A machine learning method to process voice samples for identification of Parkinson's disease

Area of Science:

• Computational neuroscience
• Speech processing
• Biomedical engineering

Background:

• Machine learning aids Parkinson's disease detection using voice data.
• Telephone-based voice recordings offer accessible data collection but yield inconsistent results.
• Phonation-related voice features are crucial for analysis.

Purpose of the Study:

• To validate telephone-collected voice recordings for Parkinson's disease detection.
• To apply a novel deep learning approach for analyzing voice spectrograms.
• To improve the classification accuracy of Parkinson's disease using speech analysis.

Main Methods:

• Collected sustained vowel /a/ recordings from 50 Parkinson's disease patients and 50 healthy controls via personal telephones.
• Applied machine learning classification using phonation-related voice features.
• Utilized a pre-trained Inception V3 convolutional neural network with transfer learning to analyze voice spectrograms, considering time and frequency intensity estimates.

Main Results:

• Demonstrated the reliability of personal telephone-collected voice recordings for Parkinson's disease detection.
• Showcased the effectiveness of the deep learning model in classifying Parkinson's disease patients.
• The novel deep learning approach outperformed traditional methods by analyzing speech intensity across time and frequency scales.

Conclusions:

• Telephone-based voice recordings are reliable for Parkinson's disease detection.
• Deep learning models, specifically Inception V3 with transfer learning, offer a superior approach for analyzing voice spectrograms.
• This method enhances the potential for widespread, non-invasive Parkinson's disease screening.