RASPER trains discharge-note summaries against the downstream predictor's own loss

Unstructured discharge notes in electronic health records carry patient-specific evidence that standardized cohort-level codes cannot capture, but the signal is buried in lengthy, noisy text written without any specific prediction in mind. Summarization is the obvious mitigation, yet generic summaries tuned for fluency routinely omit decisive evidence while keeping plausible but uninformative detail.

RASPER trains the summarizer directly against the downstream clinical task: reinforcement learning from prediction feedback supplies a reward derived from the predictor's loss, and a longitudinal encoder turns structured codes into soft prompts that incorporate each patient's clinical context. Rewarding the quality of the resulting prediction pushes the summarizer to keep evidence that complements, rather than duplicates, the structured codes.

RASPER consistently outperformed strong baselines on readmission prediction and medication recommendation across MIMIC-III and MIMIC-IV, and the paper was accepted to the EMNLP 2026 main conference.

Fluency-oriented objectives have no reason to keep the one detail that flips a prediction.

The downstream predictor's loss is a direct, automatic signal of evidential usefulness.

Soft prompts from structured codes keep summaries complementary to the codes instead of repeating them.

Per-patient context matters: the same note warrants different summaries for different prediction tasks.

When a component feeds a downstream model, train it against that model's feedback instead of a generic quality metric; fluency is not usefulness.

The paper reports consistent gains over strong baselines on readmission prediction and medication recommendation across MIMIC-III and MIMIC-IV, with acceptance to the EMNLP 2026 main conference.