Temporal Phenotype Matrix Engineering for Electronic Health Records–Enhancing Coronary Artery Disease Prediction

Nowadays, most studies still aggregate electronic health records (EHRs) into one record per patient for analysis and model development without considering temporal information, which is valuable for disease progression and outcome prediction. However, EHRs often exhibit sparsity and irregularity due to their inherent nature, and data preprocessing is needed to extract temporal information in EHRs. It is crucial to consider that imputation and aggregation techniques used during EHRs preprocessing can introduce artificial and unrealistic data, potentially leading to the loss of critical information. In this study, we proposed a temporal phenotype matrix engineering approach with auxiliary data layers (ADL) to extract important hidden information from EHRs. Our proposed approach was applied to the early prediction of coronary artery disease (CAD), one of the leading causes of death worldwide. We evaluated the performance of the long short term memory network (LSTM), convolutional neural network (CNN), and temporal convolution network (TCN) models on the CAD prediction task. Upon applying our proposed matrix engineering technique with ADL, we observed a substantial improvement, with an AUROC (area under the receiver operating characteristic) score of 0.919 ± 0.006 (a 10% increase, compared to when no ADL was included, 0.831 ± 0.011) in CNN model. In conclusion, this study highlights the benefits of the proposed temporal phenotype matrix engineering approach with ADL to address the sparsity and irregularity inherent in EHRs data.Clinical Relevance—Our findings underscore the potential of the proposed temporal phenotype matrix engineering approach with ADL for enhancing the early prediction of CAD, thereby contributing to improved patient outcomes and reduced mortality rates.