Abstract 13836: Unbiased Deep Learning Approach Utilizing Longitudinal Data in Assessing All-Cause Mortality in Patients With a De Novo or Worsened Heart Failure

• Published in: Circulation (New York, N.Y.) N.Y.), 2021-11, Vol.144 (Suppl_1), p.A13836-A13836
• Main Authors: Herman, Robert, Vanderheyden, Marc, Vavrik, Boris, Beles, Monika, Palus, Timotej, Kepesiova, Zuzana, Goethals, Marc, Verstreken, Sofie, Dierckx, Riet, Heggermont, Ward, Bartunek, Jozef
• Format: Article
• Language: English
• ISSN: 0009-7322; 1524-4539
• DOI: 10.1161/circ.144.suppl_1.13836

IntroductionHeart failure (HF) is a heterogenous syndrome with complex pathophysiology. Biomarkers and clinical risk scores often fail to capture modifications in the treatment continuity and provide suboptimal patient-level precision in the prognostic stratification. Electronic patient records provide necessary granularity yielding opportunities to develop new artificial intelligence (AI) based strategies for comprehensive prognostic re-stratification. HypothesisWe assessed the hypothesis that, utilizing longitudinal patient data in an AI approach, yields superior performance predicting all-cause mortality in a cohort of patients hospitalized with a de novo or worsened HF, compared to single observational time point predictions. MethodsIn a cohort of 2449 HF patients hospitalized between 2011-2017, we utilized 151 451 patient exams from 422 parameters. Features included clinical phenotyping, medication, ECG, laboratory, echocardiography, catheterization data or percutaneous and surgical interventions gathered on a routine clinical basis reflecting standard of care as captured in individual electronic records. AI models were developed, and their performance on the validation set was compared to industry standard clinical scores. ResultsAI models yielded performance ranging from 0.83 to 0.89 AUC on the outcome-balanced validation set in predicting all-cause mortality at 30-, 90-, 180-, 360- and 720-day time-limits. The primary endpoint, 1-year mortality prediction model, recorded 0.85 AUC on the validation set compared to 0.7 AUC (Seattle HF model) and 0.73 AUC (MAGGIC HF Score) respectively. ConclusionsOur findings present a novel, patient-level, AI-based risk prediction approach of all-cause mortality in heart failure utilizing all historical data available in electronic health records. This suggests the potential of AI based predictive models in a point-of-care approach to guide clinical risk stratification.