Evaluation of diastolic function: machine learning improves classification of left ventricular filling pressure

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Current left ventricular filling pressure (LVFP) classification relies on expert-defined echocardiographic algorithms.
• These algorithms can be limited by missing data and suboptimal parameter selection.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) in improving LVFP classification.
• To identify key echocardiographic parameters prioritized by ML models for LVFP assessment.

Main Methods:

• A multicenter study involving 250 patients undergoing echocardiography and heart catheterization.
• Eight ML models were trained and validated using nested cross-validation to classify LVFP.
• ML models performed parameter selection to optimize classification performance.

Main Results:

• ML models achieved classification accuracy of 82%-86%, outperforming the 2016 ASE/EACVI guidelines (81% accuracy, 13% unclassified).
• ML models handled missing parameter values effectively, classifying all patients.
• Key parameters identified by ML included mitral E/left atrial reservoir strain, log(NT-proBNP), and tricuspid regurgitation velocity.

Conclusions:

• ML significantly enhances the classification of LVFP with improved feasibility.
• The study highlights the value of less commonly used parameters in LVFP evaluation through ML-driven insights.