MLP combining clinical data and segmental strain improved prediction of adverse remodeling post-MI with AUC 0.83, outperforming clinical (0.55) or strain alone (0.42).
Does a machine learning model combining clinical and CMR strain data improve the prediction of left ventricular adverse remodeling in STEMI patients?
Machine learning combining clinical and CMR strain data provides high predictive accuracy for left ventricular adverse remodeling 6 months after STEMI.
Absolute Event Rate: 0% vs 0%
Abstract Background Adverse remodeling is significantly associated with heart failure, and strains derived from Cardiovascular magnetic resonance (CMR) serves as a sensitive detector in predicting this pathophysiological process after acute myocardial infarction (MI). However, strategies to identify individuals at risk of adverse remodeling focusing on segmental strains, especially strains of remote myocardium are lacking, we therefore aimed to investigate the predictive value of functional early change in infarcted, viable and remote myocardium after acute MI. Method We derived data from a cohort of 271 STEMI patients with serial CMR collected at one week and six months after. Adverse remodeling is defined as 20% increase in end-diastolic volume after 6 months of MI. According to whether the adverse remodeling occurred, stratified random sampling was used to divide 271 patients into training and testing set in 8:2 ratio. A total of 53 features (38 clinical parameters, 3 CMR traditional parameters and 12 CMR feature-tracking derived strain parameters) have been used to predict the risk of adverse remodeling. To evaluate the predictive accuracy of various machine learning techniques and feature combinations, we calculate the areas under the receiver operating characteristic curve (AUC)s, in which 1) ML with clinical features (ML-clinic), 2) ML with traditional CMR parameters (ML-LGE), and 3) ML with strain data (ML-strain), 4) ML combining clinical and traditional CMR parameters (ML-combined LGE), 5) ML combining clinical and strain data (ML-combined strain) was used as an input of five ML techniques. The SHAP (SHapley Additive exPlanations) value was used to quantity the feature importance. Results Multilayer perceptron (MLP) with residual framework perform best in ML techniques. Using MLP, addition of segmental and global strain measures to clinical data increased the AUC from 0.55 (ML-selected clinic) and 0.42 (ML-strain) to 0.83 (ML-combined strain), which is comparable to 0.73 (LGE combined model). We also found that LS in adjacent regions may serve as a more sensitive detector for adverse remodeling after a SHAP value-based comparison of global versus regional stains. Conclusion ML combined both clinical and strains were found to have high predictive accuracy for adverse remodeling, which could serve as alternative to traditional CMR parameters. Besides, we illustrated a framework for comparing the prognostic value of multiple CMR-feature tracking parameters in the context of clinical parameters.The Study Workflow SHAP feature importance from MLP model
A et al. (Sat,) reported a other. MLP combining clinical data and segmental strain improved prediction of adverse remodeling post-MI with AUC 0.83, outperforming clinical (0.55) or strain alone (0.42).