Improving the compromise between accuracy, interpretability and personalization of rule-based machine learning in...

Area of Science:

• Machine Learning
• Clinical Informatics
• Artificial Intelligence

Background:

• Developing predictive models requires clear domain knowledge and cause-effect relationships.
• Decision rules are valuable in clinical practice for building intelligent prediction models.
• Existing methods like decision trees and ensembles of trees (ETs) face a trade-off between interpretability and predictive performance.

Purpose of the Study:

• To introduce a novel component for personalizing decision rule selection in predictive modeling.
• To improve the trade-off between interpretability and predictive performance in machine learning models.
• To enhance the accuracy of selected decision rules for individual patient predictions.

Main Methods:

• Developed a new component to predict the correctness of decision rules for individual patients.
• Integrated this component into existing methodologies for simplifying ensembles of trees (ETs).
• Validated the approach using three public clinical datasets.

Main Results:

• The novel component introduces personalization into the decision rule selection process.
• Validation results indicate an increase in the predictive performance of the selected rule sets.
• The proposed method effectively improves the interpretability-predictive performance trade-off.

Conclusions:

• The personalized approach to decision rules offers a significant advancement in clinical predictive modeling.
• This method enhances the accuracy and interpretability of machine learning models in healthcare.
• The findings suggest a promising direction for developing more effective and understandable AI tools in clinical practice.