Overview
The transfer of longitudinal Electronic Health Record (EHR) representations across different healthcare systems presents significant challenges due to inherent variations in clinical coding practices, patient populations, and healthcare workflows across institutions and countries. To address this, a research initiative introduced Asymmetric Supervised Contrastive Learning (Asymmetric SupCon), a task-specific pre-training objective. This objective is designed to handle the heterogeneity of negative clinical outcomes by clustering patients who share a specific positive outcome without explicitly drawing negative trajectories closer together. The methodology also incorporated a hybrid semantic mapping pipeline to facilitate transfer across diverse clinical vocabularies.
Research Context
Cross-system transfer of longitudinal EHR representations is inherently challenging. The primary factors contributing to this difficulty include substantial differences in clinical coding standards, the demographic and clinical characteristics of patient populations, and the operational workflows within healthcare systems, both nationally and internationally. The development of an effective method to bridge these disparities is crucial for leveraging large-scale EHR datasets from varied sources for predictive modeling and other applications.
Approach
The research employed Asymmetric SupCon as a task-specific pre-training objective. This objective is motivated by the understanding that negative clinical outcomes exhibit high heterogeneity. Its design allows for the clustering of patients who experience a target positive outcome, while not actively drawing disparate negative trajectories towards a common representation. Temporal Transformer encoders were selected for pre-training. These encoders were trained on longitudinal records sourced from 3.98 million patients within the Taiwanese National Health Insurance Research Database (NHIRD). Following pre-training, these models were transferred to two distinct U.S. EHR datasets: MIMIC-IV and EHRSHOT. To manage the heterogeneity present in clinical vocabularies between the source (NHIRD) and target (U.S. datasets), a hybrid semantic mapping pipeline was developed. This pipeline combined direct mapping techniques with embedding-based retrieval methods.
Findings
- NHIRD pre-training consistently improved performance on MIMIC-IV when compared to models initialized randomly. This pre-training also substantially reduced the performance gap when compared to task-specific in-domain pre-training on MIMIC-IV.
- Transferred models demonstrated particularly strong few-shot performance for incident disease prediction when evaluated on the EHRSHOT dataset.
- A controlled objective ablation study was conducted under matched pre-training scales. This comparison indicated that Asymmetric SupCon achieved a higher mean Area Under the Precision-Recall Curve (AUPRC) than direct supervised Binary Cross-Entropy (BCE) transfer across all four evaluated tasks.
- In the same ablation study, Asymmetric SupCon also outperformed Standard SupCon on three out of four tasks. For the remaining task (readmission), Asymmetric SupCon exhibited a 0.003 AUPRC deficit compared to Standard SupCon.
Why This Matters
The described approach, utilizing asymmetric contrastive pre-training, is presented as an effective method for achieving task-specific cross-national EHR representation transfer. This capability allows for the utilization of diverse and large-scale EHR datasets across different national healthcare systems, potentially enhancing predictive model generalizability and performance, particularly in scenarios with limited in-domain data.