Characterization of cardiac resynchronization therapy response through machine learning and personalized models.

Area of Science:

• Cardiology
• Biomedical Engineering
• Computational Medicine

Background:

• Cardiac resynchronization therapy (CRT) selection for heart failure (HF) patients faces challenges with a ~30% non-responder rate.
• Current guidelines for CRT patient selection have limitations in identifying optimal candidates.

Purpose of the Study:

• To develop a novel hybrid approach integrating machine learning and personalized models for HF patient phenogrouping.
• To predict CRT response in HF patients using explainable phenogroups.

Main Methods:

• Generated personalized models from preoperative CRT patient strain curves.
• Utilized clustering for phenotype identification and random forest for CRT response classification.
• Analyzed feature importance for predicting patient response.

Main Results:

• Achieved high accuracy in simulating myocardial strain curves (RMSE 4.48%).
• Identified five distinct HF patient phenogroups with varying CRT response rates (52%-94%).
• Random forest classification yielded an AUC of 0.86, highlighting regional contractility, viability, and electrical delays as key predictors.

Conclusions:

• Patient-specific model parameter analysis offers explainable insights into HF phenogroups and CRT response.
• The hybrid approach shows promise for enhancing HF patient characterization and CRT selection.
• Improved understanding of left ventricular mechanical dyssynchrony aids in personalized treatment strategies.