Classification of Parkinson's disease utilizing multi-edit nearest-neighbor and ensemble learning algorithms with...

Area of Science:

• Biomedical Engineering
• Computational Linguistics
• Machine Learning

Background:

• Speech-based data offers a non-invasive method for Parkinson disease (PD) classification.
• Speech pattern analysis is crucial for developing tele-diagnosis and tele-monitoring models for Parkinsonism.
• Reducing noise in speech samples is key to improving PD classification accuracy and stability.

Purpose of the Study:

• To propose and examine a novel PD classification algorithm.
• To investigate the effectiveness of instance selection in optimizing speech-based PD classification.
• To enhance classification accuracy and stability in PD detection using speech data.

Main Methods:

• A hybrid algorithm combining multi-edit-nearest-neighbor (MENN) for sample selection and ensemble learning (Random Forest or Decorrelated Neural Network Ensembles) was developed.
• The MENN algorithm iteratively selects optimal training speech samples with high separability.
• Trained ensemble models were applied to test samples for PD classification, validated against existing algorithms on public datasets.

Main Results:

• The proposed algorithm achieved a 29.44% improvement in classification accuracy compared to other methods.
• The MENN algorithm alone demonstrated a substantial accuracy improvement of up to 45.72%.
• The combined MENN and Random Forest approach exhibited superior stability in PD classification.

Conclusions:

• The developed method effectively improves Parkinson disease classification using speech data.
• This approach holds promise for future research aimed at refining PD classification techniques.
• The findings support the application of this algorithm in advanced tele-diagnosis and tele-monitoring systems for PD.