Comparative performance of deep learning architectures for predicting 60-day heart failure readmission using device-derived physiologic data
V Ramos, S Sarkar, V Aaryamaan, D J Gonzalez, R A Devathu, E Godinez, K Chatterjee, M Fazal, J K Han, T Haddad, Y Cho, T BaykanerAbstract
Background
Cardiac implantable electronic devices (CIEDs) continuously record physiologic parameters reflecting heart failure (HF) recovery. Transforming these time-series signals into clinically actionable risk predictions requires model architectures capable of capturing complex temporal dynamics.
Objective
To compare two deep learning architectures, a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory (LSTM) network, for predicting 60-day HF readmission based on daily CIED diagnostic data.
Methods
Patients in the Medtronic CareLink™ database with implantable cardioverter-defibrillators or cardiac resynchronization therapy defibrillators and at least one HF hospitalization (2007–2021) were included. Temporal sequences of daily cardiac compass data (five physiologic parameters plus a discharge index) over 30 days post-discharge were analyzed. The 1D-CNN model applied convolutional filters across 7-day sliding windows to extract localized temporal features from stacked multichannel inputs (6×30 matrix per recovery window). It comprised 108,974 learnable parameters, trained for 50 epochs using cross-entropy loss and the ADAM optimizer (learning rate 0.0001), with stratified tenfold cross-validation and grid search for hyperparameter optimization. The LSTM model, built with identical input features, captured long-range temporal dependencies through gated recurrent units. Both models were trained with a 60/20/20 train/validation/test split and evaluated for binary classification of 60-day readmission.
Results
In the independent test dataset (2,494 HF events), the 1D-CNN achieved accuracy 92.8%, sensitivity 76.9%, specificity 97.8%, and AUC 0.89 (Figure, panel A). The LSTM achieved accuracy 90.9%, sensitivity 78.6%, specificity 94.7%, and AUC 0.88 (Figure, panel B). Both architectures effectively stratified recovery trajectories into high- and low-risk categories based solely on physiologic time-series data, with consistent discrimination confirmed by ROC and confusion matrix analyses.
Conclusion
Deep learning models trained on device-derived temporal data can accurately predict early heart failure readmission. Among the architectures evaluated, the one-dimensional convolutional neural network (1D-CNN) demonstrated slightly superior performance, indicating that convolutional feature extraction of short-term physiologic fluctuations may be more effective than memory-based recurrent models for near-term forecasting. Incorporating these predictive algorithms into remote monitoring systems could enable automated, real-time identification of patients at highest risk following discharge.Figure 1.Models Confusion Matrix