Electronic Health Records (EHRs) [1] are a comprehensive repository of patients’ health histories, medical visits, and related data. They are gradually transforming the traditional healthcare service model and significantly enhancing the operational efficiency of healthcare delivery [2]. Effectively extracting and utilizing the valuable information in EHR data is crucial for advancing personalized and precision healthcare services [3], [4], [5]. By analyzing patients’ historical admission records, we can predict their future health status, which provides practical patient treatment guidance [6], [7], [8], [9], [10], [11].

Recent advancements in deep learning have enhanced the development of existing methods that focus on extracting temporal disease progression patterns from structured EHR data [8], [12], [13], [14] for predicting clinical diagnostic outcomes. The success of Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and transformer frameworks in handling temporal information and text has inspired the creation of high-performing EHR prediction methods based on these techniques. However, challenges still exist in health event prediction using EHR data.

First, sparse and infrequent diagnosis codes in EHR data and their complex interconnections have complicated effective representation learning. This study uses more than 2000 diagnosis codes, fewer than 10 times, while fewer than 300 codes have frequencies exceeding 1000 times. Moreover, certain diseases, such as hypertension and heart disease, often co-occur in the same patients, leading to potential interdependencies between specific diagnosis codes due to this comorbidity. Existing approaches [7], [15], [16], [17] attempt to integrate medical ontology or large-scale knowledge graphs to capture code representation. However, these methods face challenges regarding the mismatch between diagnosis codes and knowledge graph entities. Neglecting comorbid relationships limits the complex interactions and understanding among diseases and undermines the reliability of disease code embedding representations.

Second, EHR admission records exhibit imbalance and temporal irregularity, complicating accurately modeling patient trajectories and meaningful patterns over time. Moreover, there is a significant imbalance in the frequency of patient admission. For instance, Patient 1 had four hospital admissions between December 2015 and December 2016, while Patient 2 had five admissions during the same period, as shown in Fig. 1. In addition, the timing of admission records can be irregular due to various factors, including disease outbreaks, the severity of illness, and individual patient circumstances. The time interval between adjacent admission records of these two patients ranged from 36 to 183 days. Therefore, common approaches include data normalization and feature engineering to optimize the analysis and address the above problems. The study employs time series analysis and deep learning models to model patients’ historical admission records [13], [18], [19], [20]. Using RNNs is the most straightforward approach [20], [21], [22], [23], [24], exploiting the decay of temporal information to simulate disease progression patterns. However, RNNs have limitations when handling long EHR sequences. Moreover, predicting chronic diseases and simple time decay models may not adequately capture the disease complexity.

This study proposes a temporal health event network prediction model, GLT-Net, based on Graph Learning and the Transformer framework to address the challenges of EHR data and consider the advantages of the Transformer framework and attention mechanism in processing temporal information and variable-length text. First, we leverage the inherent hierarchical information from medical ontologies to represent diagnosis codes and pre-train their initial representations effectively. Afterward, we construct a diagnosis code association graph and use graph neural networks (GNNs) to account for comorbidity associations between diseases and enhance these learning representations. In addition, we utilize a temporal Transformer to model patients’ admission histories and process admission records. This process introduces a self-attention mechanism that allocates attention weights based on the significance of previous admissions, thereby evaluating their impact. We also designed a temporal encoder to utilize temporal information better and represent the time intervals between admissions. This study presents the following contributions:

First, we developed an unsupervised learning method to derive patient representations based on patient demographic information. The representation of diagnosis codes is effectively learned using the hierarchical structure of the codes and the comorbidity relationships between diseases.

Second, we employ a Transformer to model patients’ admission records, incorporating temporal information to effectively learn patients’ latent representations and assess the impact of previous admissions. This approach delves into the relationships between adjacent admissions and incorporates a self-attention mechanism to allocate attention weights based on their significance.

Third, we performed comprehensive experiments on real-world datasets to evaluate the proposed GLT-Net model. These experiments revealed that the GLT-Net model outperformed existing models in temporal health event prediction and confirmed the method’s effectiveness.

Statement of significance