Authors :
Presenting Author: Zheng Han, PhD – University of Central Oklahoma
Priyadharsini Ramamurthy, BS – Oklahoma State University
Dursun Delen, PhD – Oklahoma State University
Zhuqi Miao, PhD – Oklahoma State University
Andrew Gin, MD – University of Oklahoma Health Sciences Center
William Paiva, PhD – Oklahoma State University
Rationale:
Traumatic brain injury (TBI) is a major risk factor for neurological disorders, including post-traumatic epilepsy (PTE), a debilitating condition associated with significant long-term consequences. The prognosis of PTE occurrence remains challenging due to the complex pathophysiology of PTE and the impracticality of traditional blood biomarker- or imaging-based screening for large populations. This study proposes a graph-based deep learning approach that leverages electronic health records (EHR) to enhance the predictive assessment of PTE risk.
Methods:
We utilized Oracle Real-World Data (ORWD) to construct a Heterogeneous Graph Attention Network (HeteroGAT)(the graph structure and model architecture are shown in Figure 1A&B) that contains patient and diagnosis nodes, with temporal information represented using patients-to-diagnosis edges, and comorbidity connectivity embedded using diagnosis-to-diagnosis edges. The HeteroGAT was trained on a cohort of 1,598,998 TBI-only patients and 102,687 individuals who developed epilepsy after TBI. Model performance was evaluated using sensitivity, specificity, macro F1-score, and area under the receiver operating characteristic curve (AUC-ROC), benchmarked against traditional machine learning models.
Results:
HeteroGAT significantly outperformed conventional models in PTE prediction by effectively integrating demographic data and comorbidity profiles spanning from 20 to 500 distinct conditions. The model’s multi-head attention mechanisms, in combination with learned comorbidity connectivity, enhanced its ability to capture complex dependencies within EHR data. HeteroGAT achieved an AUC-ROC of 0.80, outperforming the best-performing traditional model, random forest (AUC-ROC = 0.77). Figure 2 shows the performance metrics of HeteroGATs with or without diagnosis-diagnosis self-connection in comparison to conventional models using varying numbers of comorbidities.
Conclusions:
By modeling sparse EHR data through patient encounter embeddings, HeteroGAT effectively captures temporal and relational patterns in comorbidities critical for PTE prediction. Our findings highlight the potential of graph-based deep learning models, synergized with large-scale EHR data, in advancing personalized risk assessment, ultimately addressing the urgent need for more precise and proactive management of PTE in TBI patients.
Funding:
This work is funded by US Department of Defense CDMRP Award # HT9425-24-1-0481.