Towards symbolic regression for interpretable clinical decision scores

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Biology

Background:

• Medical decision-making relies on algorithms combining risk equations and rules for standardized pathways.
• Traditional symbolic regression (SR) struggles to model complex, rule-based clinical decision-making due to its focus on continuous functions.
• SR's potential for data-driven, interpretable models is promising for clinical risk score development.

Purpose of the Study:

• Introduce Brush, a novel SR algorithm designed to integrate rule-based logic into model development.
• Enable the creation of data-driven, interpretable clinical risk scores.
• Enhance the capabilities of symbolic regression for complex decision-making processes.

Main Methods:

• Developed Brush, an SR algorithm combining decision-tree-like splitting with nonlinear constant optimization.
• Integrated rule-based logic seamlessly into SR and classification models.
• Evaluated Brush on SRBench and applied it to recapitulate clinical scoring systems.

Main Results:

• Brush achieved Pareto-optimal performance on SRBench.
• Successfully recapitulated two widely used clinical scoring systems with high accuracy.
• Generated interpretable models that are simpler than those from decision trees (DTs) and random forests (RFs).

Conclusions:

• Brush enhances symbolic regression for medical decision-making by integrating rule-based logic.
• The algorithm produces accurate, interpretable, and simpler models compared to existing methods.
• Brush shows significant promise for developing next-generation data-driven clinical risk scores.