From Retinal Evidence to Safe Decisions: RETINA-SAFE and ECRT for Hallucination Risk Triage in Medical LLMs

Zhe Yu; Wenpeng Xing; Meng Han

cs.AIApr 7, 2026

From Retinal Evidence to Safe Decisions: RETINA-SAFE and ECRT for Hallucination Risk Triage in Medical LLMs

Authors: Zhe Yu, Wenpeng Xing, Meng Han

Organizations: Binjiang Institute of Zhejiang University, Hangzhou, China · Communication University of Zhejiang, Hangzhou, China · Zhejiang University, Hangzhou, China

Abstract

Hallucinations in medical large language models (LLMs) remain a safety-critical issue, particularly when available evidence is insufficient or conflicting. We study this problem in diabetic retinopathy (DR) decision settings and introduce RETINA-SAFE, an evidence-grounded benchmark aligned with retinal grading records, comprising 12,522 samples. RETINA-SAFE is organized into three evidence-relation tasks: E-Align (evidence-consistent), E-Conflict (evidence-conflicting), and E-Gap (evidence-insufficient). We further propose ECRT (Evidence-Conditioned Risk Triage), a two-stage white-box detection framework: Stage 1 performs Safe/Unsafe risk triage, and Stage 2 refines unsafe cases into contradiction-driven versus evidence-gap risks. ECRT leverages internal representation and logit shifts under CTX/NOCTX conditions, with class-balanced training for robust learning. Under evidence-grouped (not patient-disjoint) splits across multiple backbones, ECRT provides strong Stage-1 risk triage and explicit subtype attribution, improves Stage-1 balanced accuracy by +0.15 to +0.19 over external uncertainty and self-consistency baselines and by +0.02 to +0.07 over the strongest adapted supervised baseline, and consistently exceeds a single-stage white-box ablation on Stage-1 balanced accuracy. These findings support white-box internal signals grounded in retinal evidence as a practical route to interpretable medical LLM risk triage.

Figures & tables

Panel of Figure 1 — Figure 1: RETINA-SAFE Benchmark Construction and Taxonomy. Driven by raw clinical retinal records, structural attributes are extracted to formulate cases $(q,o,c,e)$ . The evaluation follows a safety-critical taxonomy: answers are either directly supported (E-Align), structurally conflicting with the prompt (E-Conflict), or lacking sufficient evidence to decide (E-Gap).

Class / Split	Samples	Ratio	Avg Q Len	Avg Ev Len
E-Align	1,149	0.092	34.15	3.00
E-Conflict	5,107	0.408	47.34	3.00
E-Gap	6,266	0.500	38.67	3.00
All	12,522	1.000	41.79	3.00

Table 1: RETINA-SAFE benchmark statistics (computed from retina_safe.jsonl ; lengths use whitespace token counts for question and evidence text).

Baseline	E-Align	E-Conflict	E-Gap	Macro-Acc
Random-choice (4-way; lower bound)	0.2500	0.2500	0.2500	0.2500
Llama 3-8B (black-box QA)	0.4412	0.2215	0.1241	0.2623
Llama 2-7B (black-box QA)	0.3842	0.1412	0.0815	0.2023
Llama 2-13B (black-box QA)	0.5124	0.1642	0.0912	0.2559
Qwen2.5-7B (black-box QA)	0.5891	0.8812	0.2114	0.5606

Table 2: RETINA-SAFE benchmark difficulty baselines (black-box MCQA with evidence). Columns report per-task accuracy and macro task accuracy.

Panel of Figure 2 — Figure 2: RETINA-SAFE benchmark exemplars (images + task semantics). Left: representative retinal images linked to evidence records across E-Align / E-Conflict / E-Gap. Right: the RETINA-SAFE textual record schema and task semantics, illustrating that labels are assigned from the evidence-conditioned decision setting (question + structured evidence), not from image appearance alone; labels are defined by the question–evidence relation rather than image severity alone.

Figure 3: ECRT Framework Overview. ECRT leverages internal representation and logit shifts under CTX/NOCTX conditions to extract three families of safety signals (Discrepancy, Deviation, Incoherence). These features train a two-stage triage engine that detects risks and attributes them to clinical contradiction or evidence-gap categories.

Family	Notation	Level	Technical Metric
Discrepancy	$\Delta\mathrm{logits}$	Prediction	Logit output shift $\Delta z_{t}$
Deviation	$\Delta\mathrm{traj}$	Latent	Hidden-state shift $\Delta h_{t}^{\ell}$
Incoherence	$\Delta\mathrm{inc}$	Structural	Per-layer KL-Divergence

Table 3: ECRT Internal Signal Families. Features are extracted via teacher-forced paired passes and aggregated into a pooled vector

\mathcal{P}(S)

Backbone	Method	U-Recall	Flag Rate	S1 BA
L3-8B	ECRT	$0.9505\pm 0.0011$	$0.8712\pm 0.0049$	$0.8405\pm 0.0193$
L3-8B	1-stage	$0.9556\pm 0.0026$	$0.8767\pm 0.0054$	$0.8124\pm 0.0201$
L2-7B	ECRT	$0.9430\pm 0.0076$	$0.8637\pm 0.0151$	$0.9229\pm 0.0132$
L2-7B	1-stage	$0.9482\pm 0.0099$	$0.8878\pm 0.0216$	$0.8926\pm 0.0117$
Mistral	ECRT	$0.9466\pm 0.0024$	$0.8636\pm 0.0048$	$0.9310\pm 0.0024$
Mistral	1-stage	$0.9458\pm 0.0024$	$0.8802\pm 0.0051$	$0.9064\pm 0.0021$

Table 4: Primary endpoint (Stage 1 Safe/Unsafe) under the strict group protocol using the target-recall policy. Values are mean

\pm

std across seeds (Llama 3-8B: 5; others: 3).

Panel of Figure 4 — Figure 4: Comparison of primary clinical endpoint (Stage-1 BA) under a target-recall policy. ECRT consistently outperforms the single-stage ablation across all evaluated backbones.

Method	Llama 3-8B				Llama 2-7B				Mistral-7B
Method	Rec.	F.R.	S1 BA	$\Delta$ BA	Rec.	F.R.	S1 BA	$\Delta$ BA	Rec.	F.R.	S1 BA	$\Delta$ BA
ECRT (Ours)	0.9505	0.8712	0.8405	–	0.9430	0.8637	0.9229	–	0.9466	0.8636	0.9310	–
FactoScope	0.9542	0.8856	0.8125	-0.0280	0.9556	0.8754	0.8546	-0.0683	0.9551	0.8723	0.9124	-0.0186
UQ Heads	0.9525	0.8924	0.7856	-0.0549	0.9512	0.8824	0.8245	-0.0984	0.9521	0.8812	0.8623	-0.0687
Lookback Lens	0.9518	0.9125	0.7152	-0.1253	0.9508	0.8912	0.7854	-0.1375	0.9518	0.8912	0.8124	-0.1186
ReDeEP	0.9531	0.9082	0.6923	-0.1482	0.9541	0.8954	0.7325	-0.1904	0.9542	0.9054	0.7423	-0.1887

Table 5: Comprehensive Stage-1 performance comparison under the strict target-recall policy (seed13). We report Unsafe-Recall (Rec.), Flag Rate (F.R.), Stage-1 Balanced Accuracy (BA), and the performance gap (

\Delta

BA) relative to ECRT. To maintain readability with full precision, results are split across two stacked blocks. ECRT consistently achieves the highest Stage-1 BA across all evaluated backbones.

Backbone	Gap Recall ( $\uparrow$ )	Contradict. Recall ( $\uparrow$ )	S2 BA ( $\uparrow$ )
L3-8B	$0.91\sim 0.94$	$0.85\sim 0.89$	$0.87\sim 0.91$
L2-7B	$0.93\sim 0.95$	$0.91\sim 0.93$	$0.92\sim 0.94$
Mistral-7B	$0.95\sim 0.97$	$0.91\sim 0.94$	$0.93\sim 0.95$
Qwen2.5-7B	$0.97\sim 0.99$	$0.94\sim 0.97$	$0.96\sim 0.98$
Qwen3-8B	$0.96\sim 0.98$	$0.93\sim 0.97$	$0.95\sim 0.99$

Table 6: Secondary endpoint (Stage-2 Contradiction-vs-Gap) performance on ground-truth unsafe cases. Values represent the robust separation capability of the Stage-2 head (ranges across multi-seed runs).

Hallucination Detection

Hallucination detection identifies unsupported, fabricated, or evidence-inconsistent content in language and multimodal model outputs.

See all 78 papers

Med-HEAL: Analyzing and Mitigating Hallucinations in Medical LLMs with Hallucination-Aware In-Context Learning
May 31, 2026Yiming Liao, Zeno Franco, Jose Eduardo Lizarraga Mazaba +1LLM Self-Correction LLM Hallucination Mitigation
University of Maryland, Baltimore County Baltimore, Maryland, USA · Concordia University Wisconsin · Medical College of Wisconsin Milwaukee, Wisconsin USA
DIASENTINEL: An Auditable Multi-Agent System for Guideline-Grounded Diabetes Risk Screening
Aug 31, 2026Yung Wei Shueh, Zhi-Jie Chen, Chia-Hsuan Hsu +9Multi-Agent LLM Systems Medical Report Generation
Far Eastern Memorial Hospital, Taiwan · University of Michigan, USA · Rutgers University, USA +1
ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning
Jun 12, 2026Sicheng Yang, Hangjie Yuan, Wenjun Zhang +5Multimodal Hallucination Clinical Reasoning
The Hong Kong University of Science and Technology (Guangzhou) · DAMO Academy, Alibaba Group · Hupan Lab +1

May 31, 2026J/K move · Enter open · S save

Med-HEAL: Analyzing and Mitigating Hallucinations in Medical LLMs with Hallucination-Aware In-Context Learning

Yiming Liao, Zeno Franco, Jose Eduardo Lizarraga Mazaba, Keke Chen

University of Maryland, Baltimore County Baltimore, Maryland, USA · Concordia University Wisconsin · Medical College of Wisconsin Milwaukee, Wisconsin USA

LLM Self-Correction LLM Hallucination Mitigation Clinical QA In-Context Learning LLM Hallucination Detection Retrieval-Augmented ICL+1

Open paper AskPDF arXiv

Hallucinations in medical large language models (LLMs) pose serious risks for clinical decision support, particularly when models must reason over complex electronic health records (EHRs). However, existing benchmarks often lack a realistic clinical context and provide limited insight into how hallucinations can be mitigated in practice. We introduce Med-HEAL, a framework for systematically identifying, analyzing, and mitigating hallucinations in medical LLMs using clinically grounded data. Building on the EHRNoteQA benchmark derived from MIMIC-IV discharge summaries, we construct a hallucination dataset by evaluating BioMistral-7B on open-ended clinical question answering tasks. Model outputs are labeled through a dual evaluation pipeline that combines LLM-as-a-Judge assessment (GPT-4o) with human auditing by medical student reviewers, producing correctness judgments and annotations of reasoning errors via a custom web-based evaluation system. We then leverage this dataset to investigate mitigation strategies: a self-critique pipeline, in which the test model reviews its own answers to detect potential errors and regenerates responses for flagged cases, and retrieval-augmented in-context learning (RA-ICL), which exposes the model to hallucinated and corrected examples. Experiments across five open-source LLMs-BioMistral, Llama-3.1, DeepSeek, Qwen2.5, and Qwen3, show that the self-critique strategy improves accuracy for three of five models (p < 0.05) without requiring parameter updates. Med-HEAL provides both a reusable hallucination dataset and a practical framework for studying and mitigating hallucinations in medical LLMs, supporting safer deployment of AI systems in clinical environments. Our code and data are publicly available at https://github.com/yimingliao-blad/med-heal.git.

Abstract

Hallucination Detection

Explore similar work

Med-HEAL: Analyzing and Mitigating Hallucinations in Medical LLMs with Hallucination-Aware In-Context Learning

DIASENTINEL: An Auditable Multi-Agent System for Guideline-Grounded Diabetes Risk Screening

ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning