Self-Explaining Neural Networks for Transparent Parkinson's Disease Screening

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Biomechanical Analysis

Background:

• Transparent clinical decision-making is crucial for AI in medical diagnosis.
• Post hoc explanation methods for deep learning models lack guaranteed faithfulness to underlying reasoning.
• Parkinson's Disease (PD) diagnosis can benefit from objective gait analysis.

Purpose of the Study:

• To develop a Self-Explaining Neural Network (SENN) for Parkinson's Disease screening.
• To enforce intrinsic interpretability in deep learning models for gait analysis.
• To ensure explanations faithfully reflect the model's diagnostic reasoning.

Main Methods:

• Utilized a residual CNN backbone with stochastic depth regularization.
• Incorporated a 16-concept encoder with diversity and stability constraints.
• Employed temperature-scaled probability calibration for reliable clinical operating points.
• Evaluated on the PhysioNet Gait in Parkinson's Disease dataset using Ground Reaction Force (GRF) data.

Main Results:

• SENN achieved a subject-level ROC-AUC of 0.916, sensitivity of 0.913, and Average Precision of 0.942.
• Interpretability constraints did not significantly reduce discriminative performance compared to baselines.
• Concept-level analysis identified key biomechanical hallmarks of parkinsonian gait.

Conclusions:

• Rigorous intrinsic interpretability and competitive predictive accuracy are achievable simultaneously in deep gait analysis.
• SENN provides clinically grounded, patient-specific explanations for PD screening.
• The developed transparent AI supports potential clinical adoption for diagnostic purposes.