Cartesian genetic programming for diagnosis of Parkinson disease through handwriting analysis: Performance vs....

Area of Science:

• Neuroscience
• Computational Biology
• Medical Informatics

Background:

• Early diagnosis of Parkinson's disease (PD) is crucial for effective patient care and improved life expectancy.
• Current high-performing diagnostic systems often use 'black-box' machine learning models, lacking interpretability and hindering clinical acceptance.
• Non-invasive, automatic methodologies for disease identification are of increasing interest.

Purpose of the Study:

• To compare various machine learning (ML) techniques for the automatic identification of PD patients using handwriting and drawing samples.
• To evaluate ML models based on both classification accuracy and interpretability.
• To identify diagnostic models that provide explicit classification rules for clinical insights.

Main Methods:

• Applied and compared multiple machine learning (ML) techniques, including white-box (Cartesian Genetic Programming, Decision Tree) and black-box approaches.
• Utilized handwriting and drawing samples from the publicly available PaHaW and NewHandPD datasets.
• Analyzed classification results focusing on accuracy and model interpretability.

Main Results:

• White-box approaches, like Cartesian Genetic Programming and Decision Trees, successfully supported PD diagnosis and generated interpretable classification models.
• These interpretable models identified specific feature subsets related to distinct tasks for classification.
• Cartesian Genetic Programming demonstrated superior accuracy compared to other white-box methods and enhanced interpretability over black-box methods.

Conclusions:

• Machine learning, particularly Cartesian Genetic Programming, can effectively support Parkinson's disease diagnosis through handwriting analysis.
• Interpretable models derived from ML offer valuable insights for developing non-invasive, cost-effective, and easily administered diagnostic protocols.
• The findings advocate for ML approaches that balance predictive performance with clinical transparency for wider adoption in healthcare settings.