Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1356
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-14T18:14:25.250055Z
status: published
lens: trace
sector: health
headline: Automating Injury Severity Score Calculation From Free-Text Discharge Diagnoses Using Domain-Specific Retrieval-Augmented Large Language Models: Retrospective Validation Study
dek: Background: Trauma registries are crucial for evaluating the prognosis of injuries and the quality of trauma care. However, the calculation of Injury Severity Score (ISS) remains a predominantly manual process, which is labor-intensive, potentially error-prone, and difficult to scale. Automating this process is a critical step toward enhancing the efficiency and quality of trauma registry data.
Objective: This study conducts a feasibility test on the use of large language models (LLMs) to automatically calculate…
gain_title: A domain-specific retriever fine-tuned on trauma ICD-10 codes raised LLM exact-match accuracy for Injury Severity Score calculation, reaching 75.3% with GPT-4.1 versus 70.0% for the ICDPIC-R baseline.
problem_title: Standard retrieval augmentation gave no consistent gains, and baseline LLMs reached only 47.3% to 64.7% exact-match accuracy, so off-the-shelf models are unreliable for this task.
trace_subject: LLM-based Injury Severity Score calculation from free-text trauma discharge diagnoses
gain_reading: A domain-specific retriever fine-tuned on trauma ICD-10 codes raised LLM exact-match accuracy for Injury Severity Score calculation, reaching 75.3% with GPT-4.1 versus 70.0% for the ICDPIC-R baseline.
gain_evidence: GPT-4.1 with domain-specific RAG reached the highest EM accuracy (75.3%, 95% CI 66.1%-82.7%) | the domain-specific RAG substantially improved accuracy (62.7% to 75.3%)
problem_reading: Standard retrieval augmentation gave no consistent gains, and baseline LLMs reached only 47.3% to 64.7% exact-match accuracy, so off-the-shelf models are unreliable for this task.
problem_evidence: Baseline state-of-the-art LLMs achieve modest exact-match (EM) accuracy, ranging from 47.3% to 64.7% | Standard RAG did not produce consistent gains (54.3% to 60.3%)
quick_read: Researchers tested whether large language models could calculate the Injury Severity Score from free-text discharge diagnoses of trauma patients at one Taiwanese hospital in 2022. They compared plain LLMs, LLMs with a standard retrieval system, and LLMs with a retriever fine-tuned on trauma diagnosis codes, scoring all of them against physician-verified answers.

The work matters because manual scoring for trauma registries is slow, error-prone and hard to scale, so automation could ease administrative burden and improve registry data. Results come from a single center and a retrospective test set, and the authors say multi-institution validation is still needed before the approach can be trusted more widely.
limitation: The study is a single-center retrospective feasibility test, so the findings need confirmation across multiple institutions.
tag: Dual reading
key_points: Baseline LLMs reached exact-match accuracy of 47.3% to 64.7% on Injury Severity Score calculation. | Standard retrieval augmentation produced inconsistent results, ranging from 54.3% to 60.3%. | The domain-specific retriever, fine-tuned on trauma ICD-10 codes, lifted accuracy to between 62.7% and 75.3%. | Gains were most pronounced in patients with moderate to severe injuries. | The best system, GPT-4.1 with domain-specific retrieval, was compared with a 70.0% deterministic ICDPIC-R baseline.
rundown: The data were deidentified free-text discharge diagnoses from trauma patients at Chang Gung Memorial Hospital in Linkou, Taiwan, from 2022. Performance was judged against a physician-verified gold standard, using exact score prediction, tolerance-based accuracy and severity categories.

The retrieval component drew on knowledge from the Abbreviated Injury Scale, and the domain-specific version had its retriever fine-tuned on trauma-related ICD-10 codes. This was contrasted with a standard retrieval setup that did not help consistently.

The authors conclude that accuracy was comparable overall to the deterministic ICDPIC-R tool, with the advantage concentrated in the most severe injuries. They suggest this could reduce coding workload and improve standardization of trauma registry data.
sources:
- peer_reviewed | JMIR Medical Informatics | https://doi.org/10.2196/93922 | 2026-10-13
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

d48d5b3ba2edf2f64c16f0d8c37d16e54c93cdf51d6c159337e4266c6441f07c

previous

0000000000000000000000000000000000000000000000000000000000000000