Deep learning for mortality prediction in patients with a de-novo or worsened heart failure

Overview

This study used deep learning on routine electronic patient records to predict all-cause mortality in 2,449 patients hospitalized with de novo or worsened heart failure between 2011 and 2017. The models drew on 151,451 patient exams covering 422 parameters, including clinical phenotyping, medication, ECG, laboratory, echocardiography and catheterization data, and percutaneous and surgical interventions. On an outcome-balanced validation set, the models predicted mortality at time limits from 30 days to 2 years with AUCs of 0.83 to 0.89, and the primary 1-year mortality model reached an AUC of 0.85. Performance was stable across heart failure phenotypes (HFpEF 0.83, HFmrEF 0.85 and HFrEF 0.86).

Published in: ESC Congress 2021 – The Digital Experience; abstract published in European Heart Journal, Volume 42, Supplement 1
Published on: 27 August 2021

Background

Heart failure is a heterogeneous syndrome with complex pathophysiology. Biomarkers and clinical risk scores often fail to provide optimal patient-level precision in prognostic stratification: because they rely on a single observational timepoint, they do not capture the entire care pathway or the variations in individual patient management. Electronic patient records offer an opportunity to develop new artificial intelligence (AI) strategies for comprehensive prognostic re-stratification that reflect diagnostic and therapeutic management. The authors set out to use deep learning to develop an unbiased predictive algorithm for all-cause mortality in patients hospitalized with de novo or worsened heart failure.

Methods

The cohort included 2,449 patients with heart failure hospitalized between 2011 and 2017. The authors used 151,451 patient exams covering 422 parameters, including clinical phenotyping, medication, ECG, laboratory, echocardiography and catheterization data, and percutaneous and surgical interventions, all gathered on a routine clinical basis as standard of care and captured in individual electronic records. Model development consisted of 101 iterations of repeated random subsampling splits into balanced training and validation sets. Models were trained to predict all-cause mortality at 30, 90, 180, 360 and 720 days, with 1-year mortality as the primary endpoint.

Results

On the outcome-balanced validation set, the AI models predicted all-cause mortality at the 30-, 90-, 180-, 360- and 720-day time limits with AUCs ranging from 0.83 to 0.89. The primary endpoint model, predicting 1-year mortality, reached an AUC of 0.85. Model performance was stable across all heart failure phenotypes: AUC 0.83 in heart failure with preserved ejection fraction (HFpEF), 0.85 in heart failure with mildly reduced ejection fraction (HFmrEF) and 0.86 in heart failure with reduced ejection fraction (HFrEF).

Conclusion

These findings present a novel, patient-level, AI-based prediction of all-cause mortality in heart failure with robust accuracy across its phenotypes, suggesting the potential of AI-based predictive models in a point-of-care approach to guide clinical risk stratification.