MPRE: Multi-perspective Patient Representation Extractor for Disease Prediction

TL;DR

The paper tackles disease prediction from sparse EHR data by introducing MPRE, a framework that separates and learns trend and variation in dynamic features through a Frequency Transformation Module based on symlets wavelets. A 2D Multi-Extraction Network then reshapes these signals into a 2D tensor to capture adjacencies and correlations across short and long horizons, while First-Order Difference Attention Mechanism adaptively weighs differences between adjacent variations. Static features are embedded and fused with dynamic representations for final prediction, trained with cross-entropy loss and Adam optimization. Empirical results on the SCRIPT CarpeDiem and Health Facts datasets show MPRE outperforms strong baselines in AUROC and AUPRC, with ablation and symmetry analyses validating the contributions of FTM, 2D MEN, and FODAM. The work demonstrates that time-frequency learned representations and multi-perspective correlation modeling can effectively overcome data sparsity in EHR-based disease prediction, offering improved diagnostic guidance signals for clinical use.

Abstract

Patient representation learning based on electronic health records (EHR) is a critical task for disease prediction. This task aims to effectively extract useful information on dynamic features. Although various existing works have achieved remarkable progress, the model performance can be further improved by fully extracting the trends, variations, and the correlation between the trends and variations in dynamic features. In addition, sparse visit records limit the performance of deep learning models. To address these issues, we propose the Multi-perspective Patient Representation Extractor (MPRE) for disease prediction. Specifically, we propose Frequency Transformation Module (FTM) to extract the trend and variation information of dynamic features in the time-frequency domain, which can enhance the feature representation. In the 2D Multi-Extraction Network (2D MEN), we form the 2D temporal tensor based on trend and variation. Then, the correlations between trend and variation are captured by the proposed dilated operation. Moreover, we propose the First-Order Difference Attention Mechanism (FODAM) to calculate the contributions of differences in adjacent variations to the disease diagnosis adaptively. To evaluate the performance of MPRE and baseline methods, we conduct extensive experiments on two real-world public datasets. The experiment results show that MPRE outperforms state-of-the-art baseline methods in terms of AUROC and AUPRC.

Figures (4)

• Figure 1: The overview of MPRE. Specifically, we adopt $\#$ as a specific dynamic feature across all visit records, and $*$ means the one specific dynamic feature. FTM is used to capture trend and variation information from the dynamic features separately. 2D MEN reshapes the trend and variation as the 2D temporal tensor and further captures the adjacent, short, and long-term correlation between the trend and variation. Fusion is used to embed all dynamic feature representations based on the outputs from 2D MEN. FODAM adaptively computes the contribution of variation differences. Static features are embedded by the linear layer. Finally, the prediction module aggregates the output of fusion, the result of FODAM, and the embedded static features to perform disease prediction.
• Figure 2: The performance of ablation studies in terms of AUROC and AUPRC. (a) shows the average performance on SCRIPT CarpeDiem Dataset. (b) presents the average performance on Health Facts Database.
• Figure 3: The performance of symlets with different vanishing moments in terms of AUPRC and AUROC.
• Figure 4: The attention scores for differences of adjacent variation in two dynamic features, i.e., systolic blood pressure and neutrophils.