PORTER: Portable Representations for Transferable Electronic Health Records

Clinical prediction models degrade across sites because differing medical and documentation policies induce distinct joint distributions over observed records and outcomes. Such variations arise from the underlying data-generating process (DGP) and cannot be directly addressed by increasing downstream model capacity alone. Instead, we focus on constructing transferable representations. We propose Portable Representations for Transferable Electronic Health Records (PORTER), a summarize-then-embed pipeline equipped with a clinical schema to reduce site-specific variation, paired with a complementary fine-tuning method that addresses variations unresolvable by the schema alone. Based on our assumptions about the data-generating process, we design the schema to guide the summarizer to preserve predictive information while removing site-induced representational noise. We then distill site-specific outcome policies using a test-site support set to account for remaining variations. Across three intensive care unit (ICU) datasets, six transfer directions, five clinical prediction tasks, four prediction heads, and four summarizers, PORTER attains the best average ranking among sixteen methods under zero-shot transfer. Furthermore, PORTER’s test-site distillation achieves the best few-shot performance, with the largest gains on tasks that perform weakest in the zero-shot setting. Our ablations show that even a simple prompt containing only the schema is competitive with the strongest baseline. We conclude that domain-specific prompting is essential for enabling LLMs to generate representations that support cross-site prediction.