An interpretable data-driven approach for rules construction: Application to cardiovascular risk assessment.

Area of Science:

• Clinical Decision Support Systems
• Machine Learning in Healthcare
• Biostatistics

Background:

• Clinical decision-making requires understandable models to support healthcare professionals.
• Extracting and integrating knowledge from datasets is crucial for enhancing clinical evidence.
• Interpretable models are essential for building trust and facilitating adoption in clinical practice.

Purpose of the Study:

• To develop interpretable models for patient condition assessment using supervised clustering.
• To discover key features representing specific patient conditions.
• To formulate simple, interpretable rules for clinical application, specifically in coronary artery disease (CAD) risk stratification.

Main Methods:

• Development of interpretable models based on supervised clustering theories.
• Feature discovery to identify the most representative patient condition indicators.
• Rule formulation for clear and understandable clinical insights.
• Application and validation on a real-world dataset of Acute Coronary Syndromes (ACS) patients.

Main Results:

• The proposed methodology generates simple and interpretable rules for clinical use.
• Achieved a 7% improvement in geometric mean performance compared to the GRACE model for 30-day mortality risk assessment in CAD patients.
• Demonstrated the effectiveness of interpretable models in a practical clinical setting.

Conclusions:

• The developed methodology provides an effective approach to interpretable patient condition assessment.
• The interpretable models offer a valuable alternative for clinical decision support, particularly in cardiovascular risk prediction.
• The findings suggest potential for broader application of this methodology across various clinical domains.