Using recurrent neural network models for early detection of heart failure onset

Abstract

Objective: We explored whether use of deep learning to model temporal relations among events in electronic health records (EHRs) would improve model performance in predicting initial diagnosis of heart failure (HF) compared to conventional methods that ignore temporality.

Materials and methods: Data were from a health system's EHR on 3884 incident HF cases and 28 903 controls, identified as primary care patients, between May 16, 2000, and May 23, 2013. Recurrent neural network (RNN) models using gated recurrent units (GRUs) were adapted to detect relations among time-stamped events (eg, disease diagnosis, medication orders, procedure orders, etc.) with a 12- to 18-month observation window of cases and controls. Model performance metrics were compared to regularized logistic regression, neural network, support vector machine, and K-nearest neighbor classifier approaches.

Results: Using a 12-month observation window, the area under the curve (AUC) for the RNN model was 0.777, compared to AUCs for logistic regression (0.747), multilayer perceptron (MLP) with 1 hidden layer (0.765), support vector machine (SVM) (0.743), and K-nearest neighbor (KNN) (0.730). When using an 18-month observation window, the AUC for the RNN model increased to 0.883 and was significantly higher than the 0.834 AUC for the best of the baseline methods (MLP).

Conclusion: Deep learning models adapted to leverage temporal relations appear to improve performance of models for detection of incident heart failure with a short observation window of 12-18 months.

Keywords: deep learning; electronic health records; heart failure prediction; patient progression model; recurrent neural network.

Figure 1.

(A) One-hot vector encoding of clinical events. t in (B) indicates the time at which the event occurs, assuming we make the prediction at time t₇. We appended the time duration feature at the end of the one-hot vector, as shown in (C).

Figure 2.

The GRU model architecture (A), and building blocks (B). Note that vectors are denoted in bold lowercase, matrices in bold uppercase, and scalars in plain lowercase letters.

Figure 3.

Two experimental settings where we alternately changed the length of the prediction window and the observation window. In (A), the prediction window was fixed at 6 months, while we varied the length of the observation window. In (B), the observation length was fixed at 9 months, while we varied the length of the prediction window.

Figure 4.

Heart failure prediction performance of the GRU and baseline models. All models were trained and tested using the dataset created from the 12-month observation window and the 6-month prediction window. The values of the AUC and the standard error are provided in the supplementary section.

Figure 5.

Heart failure prediction performance of the GRU and baseline models. All models were trained and tested using the dataset created from the 18-month observation window and 0-month prediction window. The values of the AUC and the standard error are provided in the supplementary section.

Figure 6.

Training time vs number of patients.