Multi-source sparse broad transfer learning for parkinson's disease diagnosis via speech.

Area of Science:

• Computational Linguistics
• Machine Learning
• Biomedical Signal Processing

Background:

• Diagnosing Parkinson's disease (PD) using speech analysis offers a non-invasive method for data collection.
• Limited sample sizes in PD speech datasets hinder the development of accurate recognition models.
• Overfitting remains a challenge when training models on small, specialized datasets.

Purpose of the Study:

• To introduce a novel multi-source sparse broad transfer learning (SBTL) method for enhanced PD speech recognition.
• To address the limitations of small sample sizes and overfitting in PD speech data analysis.
• To improve the accuracy and stability of PD diagnosis through speech.

Main Methods:

• Developed a multi-source sparse broad transfer learning (SBTL) method inspired by incremental broad learning.
• Utilized a sparse network for preprocessing PD speech data to identify intrinsic invariant features.
• Employed an incremental learning mechanism to evaluate transfer effectiveness and adapt model structure for positive knowledge transfer.

Main Results:

• SBTL demonstrated significant advantages over existing transfer learning methods for PD speech diagnosis.
• Achieved improvements of at least 2.58% in accuracy, 5.71% in precision, 12% in sensitivity, and 14.81% in F1-score.
• Showed comparable sensitivity to well-known transfer learning methods while maintaining superior performance in other metrics.

Conclusions:

• SBTL is an effective, efficient, and stable multi-source transfer learning method for PD speech recognition.
• The proposed method successfully balances learning capability and overfitting for limited PD speech data.
• SBTL provides more accurate assistance for clinicians in making PD diagnostic decisions.