Unsupervised machine learning generated clusters of left ventricular strain curves identifies patients in risk of heart failure and cardiovascular death following acute myocardial infarction. (25th November 2020)

• Record Type: Journal Article
• Authors: Simonsen, J.O; Skaarup, K.G; Djernaes, K; Modin, D; Lassen, M.C.H; Grove, G.L; Pedersen, S; Estepar, R.S.J; Martinez, S.S; Gislason, G; Biering-Soernsen, T
• Is Part Of: European heart journal. Volume 41:(2020)Supplement 2
• Publication Date: 2020-11-25
• Subjects: Tissue Doppler, Speckle Tracking and Strain Imaging; Cardiology -- Periodicals; Heart -- Diseases -- Periodicals; 616.12005
• DOI: 10.1093/ehjci/ehaa946.0094
• Languages: English
• ISSNs: 0195-668X

Abstract

Abstract: Background: Today myocardial deformation, also known as strain, is assessed by the global longitudinal strain (GLS) which only provides information about the maximal deformation during systole. Hence, a lot of information obtained from different patterns of deformation curves might be undiscovered. Unsupervised Machine leaning (uML) is capable of identifying similar patterns of deformation curves. Identifying different phenotypical patterns from myocardial deformation curves might provide insights into the pathophysiological development of cardiac disease and entail useful prognostic information. Purpose: To investigate whether uML can group specific patterns of myocardial deformation curves which provide prognostic information on heart failure and/or cardiovascular death (HF/CVD) following ST-segment elevation myocardial infarction (STEMI). Methods: A total of 319 STEMI patients had an echocardiogram performed at median 2 days after primary percutaneous coronary intervention (pPCI). Speckle tracking echocardiography analysis divided the left ventricle into 18 segments. Standardisation of the cardiac cycle was done using linear interpolation and complete strain data (mean of all segments) as function of time throughout the cardiac cycle was used as input for the uML algorithm. Clusters were identified using a K-means cluster analysis algorithm. Primary endpoint was the composite of heart failure (HF) and/or cardiovascular death (CVD). Median follow-up time was 1423 days (IQR: 91; 1660). Results: Mean age was 62 years, 75% were male and 130 (41%) suffered incident HF/CVD during follow-up. The uML algorithm grouped patients into three clusters containing 97, 104, and 118 patients respectively. GLS curves of the three clusters are illustrated in the Figure 1. Incidence of HF/CVD increased significantly from cluster 1 through 3 (24% vs. 39% vs. 60%, P<0.001). In multivariable Cox regressions adjusting for the variables in the score risk chart model all three clusters were significantly associated with future HF/CVD (Figure 1). Cluster models provided significant incremental prognostic information when comparing C-statistics (0.64 vs. 0.62, p=0.029) Conclusion: Unsupervised Machine Learning clusters of left ventricular deformation curves identifies patients in risk of HF/CVD following STEMI treated with pPCI, and provides incremental prognostic information to the score risk chart model. Funding Acknowledgement: Type of funding source: None