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

CLIN-SUMM is a framework that incrementally constructs structured, date-partitioned longitudinal summaries of clinical notes to reduce data redundancy while providing a scalable representation layer that improves disease prediction and longitudinal reasoning.

The Gist

When a person receives medical care over many years, doctors write thousands of notes about their visits. These notes contain vital details: new symptoms, changes in medication, or results from lab tests. However, for a doctor trying to understand a patient's long-term health, these notes can be overwhelming. They are often repetitive, with much of the same information copied from one visit to the next, and the most important updates can be buried under mountains of text.

The researchers in this paper developed a system called CLIN-SUMM to solve this problem. Instead of trying to summarize a patient's entire history into one static paragraph, the system builds a growing, organized timeline. Every time a new medical note is written, the system looks at it and identifies only the new information. It then adds this new data into specific categories, such as "Diagnosis" or "Medications," and stamps it with the date. This creates a structured record that evolves alongside the patient, much like a ledger that only records new transactions rather than rewriting the entire history of an account every day.

To build this, the researchers used a large language model. They designed a process where the system first creates an initial summary of a patient's first visit. For every visit after that, the system compares the new note to the existing summary and only extracts what has changed. To keep the process efficient, the system also filters out notes that are nearly identical to previous ones.

The authors tested this system using data from over 12,000 patients at Massachusetts General Hospital. They found that CLIN-SUMM reduced the amount of text by about 70%, making the information much easier for a computer to process while still keeping the details accurate. When doctors reviewed the summaries, they rated them highly for correctness and completeness, noting that the system rarely invented information that wasn't in the original notes.

The researchers then used these compressed summaries to see if they could help predict health issues. They focused on dementia as a case study. They trained a machine learning model using only the CLIN-SUMM summaries—not the original, massive pile of notes. The model was able to identify dementia cases with high accuracy. More importantly, the model could predict the risk of a dementia diagnosis up to three years before it actually happened. By looking at the summaries, the model picked up on subtle patterns in the text, such as mentions of memory issues, changes in walking, or dizziness, which are often documented in notes long before a formal diagnosis is made.

The paper also shows that these summaries are better at capturing medication details than the standard structured databases used in hospitals. The researchers found that the summaries identified many more instances of specific medications, like Donepezil, because the information was often written in the narrative notes but missed by the automated coding systems.

Ultimately, the researchers suggest that CLIN-SUMM acts as a middle layer. It transforms messy, repetitive, and massive amounts of text into a clean, organized, and time-stamped format that can be used for both quick human review and advanced computer modeling.

Technical Summary

Technical Summary: CLIN-SUMM (Clinical Longitudinal Insight from Notes using Summarization)

1. Problem Statement

Electronic Health Records (EHRs) contain vast amounts of longitudinal clinical notes that capture essential nuances (symptoms, reasoning, lifestyle) often missing from structured data (ICD codes). However, these notes present three major challenges for both clinicians and machine learning (ML) models:

• Volume and Redundancy: Patients accumulate thousands of notes, often containing over 50% duplicated text due to "copy-forward" practices.
• Fragmentation: Critical clinical updates are scattered across different encounters and specialties, making it difficult to track a patient's evolving trajectory.
• Computational Inefficiency: Existing Large Language Model (LLM) approaches typically summarize individual notes or episodes in isolation, failing to preserve temporal fidelity or creating static summaries that obscure how a patient's condition changes over time.

2. Methodology

The authors propose CLIN-SUMM, a framework that treats summarization as a longitudinal representation problem rather than a simple text-compression task.

• Incremental Sequential Summarization: Instead of "squashing" a history into one paragraph, CLIN-SUMM builds a cumulative, date-partitioned representation.
  • Initial Step: An LLM (specifically Qwen3 15) uses a baseline prompt to organize the first clinical note into seven structured sections (e.g., Chief Complaints, Medications, Diagnosis, etc.).
  • Incremental Update: For every subsequent encounter, a "change-focused" prompt inputs the existing summary and the new note. The LLM adds only novel information to the relevant sections and populates a "Changes over Time" section to capture temporal shifts (e.g., medication titration or new symptoms).
• Technical Safeguards:
  • Redundancy Filtering: Uses Jaccard similarity to skip near-identical notes.
  • Context Management: Employs a sliding window mechanism to handle long histories without exceeding LLM context limits.
  • Temporal Fidelity: Notes are processed strictly chronologically, ensuring the model never accesses "future" data, which preserves causal direction for predictive modeling.
• Evaluation Framework: The authors used a cohort of 12,356 patients from Massachusetts General Hospital, using dementia diagnosis and 3-year risk prediction as primary use cases. They fine-tuned Clinical ModernBERT on these summaries to test downstream utility.

3. Key Contributions

• A New Representation Layer: CLIN-SUMM acts as a "durable" intermediate layer between raw, noisy text and downstream tasks, decoupling narrative processing from prediction.
• Temporal-Aware Summarization: Unlike standard abstractive summarization, this method explicitly encodes how a patient's clinical state evolves.
• Open-Source Viability: Demonstrated that open-source models (Qwen3) can achieve near-parity with frontier proprietary models (GPT-4o) in clinical utility while being significantly cheaper and more privacy-compliant (deployable on-premise).

4. Key Results

• Compression Efficiency: Achieved an overall word count reduction of 69.86% (and ~52.6% at the patient level) while maintaining high clinical fidelity.
• Clinical Accuracy: Clinician reviews yielded high scores for correctness (4.69/5) and completeness (4.65/5), with a low hallucination rate (<4%).
• Predictive Performance (Dementia):
  • Diagnosis Model: Achieved an AUROC of 0.86.
  • 3-Year Risk Prediction: Achieved an AUROC of 0.81.
  • Both significantly outperformed traditional multifactorial models (like CAIDE) that rely solely on structured demographics and vitals.
• Early Detection: The diagnosis model maintained strong discriminatory power even when applied to data truncated years before the actual diagnosis (e.g., AUROC of 0.82 at 1 year prior).
• Medication Extraction: CLIN-SUMM identified substantially more medication events (e.g., Donepezil) than structured EHR tables, suggesting it can mitigate "missingness" in standard medical databases.

5. Significance

CLIN-SUMM represents a shift from "summarizing text" to "encoding clinical trajectories." Its significance lies in its scalability and reusability: a single set of CLIN-SUMM summaries can be computed once and used for diverse tasks—from manual clinician chart review to automated risk stratification and medication tracking. By transforming unstructured, redundant narratives into a structured, compressed, and temporally accurate format, CLIN-SUMM provides a scalable infrastructure for longitudinal clinical reasoning and advanced disease modeling in real-world healthcare systems.