Collaborative Innovation Center for Civil Affairs of Guangzhou, Guangzhou 510000, China; 3. Institute of Gerontology, Guangzhou Geriatric Hospital, Guangzhou 510000, China
- Keywords:
- Alzheimer’s disease; voice analysis; random forest; support vector machine; sequential forward selection
- PACS:
- R318；TP181；R749.16
- DOI:
- 10.3969/j.issn.1005-202X.2025.05.020
- Abstract:
- Objective To identify key acoustic features associated with the progression of Alzheimer’s disease (AD) through voice analysis combined with machine learning and feature selection techniques, thereby constructing classification models that serve as candidate tools for the early screening of AD. Methods Voice samples from AD, mild cognitive impairment (MCI) and healthy (HC) elderly individuals were obtained from the NCMMSC2021 AD voice dataset. The voice samples underwent data preprocessing, followed by feature extraction from the eGeMAPS feature set via the OpenSmile toolkit. Classification models were obtained utilizing random forest and support vector machine (SVM) algorithms. Significance testing and feature importance ranking were conducted using Python, and the further selection of the optimal features was performed through sequential forward selection (SFS). The classification performance before and after feature selection was compared and evaluated using accuracy and the area under the receiver operating characteristic curve (AUC). Results The significant acoustic features in the classification models primarily derived from spectral slope, formant, fundamental frequency, and loudness. The optimal classification performance was achieved with the SVM model following SFS feature selection, with recognition accuracies of 0.926 (AUC=0.974) for AD/MCI group, 0.875 (AUC=0.956) for AD/HC group, and 0.879 (AUC=0.904) for MCI/HC group. Conclusion SVM model performs better than random forest model, and the use of SFS for feature selection can effectively enhance model performance. Voice analysis has the potential to serve as a valuable supplementary tool for the rapid AD assessment and screening.
Last Update: 2025-06-03