CLIN-SUMM: Incremental Longitudinal Summarization of Clinical Notes Enables Scalable Representation and Early Disease Prediction

CLIN-SUMM restructures longitudinal clinical notes for 12,356 Massachusetts General Hospital patients, cutting note text by 75% while preserving clinician-rated correctness and completeness.

• Why it matters: EHR notes are fragmented and redundant, limiting scalable modeling and making it hard to reuse narrative information for longitudinal prediction without losing temporal order.
• What they did: The framework incrementally builds date-partitioned, categorized patient representations by summarizing only newly documented content at each encounter, then reuses the standardized outputs for downstream models.
• The result: For dementia, fine-tuning Clinical ModernBERT on CLIN-SUMM summaries gives AUROC 0.84 for diagnosis and improves 3-year risk prediction to AUROC 0.66 versus 0.63 from raw notes, with combined structured variables reaching AUROC 0.69 (0.74 for patients under 75).

AI summary of the abstract · may contain errors · not medical advice · read the paper.