Deep Learning-Based Prediction of Heart Failure Readmission Using Longitudinal Cardiac Device Data

Heart failure (HF) is a leading cause of hospitalization and readmission. Cardiac implantable electronic devices (CIEDs) continuously capture physiologic diagnostics that change during and after HF decompensation and may enable early post-discharge risk stratification.

To predict 60-day HF readmission using deep learning models trained on post-discharge temporal behavior of CIED diagnostic parameters.

We performed a retrospective analysis of patients with implantable cardioverter defibrillators (ICDs) or cardiac resynchronization therapy defibrillators (CRT-Ds) devices enrolled in remote monitoring (2007-2021) with HF hospitalizations identified from linked electronic health record and remote monitoring databases. Five daily device diagnostics (activity, impedance, heart rate variability, and night/day heart rates) were aligned to the index HF discharge (day 0) and analyzed over the subsequent 30 days. Patients readmitted within 7 days and those with substantial missing device data were excluded; remaining recovery windows were labeled by the occurrence of HF readmission 8-60 days post-discharge. We trained and evaluated two architectures of neural networks.

The cohort included 5734 patients with 12 369 HF hospitalization events; 2489 index hospitalizations were followed by HF readmission within 60 days. Across the 30-day recovery window of device parameters, temporal characteristics of CIED cardiac compass parameters differed significantly between readmitted and non-readmitted patients (p < 0.001), consistent with slower physiologic recovery among readmitted events. On an independent data set, the 1-dimensional convolutional neural network architecture (1D CNN) achieved 0.77 sensitivity, 0.98 specificity, and area under the ROC curve (AUROC) of 0.89.

Deep learning models leveraging post-discharge trajectories of routinely captured CIED diagnostics can identify patients at higher risk of 60-day HF readmission and may support targeted early follow-up after HF hospitalization.