4302Machine-learning integration of complex echocardiographic patterns and clinical parameters from cohorts and trials

• Published in: European heart journal 2019-10, Vol.40 (Supplement_1)
• Main Authors: Cikes, M, Sanchez Martinez, S, Claggett, B, Solomon, S D, Bijnens, B
• Format: Article
• Language: English
• ISSN: 0195-668X; 1522-9645
• DOI: 10.1093/eurheartj/ehz745.0147

Abstract Introduction Contemporary machine learning (ML) allows comparing individuals based on a wide range of heterogenous data. We aimed to use similarity-based unsupervised ML to integrate clinical variables and complex imaging patterns from a population-based cohort (Atherosclerosis Risk in Communities study - ARIC) with a HFPEF (Treatment of Preserved Cardiac Function Heart Failure with an Aldosterone Antagonist - TOPCAT) as well as a HFREF (Multicenter Automatic Defibrillator Implantation Trial with Cardiac Resynchronization Therapy - MADIT-CRT) clinical trial and to predict the outcome of the patients based on their location in the derived ML universe, rather than on their original cohort. Methods An unsupervised ML for dimensionality reduction algorithm (multiple kernel learning) was trained using 2123 subjects enrolled in ARIC. The algorithm positioned them based on similarities in longitudinal strain patterns of 12 LV regions and an LV volume curve, together with 17 clinical parameters (age, sex, myocardial infarct, QRS width, prior HF, LVEDVi, LAVi, race, heart rate, SBP, BP meds, diabetes, smoking status, BMI, GLS, LVMi, LVEF). Using the same descriptors, 105 untreated TOPCAT and 429 untreated MADIT-CRT patients were projected onto the learned space (upper-left figure). A k-means algorithm was subsequently used to partition the resulting space into 3 clusters (broadly speaking, the number of expected clinical categories: controls (ARIC), HFPEFs (TOPCAT) and HFREFs (MADIT-CRT) (upper-right figure). Finally, cluster-wise event rates (composite of death or HF hospitalization) were computed for the combination of the two trials (bottom-left figure) and only considering subjects from ARIC (bottom-right figure). Results The visualisation of the low-dimensional space (upper-left figure) showed a clear aggregation of individuals according to the study, highlighting their different clinical and deformation properties. The individuals located around the MADIT-CRT patients (predominantly in cluster 3) showed the highest event rate (26.6%, similar to the MADIT-CRT event rate 25.4%) while those around the majority of the ARIC cohort showed the lowest rates (5.3%, similar to the ARIC event rate 5.7%). The MADIT-CRT-like ARIC individuals showed a much higher event-rate compared to the one expected from the cohort (21.1% vs 5.7%). Figure 1 Conclusion Our results serve as a proof-of-concept that unsupervised ML based analysis of complex clinical and imaging dat