Left Atrial Strain in the Assessment of Diastolic Function in Heart Failure: A Machine Learning Approach

Background: Diastolic dysfunction (DD) assessment in heart failure is still challenging. Peak atrial longitudinal strain (PALS) is strongly related to end-diastolic pressure and prognosis, but it is still not part of standard DD assessment. We tested the hypothesis that a machine learning approach would be useful to include PALS in DD classification and refine prognostic stratification. Methods: In a derivation cohort of 864 heart failure patients in sinus rhythm (age, 66.6±12 years; heart failure with reduced ejection fraction, n=541; heart failure with mildly reduced ejection fraction, n=129; heart failure with preserved ejection fraction, n=194), machine learning techniques were retrospectively applied to PALS and guideline-recommended diastolic variables. Outcome (death/heart failure rehospitalization) of the identified DD-clusters was compared with that by guidelines-based classification. To identify the best combination of variables able to classify patients in one of the identified DD-clusters, classification and regression tree analysis was applied (with DD-clusters as dependent variable and PALS plus guidelines-recommended diastolic variables as explanatory variables). The algorithm was subsequently validated in a prospective cohort of 189 heart failure outpatients (age, 65±13 years). Results: Three distinct echocardiographic DD-clusters were identified (cluster-1, n=212; cluster-2, n=376; cluster-3 DD, n=276), with modest agreement with guidelines-recommended classification (kappa=0.40; P <0.001). DD-clusters were predicted by a simple algorithm including E/A ratio, left atrial volume index, E/e′ ratio, and PALS. After 36.5±29.4 months follow-up, 318 events occurred. Compared to guideline-based classification, DD-clusters showed a better association with events in multivariable models (C-index 0.720 versus 0.733, P =0.033; net reclassification improvement 0.166 [95% CI, 0.035–0.276], P =0.013), without interaction with ejection fraction category. In the validation cohort (median follow-up: 18.5 months), cluster-based classification better predicted outcome than guideline-based classification (C-index 0.80 versus 0.78, P =0.093). Conclusions: Integrating PALS by machine learning algorithm in DD classification improves risk stratification over recommended current criteria, regardless of ejection fraction status. This proof of concept study needs further validation of the proposed algorithm to assess generalizability to other populations.