Machine learning applied to echocardiographic deformation curves for prediction of heart failure and cardiovascular death

BackgroundMany patterns from the entire strain and TDI curves remain uninvestigated. ObjectivesWe investigated whether applying signal analysis with supervised machine learning (ML) to strain and TDI curves would improve the prediction of heart failure (HF) or cardiovascular death (CV death) compared with conventional echocardiographic parameters, and whether it would reveal previously unrecognized prognostic signal-derived parameters. MethodsIn total, 744 novel signal-analytical parameters extracted from 18 strain curves and 6 TDI curves were evaluated using ML and compared with a baseline ML model trained on 17 conventional echocardiographic parameters.The primary endpoint was a combined endpoint of incident HF or CV death within five years. ResultsThe analysis included 2,589 subjects from the Copenhagen City Heart Study (CCHS) for training the ML model and 916 subjects from the LOOP Study for external testing.In total, 126 subjects (4.9%) met the primary endpoint in CCHS, whereas 59 subjects (6.4%) met the primary endpoint in the LOOP Study.The models incorporating signal analysis, both standalone and combined with conventional parameters, significantly increased discrimination compared to the baseline model based on conventional parameters in the external test cohort (Baseline model: AUC 0.721, CI 0.639:0.802vs. signal analytical model: AUC 0.790, CI 0.730:0.850,p = 0.031 and combined model: AUC 0.788, CI 0.727:0.849,p = 0.021). ConclusionA ML model based on signal analysis of strain and TDI curves from echocardiographic examinations modestly improved 5-year discrimination for incident HF or CV death compared to a model based on conventional echocardiographic parameters.