Post-infarct cardiac remodeling predictions with machine learning. (15th May 2022)

• Record Type: Journal Article
• Authors: Dieu, Xavier; Chabrun, Floris; Prunier, Fabrice; Angoulvant, Denis; Mewton, Nathan; Roubille, François; Reynier, Pascal; Ferre, Marc; Moal, Valérie; Cottin, Laurane; Furber, Alain; Garcia, Gabriel; Bière, Loïc; Mirebeau-Prunier, Delphine
• Is Part Of: International journal of cardiology. Volume 355(2022)
• Page Start: 1
• Page End: 4
• Publication Date: 2022-05-15
• Subjects: Left ventricular remodeling -- Myocardial infarction -- Machine learning -- Deep learning -- Neural networks -- Cardiac magnetic resonance; Cardiology -- Periodicals; Electronic journals; 616.12
• DOI: 10.1016/j.ijcard.2022.02.009
• Languages: English
• ISSNs: 0167-5273

Abstract

Abstract: Background: We sought to improve the risk prediction of 3-month left ventricular remodeling (LVR) occurrence after myocardial infarction (MI), using a machine learning approach. Methods: Patients were included from a prospective cohort study analyzing the incidence of LVR in ST-elevation MI in 443 patients that were monitored at Angers University Hospital, France. Clinical, biological and cardiac magnetic resonance (CMR) imaging data from the first week post MI were collected, and LVR was assessed with CMR at 3 month. Data were processed with a machine learning pipeline using multiple feature selection algorithms to identify the most informative variables. Results: We retrieved 133 clinical, biological and CMR imaging variables, from 379 patients with ST-elevation MI. A baseline logistic regression model using previously known variables achieved an AUC of 0.71 on the test set, with 67% sensitivity and 64% specificity. In comparison, our best predictive model was a neural network using seven variables (in order of importance): creatine kinase, mean corpuscular volume, baseline left atrial surface, history of diabetes, history of hypertension, red blood cell distribution width, and creatinine. This model achieved an AUC of 0.78 on the test set, reaching a sensitivity of 92% and a specificity of 55%, outperforming the baseline model. Conclusion: These preliminary results show the value of using an unbiased data-driven machine learning approach. We reached a higher level of sensitivity compared to traditional methods for the prediction of a 3-month post-MI LVR. Highlights: Good sensitivity for the prediction of 3-month post-MI Left Ventricular Remodeling Machine learning approach leads to better results than a traditional statistical one. Use of multiple feature selection algorithms allowed better variable selection. Known predictors of post-MI LVR were selected, validating our approach. Unexpected variables were also selected, requiring further exploration.