SymScore: Machine learning accuracy meets transparency in a symbolic regression-based clinical score generator

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Self-report questionnaires are vital for disease risk assessment but lengthy questionnaires can impact data quality.
• Machine learning (ML) shortened questionnaires offer accuracy but lack transparency and require specialized expertise, hindering clinical adoption.
• Healthcare professionals need interpretable tools for trust and effective decision-making in clinical workflows.

Purpose of the Study:

• To develop a novel method, SymScore, that generates interpretable score tables for shortened clinical questionnaires.
• To preserve both predictive accuracy and clinical interpretability in risk assessment tools.
• To provide a user-friendly alternative to complex ML models for healthcare professionals.

Main Methods:

• SymScore utilizes symbolic regression to generate score tables for shortened questionnaires.
• It optimally groups responses, assigns weights based on predictive importance, and applies constraints.
• Performance was compared against established ML-based shortened questionnaires (MCQI-6, SLEEPS) for sleep disorder assessment.

Main Results:

• SymScore demonstrated comparable performance to MCQI-6 (MAE=10.73, R²=0.77 vs. MAE=9.94, R²=0.82).
• Achieved high AUROC values (0.85-0.91) for sleep disorders, closely matching SLEEPS (0.88-0.94).
• Generated interpretable score tables that clinicians can easily understand and trust.

Conclusions:

• SymScore offers an accurate and interpretable solution for clinical risk assessment using shortened questionnaires.
• It advances explainable AI in healthcare by bridging the gap between predictive accuracy and clinical usability.
• SymScore enhances workflow efficiency and patient outcomes by providing a resource-efficient, trustworthy tool for clinicians.