AI in the Clinical Documentation Loop: Effects of LLM-Induced Note Bloat and Diagnostic Overcall on Downstream LLMs

Jinghui Liu; Anthony Nguyen; Taylan Gurgenci

cs.CLNov 7, 2026

AI in the Clinical Documentation Loop: Effects of LLM-Induced Note Bloat and Diagnostic Overcall on Downstream LLMs

Authors: Jinghui Liu, Anthony Nguyen, Taylan Gurgenci

Organizations: Australian e-Health Research Centre, CSIRO · Mater Research Institute – University of Queensland · School of Medicine and Dentistry, Griffith University

Abstract

Large language models (LLMs) are increasingly used to process medical records for healthcare tasks such as clinical documentation. Because these tools also generate documentation, the relationship is now bidirectional: models both read from and write to the record. Understanding how LLMs behave when reading records that already contain LLM-authored content is therefore critical for their safe deployment. In this study, we examine how LLM-authored content in the medical record affects downstream LLMs that read the same record. We consider two types of LLM-induced artifacts: bloat, in which LLM-generated summaries accumulate in the record and change the length and provenance of its content over time; and distortion, in which LLMs subtly alter clinical findings rather than fabricating facts outright. As a case of distortion, we study diagnostic overcall (documenting an unconfirmed differential diagnosis in the record) and test whether it propagates to the reader LLM's output. Across seven reader models from four families, we find evidence of both effects. First, LLM-generated summary notes modestly but consistently degraded downstream hospital-course generation compared with copy-forward and templated text of matched length. Second, injected overcalls reappeared in downstream outputs at rates that varied by reader model, without detectable changes in standard quality metrics. These findings show the need to monitor AI-generated content in medical records and to account for it when those records are later processed by AI.

Figures & tables

Figure 1: Study design for controlled analyses of bloat and distortion artifacts in AI-generated documentation and their impact on downstream record readers. BHC: Brief Hospital Course, the hospital-course section of the discharge summary.

Figure 2: Effect of note bloat on downstream clinical summarization (main reader, Gemma-4-26B-A4B). X-axis: inserted share of the final record (dose). Y-axis: change from the same reader’s output on the original record, in percentage points (pp) except rated accuracy (1–5 scale). Shades are 95% bootstrap CIs.

LLM evaluation: induced propagation (%)
	Related	Unrelated
Main reader	24.6 [16.8, 32.4]	16.1 [10.9, 22.2]
GPT-OSS evaluator	23.8 [16.8, 31.4]	15.6 [10.2, 21.1]
Nemotron generator	10.2 [5.1, 15.8]	8.3 [4.4, 12.7]
Five other readers	5.8–31.2	2.5–23.9
Clinician review (24 main-reader pairs)

Table 1: Diagnostic overcall propagation. Unless noted, Qwen3.8-27B wrote the overcalls and evaluated the summaries of the main reader (Gemma-4-26B-A4B). Brackets: 95% intervals, Bonferroni-adjusted over the three main-reader rows. Results for other readers in Appendix Table 3 .

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Role	Model	Maximum output
Artifact generators	Qwen3.8-27B (main); Nemotron-3.5-Lightning-30B-A3B (NVFP4)	768/1024 tokens
Main summarization reader	Gemma-4-26B-A4B (FP8)	1536 tokens
Further summarization readers	Gemma-4-12B, Gemma-4-31B (FP8); Nemotron-3.5-Lightning-30B-A3B (NVFP4);	1536 tokens
	Llama-3.3-70B-Instruct, Llama-4-Scout-17B-16E-Instruct (FP8); Qwen3.8-27B
General evaluator	Qwen3.8-27B	step-specific
Targeted evaluators	Qwen3.8-27B; GPT-OSS-120B	300 tokens

Appendix

Table 2: Model roles in this study. Generator budgets are for bloat notes and overcall edits, respectively. Nemotron-3.5 generated artifacts for the main reader only.

Figure 3: Dose–response effect of note bloat with Gemma-4-31B as the reader model. Axes as in Figure 2 .

	Bloat, dose 0.25 (pp)		Overcall propagation (%)			Rated
Reader	$\Delta$ ROUGE-L	$\Delta$ supported share	Related	Unrelated	Hardened	accuracy
Artifacts written by Qwen3.8-27B
Gemma-4-12B	$-0.76$	$-3.03$	$10.2$	$6.0$	0.0 / 0.4	$2.35$
	[ $-0.96$ , $-0.58$ ]	[ $-4.02$ , $-2.04$ ]	[ $6.2$ , $15.3$ ]	[ $3.2$ , $9.2$ ]
Gemma-4-26B-A4B (main)	$-0.66$	$-3.71$	$24.6$	$16.1$	0.0 / 0.0	$2.66$
	[ $-0.86$ , $-0.45$ ]	[ $-4.67$ , $-2.77$ ]	[ $18.4$ , $31.3$ ]	[ $11.7$ , $21.0$ ]

Appendix

Table 3: Both effects across readers and generators, with 95% percentile-bootstrap intervals below each value. Bloat: AI-red minus problem-list H-red at dose 0.25 (

n=300

; 299 for Qwen3.8-27B). Overcall propagation: induced propagation for related and unrelated targets, and the share of pairs in which the propagated condition was stated as confirmed (hardened; related / unrelated), scored by the Qwen evaluator (

n=171

–179 related and 243–252 unrelated pairs). Rated accuracy: judge-rated accuracy (1–5) of the reader’s summaries of the original records. Qwen3.8-27B as a reader was scored for similarity only, because it is also the evaluator.

	Qwen-written notes			Nemotron-written notes
Measure ( $\texttt{AI-red}-\texttt{H-red}$ )	0.10	0.25	0.50	0.10	0.25	0.50
ROUGE-L (pp)	$-0.35$ ∗	$-0.66$ ∗	$-0.93$ ∗	$-0.45$ ∗	$-0.57$ ∗	$-1.22$ ∗
Summary length (pp of reference)	$+6.1$ ∗	$+13.0$ ∗	$+17.2$ ∗	$+4.8$ ∗	$+5.5$ ∗	$+15.6$ ∗
Claims per summary	$+2.0$ ∗	$+4.1$ ∗	$+5.4$ ∗	$+1.4$ ∗	$+2.7$ ∗	$+5.1$ ∗
Supported-claim share (pp)	$-1.2$ ∗	$-3.7$ ∗	$-3.5$ ∗	$-0.9$ ∗	$-2.7$ ∗	$-4.5$ ∗
Contradicted-claim share (pp)	$-0.1$	$-0.3$	$+0.6$ ∗	$0.0$	$0.0$	$-0.2$

Appendix

Table 4: AI-red minus problem-list H-red for the main reader, by dose and generator. ∗ : 95% interval excludes zero.

n=300

at doses 0.10 and 0.25; at 0.50,

n=274

(Qwen) and 258–259 (Nemotron).

	R-red $-$ H-red (wording only)			AI-red $-$ R-red (content)
Measure	0.10	0.25	0.50	0.10	0.25	0.50
ROUGE-L (pp)	$-0.07$	$-0.07$	$0.00$	$-0.28$ ∗	$-0.59$ ∗	$-0.92$ ∗
BERTScore F1 (pp)	$+0.04$	$-0.02$	$-0.11$	$-0.36$ ∗	$-0.49$ ∗	$-0.27$ ∗
Summary length (pp of reference)	$-0.2$	$+2.4$	$+2.3$	$+6.3$ ∗	$+10.6$ ∗	$+15.0$ ∗
Claims per summary	$-0.2$	$+0.8$	$-0.2$	$+2.2$ ∗	$+3.3$ ∗	$+5.7$ ∗
Supported-claim share (pp)	$-0.5$	$-0.8$ ∗	$-0.8$	$-0.6$	$-2.9$ ∗	$-2.7$ ∗

Appendix

Table 5: Rephrased control ( R-red ) for the main reader, by dose. ∗ : 95% interval excludes zero.

n=300

at doses 0.10 and 0.25 and 272 at 0.50.

Reader	Arm	ROUGE-L (pp)	BERTScore F1 (pp)	Accuracy	Coverage	Overall
Gemma-4-26B-A4B	Related	$+0.02$ $[-0.19,+0.21]$	$-0.05$ $[-0.21,+0.10]$	$-0.12$ $[-0.25,+0.02]$	$-0.01$ $[-0.09,+0.06]$	$-0.04$ $[-0.14,+0.05]$
	Unrelated	$-0.03$ $[-0.22,+0.15]$	$-0.00$ $[-0.15,+0.14]$	$-0.07$ $[-0.20,+0.07]$	$-0.02$ $[-0.09,+0.05]$	$-0.03$ $[-0.12,+0.05]$
Gemma-4-31B	Related	$+0.10$ $[-0.10,+0.30]$	$+0.01$ $[-0.14,+0.16]$	$-0.02$ $[-0.16,+0.14]$	$-0.06$ $[-0.14,+0.02]$	$-0.05$ $[-0.15,+0.04]$
	Unrelated	$-0.03$ $[-0.21,+0.15]$	$-0.01$ $[-0.15,+0.12]$	$-0.05$ $[-0.17,+0.08]$	$-0.05$ $[-0.11,+0.02]$	$-0.05$ $[-0.13,+0.04]$

Appendix

Table 6: Changes in document-level quality for the distorted summary relative to its matched neutral-rephrase control. Similarity metrics are reported in percentage points; judge metrics are reported on their original 1–5 scales. Values are mean paired differences with 95% percentile-bootstrap intervals. Related-arm estimates used 191 pairs and unrelated-arm estimates used 258 pairs. The Qwen evaluator supplied the judge scores.

Medical LLMs

Medical large language models are AI systems adapted to clinical and biomedical data for health reasoning, information extraction, and communication.

See all 418 papers

Oct 6, 2026

Large language models are vulnerable to incidental information in clinical documentation and reasoning

Large language models (LLMs) are increasingly relied upon to support ambient documentation and clinical reasoning. Here we examine the impact of a failure mode shared between these two applications by assessing their sensitivity to information incidental to the patient encounter. In 576 patient-clinician dialogues, we found that frontier models inserted small-talk exchanges into 35% of notes, while mean quality scores changed by at most 0.20 points on five-point scales. In 3.7% of frontier notes, models misattributed the asides or used them clinically. In 57 mock recorded consultations, background speech from a separate patient encounter at -10 dB leaked into 48.2% of transcripts, with contamination detected in 5.3% of downstream notes generated by four open-weight models. We propose a dual encoding hypothesis of clinical reasoning and distraction in LLMs, with preliminary evidence that LLM components associated with disruption by incidental information also support clinical reasoning. These findings support evaluating resistance to incidental information before clinical use, with safeguards that prevent contamination while preserving clinical reasoning.

Krithik Vishwanath, Brandon Ye, Anton Alyakin +4

Department of Neurosurgery, NYU Langone Health, New York, NY, USA · Johns Hopkins University School of Medicine, Baltimore, MD, USA · Washington University School of Medicine in St. Louis, St. Louis, MO, USA +7

Medical Report Generation Clinical Language Model Evaluation Language Model Robustness Medical LLMs

Ask PDF

May 18, 2026

Systematic Evaluation of the Quality of Synthetic Clinical Notes Rephrased by LLMs at Million-Note Scale

Large language models (LLMs) can generate or synthesize clinical text for a wide range of applications, from improving clinical documentation to augmenting clinical text analytics. Yet evaluations typically focus on a narrow aspect -- such as similarity or utility comparisons -- even though these aspects are complementary and best viewed in parallel. In this study, we aim to conduct a systematic evaluation of LLM-generated clinical text, which includes intrinsic, extrinsic, and factuality evaluations of synthetic clinical notes rephrased from MIMIC databases at million-note scale. Our analysis demonstrates that synthetic notes preserve core clinical information and predictive utility for coarse-grained tasks despite substantial linguistic changes, but lose fine-grained details for task like ICD coding. We show this loss of detail can be substantially mitigated by rephrasing notes by chunks rather than by the whole note, but at the cost of reduced factual precision under incomplete context. Through fact-checking and error analysis, we further find that synthesis errors are dominated by misinterpretation of clinical context, alongside temporal confusion, measurement errors, and fabricated claims. Finally, we show that the synthetic notes -- despite their task-agnostic nature -- can effectively augment task-specific training for rare ICD codes.

Jinghui Liu, Sarvesh Soni, Anthony Nguyen

Australian e-Health Research Centre, CSIRO, Australia · National Library of Medicine, National Institutes of Health, USA

Synthetic Data Augmentation Automated Medical Coding Synthetic Data Generation Clinical Language Model Evaluation Clinical NLP Synthetic Data Evaluation

Ask PDF

Mar 27, 2026

Phase 1 Implementation of LLM-generated Discharge Summaries showing high Adoption in a Dutch Academic Hospital

Writing discharge summaries to transfer medical information is an important but time-consuming process that can be assisted by Large Language Models (LLMs). This prospective mixed methods pilot study evaluated an Electronic Health Record (EHR)-integrated LLM to generate discharge summaries drafts. In total, 379 discharge summaries were generated in clinical practice by 21 residents and 4 physician assistants during 9 weeks in our academic hospital. LLM-generated text was copied in 58.5% of admissions, and identifiable LLM content could be traced to 29.1% of final discharge letters. Notably, 86.9% of users self-reported a reduction in documentation time, and 60.9% a reduction in administrative workload. Intent to use after the pilot phase was high (91.3%), supporting further implementation of this use-case. Accurately measuring the documentation time of users on discharge summaries remains challenging, but will be necessary for future extrinsic evaluation of LLM-assisted documentation.

Nettuno Nadalini, Tarannom Mehri, Anne H Hoekman +6

Department of Health Information Office, Data & Digitalization, University Medical Center, Groningen, the Netherlands · Department of Trauma Surgery / Orthopedics, University Medical Center, Groningen, the Netherlands · Department of Orthopaedic Surgery, Massachusetts General Hospital, Boston, MA, USA +5

Medical Report Generation Medical LLMs Electronic Health Records

Ask PDF

Abstract

Medical LLMs

Explore similar work

Large language models are vulnerable to incidental information in clinical documentation and reasoning

Systematic Evaluation of the Quality of Synthetic Clinical Notes Rephrased by LLMs at Million-Note Scale

Phase 1 Implementation of LLM-generated Discharge Summaries showing high Adoption in a Dutch Academic Hospital