🤖 AI

WHBench: Evaluating Frontier LLMs with Expert-in-the-Loop Validation on Women's Health Topics

The paper introduces WHBench, a specialized benchmark comprising 47 expert-crafted scenarios across 10 women's health topics that evaluates 22 frontier large language models and reveals that even the best-performing models fail to exceed 72.1% accuracy, highlighting significant gaps in clinical safety, equity, and guideline adherence that necessitate expert oversight for deployment.

Original authors: Sneha Maurya, Pragya Saboo, Girish Kumar

Published 2026-04-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sneha Maurya, Pragya Saboo, Girish Kumar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're asking a super-smart, well-read robot for advice on a very personal and complicated medical issue, like whether it's safe to get a vaccine while trying to have a baby. You hope the robot knows the latest rules, understands your unique situation, and doesn't accidentally give you dangerous advice.

This paper introduces a new "report card" called WHBench (Women's Health Benchmark) to test exactly how good these AI robots are at answering those specific questions.

Here is the story of the paper, broken down into simple parts:

1. The Problem: The "Textbook" vs. The "Real World"

Think of previous AI tests like MedQA as a multiple-choice quiz. They ask the AI: "What is the correct dosage for drug X?" The AI just has to pick the right letter (A, B, C, or D). It's like a student memorizing a textbook.

But real life isn't a multiple-choice quiz. Real life is a messy conversation. A patient might say, "I'm 35, I have a history of blood clots, I'm on a tight budget, and I'm worried about my race affecting my care. What should I do?"

The authors realized that while AI is great at picking letters, it often fails at:

Old News: Using medical rules that changed five years ago.
Missing the Details: Forgetting that a patient's age or weight changes the answer.
The "Gender Gap": Medical research has historically ignored women, so the AI's "brain" has blind spots about women's health.
Equity: Failing to understand that a poor patient might not be able to afford the "perfect" treatment.

2. The Solution: A New "Driving Test"

The authors created WHBench, which is like a driving test for AI, rather than a written exam.

The Questions: They wrote 47 real-life scenarios (like a patient asking about fertility or contraception).
The Experts: Instead of just using a textbook, they hired real doctors (OB/GYNs, oncologists, nurses) to write the "perfect" answers.
The Trap: Each question was designed to catch a specific mistake. For example, one question is designed to see if the AI gives an outdated dosage, while another checks if it ignores racial health disparities.

3. The Grading System: The "Safety First" Rubric

How do you grade an AI's answer? The authors created a strict checklist with 23 rules.

The "One Strike" Rule: If the AI gives dangerous advice (like a wrong dosage), it gets a massive penalty. It's like a driving test where you get an automatic fail if you run a red light, even if you parked perfectly.
The "Fairness" Check: They added a special section to see if the AI considers things like insurance costs or race. This is the first time a medical AI test has really focused on this.
The Judges: They used two other super-smart AIs to grade the answers, but they made sure the graders didn't know which model they were judging (to keep it fair).

4. The Results: The AI is Still a "Learner"

They tested 22 of the smartest AI models available (including the latest from OpenAI, Google, and Anthropic). Here is what they found:

No One Passed with Flying Colors: The highest score any model got was 72.1%. In a school, that's a "C." In medicine, that's not good enough to trust blindly.
The "Safe" vs. "Smart" Gap: Some models were very good at sounding smart but gave dangerous advice. Others were safe but missed important details.
The "Blind Spot" is Huge: Every single model failed miserably at understanding social factors (like poverty or race). They were great at saying "don't be racist" (using polite words), but terrible at actually doing something to help a patient who can't afford their medicine.
The Top Tier is Clumped: The best models are all very close to each other. There is no "superhero" AI yet that dominates the field.

5. The Big Takeaway

The paper concludes that we cannot trust AI to give medical advice on women's health yet.

Think of these AIs like interns in a hospital. They have read all the books and can recite facts, but they haven't learned how to handle the messy, emotional, and complex reality of real patients. They often miss the "human" part of the equation.

Why does this matter?
If we let these AIs give advice without a human doctor checking their work, we risk:

Giving outdated advice that hurts patients.
Ignoring the specific needs of women from different backgrounds.
Creating a false sense of security.

The authors released this test (WHBench) to the public so that developers can use it to fix these flaws. It's a tool to help them build AI that is not just "smart," but also safe, fair, and ready for the real world.

1. Problem Statement

Large Language Models (LLMs) are increasingly used for medical information retrieval, yet existing benchmarks fail to adequately evaluate their performance in women's health. This domain presents unique challenges that current medical benchmarks (e.g., MedQA, MMLU) do not capture:

Dynamic Guidelines: Clinical protocols (e.g., thyroid thresholds in pregnancy) shift frequently, leading to outdated model knowledge.
Complex Reasoning: Treatment decisions depend on intersecting patient factors (age, BMI, race, history) requiring integrative reasoning rather than simple pattern matching.
Gender Data Gap: Medical literature historically underrepresents women, causing models trained on this data to inherit blind spots regarding health disparities.
Safety Risks: Errors in women's health (e.g., fertility, pregnancy) can have immediate, severe consequences, yet no benchmark specifically targets failure modes related to safety, dosage, or health equity in this domain.

2. Methodology

A. Benchmark Design (WHBench)

The authors constructed a benchmark consisting of 47 expert-crafted clinical scenarios covering 10 women's health topics (Fertility, Hormonal Health, Pregnancy, PCOS, Contraception, Endometriosis, Cancer Screening, Vaginal Health, Mental Health, Bone Health).

Failure-Mode Targeting: Each question is explicitly mapped to a specific LLM failure mode (Table 1), including:
- Missing information (e.g., omitting follow-up).
- Factual/Outdated guidelines (e.g., pre-2017 TSH thresholds).
- Health equity gaps (ignoring racial disparities).
- Incorrect treatment, contraindications, or dosage errors.
Difficulty Calibration: Questions range from Level 2 (factual recall) to Level 5 (frontier reasoning with conflicting evidence).
Expert Reference Answers: Each question was answered independently by 4–6 board-certified clinicians (OB/GYNs, gynecologic oncologists, fertility nurses, etc.). These answers serve as the ground truth, averaging 94 words (concise, clinical style).

B. Scoring Rubric

Evaluation is based on a 23-criterion rubric organized into 8 categories (Table 2). Key design features include:

Asymmetric Penalties: Safety failures carry heavy negative weights (e.g., -5 for unsafe commission) compared to positive weights for passing, ensuring a clinically dangerous response cannot achieve a high score.
Equity Dimension: The benchmark includes the first dedicated Equity & Inclusivity dimension in a medical LLM benchmark, assessing social determinants of health (F18a) and bias avoidance (F18b).
Ordinal Scoring: Four criteria use 3-level ordinal scales to capture partial credit.
Normalization: Raw scores range from -58 to +92, normalized to a 0–100% scale. Responses are classified as Correct (≥80%), Partially Correct (45–79%), or Incorrect (<45%).

C. Evaluation Protocol

Models Evaluated: 22 models across frontier (GPT-5.4, Claude Opus 4.6, etc.), reasoning-specialized, and open-source categories.
Setup: Zero-shot, closed-book setting with a system prompt instructing the model to act as a board-certified women's health physician. Temperature set to 0.
Judging Pipeline:
- Primary Judge: Claude Sonnet 4.6 operates under a "default to fail" philosophy.
- Secondary Judge: GPT-5.4 for inter-rater reliability.
- Blinding: Judges receive the question, expert references, and a checklist but not the model name.
- Verification: Raw scores are recalculated server-side from pass/fail verdicts to prevent arithmetic errors.

3. Key Contributions

Failure-Mode-Targeted Scenarios: 47 open-ended questions designed to expose specific LLM weaknesses (outdated info, dosage errors, equity gaps) rather than general knowledge.
Multi-Dimensional Rubric: A novel scoring system with asymmetric safety penalties and the first dedicated Equity & Inclusivity dimension in medical LLM evaluation.
Large-Scale Evaluation: Evaluation of 22 models resulting in ~3,100 scored responses, providing a robust dataset for analysis.
Expert-in-the-Loop Validation: Reference answers authored by a diverse panel of specialists, ensuring clinical realism and accuracy.

4. Key Results

A. Overall Performance

Capability Ceiling: No model exceeded a mean score of 75%. The top model, Claude Opus 4.6, achieved 72.1%.
Low "Fully Correct" Rates: Even the best model was fully correct (≥80%) in only 35.5% of cases. This indicates that even state-of-the-art (SOTA) models require human oversight for clinical deployment.
Frontier Clustering: The top 7 models (Frontier tier) are tightly clustered (72.1% to 60.7%), suggesting a performance ceiling not visible in saturated multiple-choice benchmarks.
Performance Drop-off: A significant gap exists between the top tier (avg 65.2%) and the bottom tier (avg 38.6%).

B. Safety and Harm

Variable Harm Rates: Aggregate scores mask significant safety risks. The "Harm%" (percentage of responses with unsafe commission or omission) varies drastically:
- Lowest Harm: Claude Opus 4.6 (12.8%).
- Highest Harm: Gemini 2.5 Pro (90.8%) and Llama 4 Maverick (83.7%).
Safety vs. Accuracy: High aggregate scores do not guarantee safety; for example, GPT-5.4 scored 66.8% overall but had a 47.5% harm rate.

C. Category and Topic Analysis

Equity Blind Spot: Performance on Social Determinants of Health (F18a) was universally poor (0.7%–19.1% pass rate), whereas Bias Avoidance (F18b) was high (78%–92.9%). Models can avoid biased language but fail to proactively integrate structural barriers (insurance, race, access) into advice.
Completeness: Models frequently omitted follow-up monitoring and alternative treatment plans (B7), a critical gap for shared decision-making.
Topic Variance: Contraception was the most difficult topic overall. Pregnancy and Cancer Screening showed high variance, likely due to guideline recency and interpretation complexity.

D. Inter-Rater Reliability

Label Agreement: Modest at the final label level (Correct/Partial/Incorrect) with $\kappa = 0.238$ .
Ranking Stability: Very strong consistency in model ranking ( $\rho = 0.916$ ), indicating the benchmark is stable for comparing systems despite noisy individual response labels.
Dimensional Agreement: Highest on concrete criteria (Instruction Follow, Completeness) and lowest on subjective dimensions (Equity, Uncertainty).

5. Significance and Implications

Clinical Reality Check: WHBench demonstrates that high performance on exam-style benchmarks (MedQA) does not translate to safe, open-ended clinical counseling. The gap between "recognition" (multiple choice) and "generation" (clinical advice) is substantial.
Safety is Non-Linear: The benchmark reveals that safety is not a byproduct of general intelligence; models can be highly capable yet dangerously unsafe in specific contexts (e.g., dosage errors, missed contraindications).
Equity Gap: The results highlight a critical failure in current LLMs to address social determinants of health, suggesting that training on existing medical literature perpetuates systemic blind spots.
Future Direction: The paper argues that medical fine-tuning alone is insufficient; future models need specialized reasoning capabilities, better handling of evolving guidelines, and explicit training on health equity.

Conclusion: WHBench establishes a new standard for evaluating clinical AI in women's health, revealing that even the most advanced frontier models are not yet ready for autonomous clinical use due to persistent safety risks and equity gaps. It calls for continued expert oversight and the development of benchmarks that prioritize safety and inclusivity over raw knowledge retrieval.