Learning from Longitudinal Data in Electronic Health Record and Genetic Data to Improve Cardiovascular Event Prediction

Abstract

Current approaches to predicting a cardiovascular disease (CVD) event rely on conventional risk factors and cross-sectional data. In this study, we applied machine learning and deep learning models to 10-year CVD event prediction by using longitudinal electronic health record (EHR) and genetic data. Our study cohort included 109, 490 individuals. In the first experiment, we extracted aggregated and longitudinal features from EHR. We applied logistic regression, random forests, gradient boosting trees, convolutional neural networks (CNN) and recurrent neural networks with long short-term memory (LSTM) units. In the second experiment, we applied a late-fusion approach to incorporate genetic features. We compared the performance with approaches currently utilized in routine clinical practice - American College of Cardiology and the American Heart Association (ACC/AHA) Pooled Cohort Risk Equation. Our results indicated that incorporating longitudinal feature lead to better event prediction. Combining genetic features through a late-fusion approach can further improve CVD prediction, underscoring the importance of integrating relevant genetic data whenever available.

Figure 1

AUROC and AUPRC of gold standard and machine learning/deep learning models for predicting 10-year CVD risk on 10-fold cross validation in Experiment I. The mean values of the AUROC and AUPRC and the standard error are provided in Supplementary Table 1.

Figure 2

AUROC and AUPRC of gold standard, GBT model on EHR feature and late fusion on EHR and genetic feature for predicting 10-year CVD risk on 50 iterations in Experiment II. The mean values of the AUROC and AUPRC and the standard error are provided in Supplementary Table 2.

Figure 3

Study design of experiment I. The figure illustrates how we defined the observation and prediction window. It also shows how we modeled longitudinal EHR features: i) We aggregated each feature across the 7-year observation window (e.g. median, max, min and SD of HDL); ii) we extracted each year value of each feature and concatenated the temporal values from all patients into a two-dimensional matrix for a classifier (e.g. LR, RF, GBT); we then constructed a tensor representation on temporal values from all patients for CNN and LSTM.

Figure 4

Flowchart of selecting cohort for late-fusion approach.

Figure 5

Framework for proposed late fusion approach to combine the genetic features with longitudinal EHR features.