A Patient Similarity Network (CHDmap) to Predict Outcomes After Congenital Heart Surgery: Development and Validation...

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Evidence-based medicine (EBM) struggles with complex clinical scenarios where standard evidence is inapplicable.
• Medicine-based evidence (MBE) leverages big data and machine learning for personalized treatment recommendations based on similar patients.
• Translating the MBE framework into practical clinical tools remains a significant challenge.

Purpose of the Study:

• To technically implement the MBE conceptual framework into a functional tool.
• To evaluate the performance of this MBE tool in supporting clinical decisions for congenital heart disease (CHD) surgery outcomes.
• To develop a decision support system for personalized treatment predictions in CHD.

Main Methods:

• Collected data from 4774 CHD surgeries, extracting 66 indicators and diagnoses using natural language processing.
• Developed a patient similarity network (PSN) called CHDmap, measuring patient distances using calculation formulas and expert-modulated fusion.
• Integrated machine learning models with the PSN for personalized predictions and analogical reasoning.

Main Results:

• CHDmap achieved better prediction results than clinicians in binary and multiple classification tasks for postoperative outcomes.
• Logistic regression models using PSN-derived similar patients improved prediction performance (best AUCs of 0.810 and 0.926).
• Clinicians demonstrated substantially improved predictive capabilities with CHDmap support.

Conclusions:

• CHDmap shows competitive performance against clinical experts, even without individual optimization.
• The system enhances clinician decision-making by combining AI with human cognitive abilities.
• The MBE approach is viable for clinical practice, with significant potential for realizing personalized medicine.