Bridging the EHR Divide: Asymmetric Contrastive Learning for Cross-National Medical Representation Transfer

Author: Qingyang Zhang

Clinical prediction tasks often have heterogeneous negative populations: patients who do not develop a particular disease may still have very different clinical histories. I introduce Asymmetric Supervised Contrastive Learning (Asymmetric SupCon), a task-specific pre-training objective that brings together patient trajectories sharing a positive future outcome without explicitly attracting negative trajectories to one another.

The study pre-trains temporal Transformer encoders using a Taiwanese NHIRD cohort of 3.98 million patients, with up to five years of history per training example, and evaluates transfer to MIMIC-IV and EHRSHOT. To bridge differences in clinical vocabularies, a hybrid alignment pipeline combines direct code mappings with semantic retrieval using Qwen3-Embedding-8B and FAISS.

On MIMIC-IV, NHIRD pre-training improves over random initialization across all three evaluated tasks. At EHRSHOT’s k=16 setting, the transferred models exceed the strongest reported baseline on all five incident-disease prediction tasks, including Acute MI (AUPRC 0.153 vs. 0.118). A controlled ablation on approximately 400K NHIRD patients achieves the highest mean AUPRC on three of four tasks, supporting the asymmetric objective’s value for cross-national transfer.