Abstract 4138176: Identifying Clinical Stability After Heart Failure Discharge: A machine learning-based decision aid based on continuous ECG and accelerometry

Introduction: Heart failure (HF) is typified by disease exacerbations resulting in high readmission rate and mortality. Continuous electrocardiographic (ECG) and activity (ACT) data offer promise of a scalable noninvasive solution that may reduce 30-day readmissions. Using a data-driven approach, we developed a machine learning (ML) model including the FDA-cleared multivariate change index (MCI) to identify HF patients at low risk for near-term decompensation. Methods: The LINK-HF clinical trials used the Prolaio system to continuously record physiological parameters using a wearable device. We used ML to build predictive models of a composite outcome of near-term (7-day) risk of emergency room use, hospitalization or death. The model was derived using data from the previously published LINK-HF1 study (2015-2016), and externally validated using data from the control arm of the LINK-HF2 trial (2021-2024). Covariates were developed in LINK-HF1 and a selection procedure used the bootstrap to identify the 100 covariates most commonly chosen when training a ML model to classify patients into high vs. low near-term risk. All combinations of the top 3 covariates were then used to develop the final model. The best-performing model in the training set was evaluated in the validation set. Results: Patient data were segmented into 7-day windows, yielding 4,243 training windows (200 [4.7%] with events) and 3,551 validation windows (109 [3.1%] with events) (Fig. 1A). The chosen model utilized covariates related to changes in heart rate, atrial fibrillation burden, and MCI to achieve an area under the receiver operating characteristic curve (AUROC) of 73% (95% Confidence Interval [CI], 69-76%) in the validation set. The model stratified patients into high and low risk, with a Hazard Ratio (HR) for the outcome of 3.9 (95% CI, 2.7-5.6; P-value, 4 x 10-13) in the validation set (Fig. 1B). Conclusions: Using multivariable wearable sensor data and ML, we developed a model able to identify HF patients at low vs. high near-term risk for decompensation. Future work will determine how this knowledge may optimize and personalize HF care, improving outcomes and efficiency in care delivery.