Interval Coded Scoring: a toolbox for interpretable scoring systems.

Area of Science:

• Medical Informatics
• Machine Learning
• Biostatistics

Background:

• Clinical decision support systems (CDSS) are crucial for managing medical information overload.
• Existing CDSS often lack interpretability (black box models), hindering trust and legal responsibility.
• Traditional medical scoring systems are interpretable but may oversimplify complex data.

Purpose of the Study:

• To introduce Interval Coded Scoring (ICS) as a method to create interpretable scoring systems from training data.
• To present a toolbox for applying ICS to binary classification problems and generating risk profiles.
• To enable a balance between model complexity and performance through expert knowledge integration.

Main Methods:

• Utilizes sparse optimization via linear programming or elastic net formulations.
• Develops interpretable models with main effects and interactions.
• Incorporates semi-automatic training for expert knowledge integration.
• Assesses classification performance using accuracy, ROC curves, and calibration curves.

Main Results:

• ICS models demonstrate comparable classification and calibration performance to standard machine learning methods (Naive Bayes, SVM).
• The ICS toolbox facilitates easy application to diverse datasets with minimal code.
• Visualizations aid in manual application and expert validation of ICS models.

Conclusions:

• ICS provides a valuable, interpretable alternative to complex "black box" models in clinical decision support.
• The toolbox enhances the practical application of ICS in medical settings.
• ICS is a suitable choice when model interpretability is a primary concern.