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

Abstract

Aims Current recommendations for echocardiography-based classification of left ventricular filling pressure (LVFP) as normal or elevated, are based on an algorithm and parameter selection determined by human experts. We tested whether machine learning (ML) can improve classification of LVFP and investigated which parameters were deemed most important by different ML models.Methods and results In a multicentre study, echocardiography was performed simultaneously with, or within 24 h of, heart catheterization in 250 patients. Eight different ML models were trained and tested using a nested cross-validation procedure to classify LVFP as normal or elevated. The training included a search and selection of the most useful parameters. Performance was assessed from the test sets not seen during training. The eight ML models could classify all patients regardless of missing parameter values with accuracy ranging from 82% to 86%. The 2016 ASE/EACVI guidelines algorithm left 13% unclassified due to missing values and had an accuracy of 81% in the remaining patients. On average, the eight ML models selected 13 parameters, and left atrial strain was included in three of these. The five highest ranked parameters by the ML models were mitral E/left atrial reservoir strain, log(NT-proBNP), tricuspid regurgitation velocity, septal E/e', and E/A.Conclusion ML can improve classification of LVFP, particularly with a higher feasibility. The study unveiled less used parameters as some of the most valuable for evaluating LVFP.