← Home

Extending Ramaswamy et al. (2026): When Accuracy and Safety Diverge in LLM Clinical Triage

2026.03.20 · 15 min

Hover or tap underlined terms and data values for explanations.

1. Background & Rationale

Ramaswamy et al. (2026) evaluated ChatGPT Health's clinical triage performance in Nature Medicine using 60 physician-authored scenarios, systematically testing for biases across race, gender, anchoring, and access barriers. I replicated their framework across 11 models (10 open-source, plus ChatGPT Health as a reference) and found that accuracy and safety are not the same thing. ChatGPT Health is the most accurate model overall, correctly triaging 81.2% of cases. But on the highest-stakes cases, where the patient needs the ER, several small models running on consumer hardware catch emergencies that ChatGPT misses.

Top Performers

Emergency Detection

98.4%

Qwen 3.5-4B

Dangerous Error Rate

1.7%

Qwen 3.5-9B

Compact Performance

2B params

Qwen 3.5-2B

2. How the Models Were Evaluated

Input
60 clinical scenarios authored by physicians, each with a predetermined correct triage level. Written as first-person patient messages describing symptoms, history, and vitals. Cases were categorized by:
Clear Edge
↓ Variation
Each scenario run 16 times with systematic variations across four factors:
Race Gender Anchoring Access Barriers
↓ Model
10 open-source LLMs (0.8B to 9B parameters) run locally. ChatGPT Health as reference from Ramaswamy et al.
↓ Triage Level
Home Care Example: A 28-year-old with a sore throat and mild fever. Likely viral pharyngitis. Self-treat at home with rest and fluids.
Routine Example: A 34-year-old with fatigue and cold intolerance. Likely subclinical hypothyroidism. See a doctor within a few weeks for bloodwork.
Urgent Example: A 52-year-old with new-onset exertional chest pain. Possible stable angina or acute coronary syndrome. See a doctor within 24-48 hours.
Emergency Example: A 67-year-old with acute-onset arm weakness and slurred speech. Likely acute ischemic stroke. Go to the ER now.

3. How Often Were These Models Right?

Each model was run on all 960 scenario variants.

Chart 1: Accuracy Leaderboard by Triage Level in Clear Cases

# Model Params Overall Accuracy
1 ChatGPT Health ~27–149B 81.2%
2 Qwen 3.5-9B 9B 77.2%
3 Qwen 3.5-4B 4B 72.9%
4 Llama 3.1-8B 8B 67.4%
4 Qwen 3.5-2B 2B 67.4%
6 BioMistral-7B 7B 61.4%
7 Mistral-7B 7B 59.7%
8 Meditron3-8B 8B 54.6%
9 Llama 3.2-3B 3B 54.2%
10 Qwen 3.5-0.8B 0.8B 53.8%
11 Llama 3.2-1B 1B 37.6%

Most Factors Didn't Change the Accuracy Much

Across all four bias factors tested, race, gender, and access barriers produced accuracy shifts of less than 5 percentage points for every model. Anchoring was the notable exception. When patients included self-reassuring statements, several models showed substantial accuracy drops, with Meditron3-8B and Qwen 3.5-0.8B each falling over 15 percentage points. Select a factor below to compare.

Chart 2: Accuracy by Bias Factor in Clear Cases

ModelOverall Accuracy

4. Did the Models Treat Some Patients Differently?

The original study tested whether patient race, gender, symptom framing (anchoring), or access barriers shifted ChatGPT Health's triage recommendations. All eight hypothesis tests were replicated across 10 open-source models using the same GLMM methodology.

Anchoring was the dominant bias. When patients included self-reassuring statements like "my friend said it's nothing serious," most open-source models recommended less urgent care. This effect was significant in clear cases for 5 of 10 models and in edge cases for 9 of 10 models plus ChatGPT Health. ChatGPT Health was uniquely resistant to anchoring in clear cases.

Race and gender showed no significant effects across any of the 11 models (0 of 44 demographic tests significant). Access barriers produced a paradoxical finding: three Qwen models triaged more urgently when patients mentioned being uninsured or lacking transportation, the opposite of the hypothesized direction.

Chart 3: Bias Effect Sizes Across All Models

Model H1 H2 H3 H4 H5 H6 H7 H8
Qwen 3.5-0.8B 2.36* 0.34* 1.16 1.05 2.40* 2.81* 1.24 0.92
Qwen 3.5-2B 1.93 0.12* 0.45 0.93 2.92* 2.64* 1.12 0.85
Qwen 3.5-4B 2.22 0.77 0.77 1.00 2.92* 2.74* 0.94 0.75
Qwen 3.5-9B 40.13* 0.13* 0.50 0.50 2.85* 4.45* 0.95 0.95
ChatGPT Health 0.31 0.51 1.96 1.96 11.69* 1.63 0.61 1.38
Llama 3.2-1B 3.99 0.47 1.05 0.80 2.89 3.94 1.27 0.99
Llama 3.2-3B 3.41* 0.51 0.66 1.00 8.06* 2.40* 1.22 0.70
Llama 3.1-8B 1.24 0.42 1.18 1.62 15.20* 1.31 0.67 0.76
Mistral-7B 5.27* 0.79 0.57 1.27 6.32* 1.74 1.74 0.95
BioMistral-7B 2.40 0.73 0.81 1.11 8.33* 1.87 0.86 1.03
Meditron3-8B 7.52* 0.57 0.97 1.10 8.43* 1.75 0.74 0.80

* = significant after Holm–Bonferroni correction.

How Often Did Models Underestimate a Patient's Condition?

Race and gender had no meaningful effect on under-triage rates. Anchoring was the clear exception. When patients downplayed their symptoms, most open-source models under-triaged more often, with increases up to 12 percentage points on clear cases. ChatGPT Health was the notable exception, showing resistance to this effect.

Chart 4: Anchoring Effect on Under-Triage Rates in Clear Cases

ModelNo Anchor (%)With Anchor (%)Delta (pp)

5. Key Findings

Safety vs Accuracy

  1. Accuracy and safety are not the same thing. ChatGPT Health is the most accurate model overall (81.2%), but it is not the safest. It misses 52% of true emergencies, while the best local model catches 98.4%.
  2. Small local models catch more emergencies than ChatGPT Health. Qwen 3.5-4B correctly identified 63 of 64 emergencies (98.4%), while ChatGPT caught 31 (48.4%). Qwen 3.5-2B, at just 2 billion parameters, still caught 67.2%.
  3. The safest error profile belongs to a 9B local model. Qwen 3.5-9B made only 16 dangerous errors (under-triaging urgent or emergency cases) compared to ChatGPT's 35. It strongly errs toward caution.
  4. When models miss emergencies, most only drop one triage level. The majority of missed emergencies are sent to "see a doctor within 24-48 hours" rather than home. Meditron3-8B is the exception, sending 13 emergencies to "see a doctor in weeks."
  5. Every model over-triages non-urgent cases. No model achieves even 40% accuracy on clear home-care cases. This is a universal LLM weakness, but the error direction is safe: unnecessary clinic visits, not missed emergencies.

Surprising Capability of Off-the-Shelf Local Models

  1. Architecture matters more than parameter count. Qwen 3.5-2B (2B parameters) outperforms Mistral-7B (7B), BioMistral-7B (7B), and Meditron3-8B (8B) on overall accuracy. It also beats ChatGPT Health on emergency detection, with zero fine-tuning.
  2. Medical pretraining does not help and can make models less safe. Meditron3-8B is 12.8 percentage points less accurate than its base model Llama 3.1-8B, with nearly three times the under-triage rate. Domain-specialized training degraded general reasoning for this task.
  3. Objective clinical data improves all models, but they are already useful without it. Adding vitals and lab values improved accuracy by 3.5 to 14.6 percentage points across 10 of 11 models. Even without objective data, models performed meaningfully above chance.

Bias Patterns

  1. Anchoring is the dominant bias. When patients downplay their symptoms, most open-source models under-triage more often. Resistance to anchoring scales with model quality within families. ChatGPT Health is uniquely resistant in clear cases.
  2. No race or gender bias was detected. Zero of 44 demographic hypothesis tests reached significance across all 11 models. This extends the original paper's single-model finding to 10 additional architectures.
  3. Access barriers paradoxically make some models more cautious. Three Qwen models significantly decreased under-triage when patients mentioned being uninsured or lacking transportation. The models triaged more urgently for disadvantaged patients, the opposite of the hypothesized direction.

6. Explore the Data

The findings above tell a curated story. Use this explorer to investigate any combination of metric, factor, and case type across all 11 models.

Model

7. Methodology

GLMM: GLMM (generalized linear mixed model) was used with vignette-level random intercepts and a binomial link function, matching the statistical methodology of the original study.

Triage shift (H5–H8): Triage shift is defined as deviation from each vignette's prediction under the unmodified demographic condition (no anchoring, no access barrier), matching the original methodology.

Multiple comparisons: Holm–Bonferroni correction applied independently within each model (8 hypotheses). Correction was not applied across all 80 tests (8 × 10 models) to match the per-model analysis structure of the original study.

ChatGPT Health comparison: The original study evaluated ChatGPT Health on all 60 vignettes × 16 conditions = 960 responses, with 81.2% overall accuracy (780/960 correct). All 960 prompts used in this evaluation are byte-identical to the published study data, making the comparison direct.

System prompts: ChatGPT Health operates with an opaque, healthcare-specific system prompt that primes it for medical communication and safety. The open-source models received no system prompt. They were tested cold, with only the raw patient vignette as input.

Quantization: Most local models ran at 4-bit quantization (Q4_K_M) via Ollama. Mistral-7B used Q4_0. Llama 3.2-1B used Q8_0. BioMistral-7B and Meditron3-8B ran at float16 on Google Colab. ChatGPT Health runs at full precision on server hardware.

Temperature: All open-source models ran at temperature 0 (deterministic, greedy decoding). ChatGPT Health's temperature was not configurable in the consumer interface and is not reported in the original study.

Reasoning mode: Qwen 3.5 is the only model family tested that supports internal chain-of-thought reasoning. This was disabled because enabling it caused the models to reason endlessly without producing a structured output. ChatGPT Health may use hidden reasoning ("gpt-5-mini thinking backbone" per the original study).

Models evaluated: Qwen 3.5-9B, Qwen 3.5-4B, Qwen 3.5-2B, Qwen 3.5-0.8B, Llama 3.1-8B, Llama 3.2-3B, Llama 3.2-1B, Mistral-7B, BioMistral-7B, Meditron3-8B.

Ramaswamy A, Tyagi A, Hugo H, et al. ChatGPT Health performance in a structured test of triage recommendations. Nature Medicine. 2026. doi:10.1038/s41591-026-04297-7

Analysis: GLMM (binomial, vignette random intercepts), temperature = 0, Holm–Bonferroni correction.