Interpretable Deep Learning Model for Longitudinal Electronic Health Records and Applications to Heart Failure Onset Prediction
Mission: To develop interpretable deep learning models on large-scale electronic health record (EHR) data and apply the proposed models to detect heart failure (HF) one to two years before usual diagnosis.
HF is a highly disabling and costly disease with a high mortality rate.1,2 There has been little progress in slowing progression of this disease in those at risk largely because it is difficult to detect HF in the pre-diagnostic stage. In part, we hypothesize that previous models have failed because the temporal sequence of events is ignored. Advances in deep learning have opened unique opportunities to address limitations to models that rely on aggregate features. EHRs, now commonplace in U.S. healthcare, have motivated substantial interest in using predictive models to detect health problems or events before they occur. However, a limitation to modeling strategies is that the data used to predict a patient outcome or a need are treated as a static ensemble of features, ignoring the fundamental importance of temporality of events and measures. Deep learning algorithms have shown tremendous success in many challenging applications such as image classification, video analysis, natural language processing, and game play. The success of deep learning in healthcare will depend on access to advanced parallel computing to large, heterogeneous, and longitudinal EHR data.
Our preliminary work has demonstrated that the application of deep learning models to EHR data improves prediction performance.3–6 However, the output from these models is not readily interpretable. One approach with complex black-box models such as recurrent neural networks (RNN) is to use their superior prediction to conduct surveillance in health systems for individuals at elevated risk, and then rely on less robust but interpretable models (e.g., logistic regression and decision trees) to facilitate care management and decision support. Alternatively, the rich output from RNN models, including information on the temporal relations of events, could be structured to be interpretable and facilitate clinical care of individual patients. We propose to develop clinical analytic software for healthcare applications by providing accurate and interpretable deep learning models that can be readily used by health systems. We propose to focus on early detection of HF onset:
Aim 1 - Prediction: To develop methods for the application of deep learning models to EHR data for early detection of HF diagnosis in individuals with HF.
● Hypothesis 1: Accounting for temporality (i.e., temporal order of clinical events) can improve accuracy of clinical prediction models such as heart failure onset.5
● Hypothesis 2: Structured and unstructured EHR data can be summarized into low-dimensional representation to enhance prediction performance. We propose to extend algorithms for processing text documents, known as word2vec, to model medical codes (e.g., diagnosis, procedure, medication) and clinical notes based on co-occurrence information within patient records.
The product of Aim 1 will include research software with algorithms for RNN and representation learning for EHR data. The models will be validated for HF onset prediction.
Aim 2 - Interpretation: To develop methods that translate interpretable results from deep learning methods for HF early detection.
● Hypothesis 3: Clinical risk factors and their timing offer the means to provide intuitive clinical interpretation of the deep learning models. We propose to develop a two-level neural attention model and an interactive visualization that detects influential past encounters and significant clinical variables within those encounters (e.g., key diagnoses).
Aim 2 will contribute interpretable deep learning algorithms into the software. Those algorithms will be validated on HF onset prediction.
Collaboration with Sutter Health: