Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Surgeons Are Indian Males and Speech Therapists Are White Females: Auditing Biases in Vision-Language Models for Healthcare Professionals

Zohaib Hasan Siddiqui, Dayam Nadeem, Mohammad Masudur Rahman, Mohammad Nadeem, Shahab Saquib Sohail, Beenish Moalla Chaudhry

TL;DR

This study audits demographic biases in vision-language models for healthcare professions using a profession-aware taxonomy, a balanced FairFace dataset, and four off-the-shelf VLMs (CLIP and OpenCLIP). It employs neutral prompts and top-$k$ retrieval alongside JS-divergence-based bias scores to quantify gender, race, and age associations across 33 healthcare roles. The findings show systematic but model-dependent biases, with age bias being especially dominant and notable intersectional patterns that could impact hiring and workforce analytics. The work highlights the need for model-specific fairness audits, intersectional benchmarks, and governance to ensure equitable and trustworthy AI deployment in healthcare.

Abstract

Vision language models (VLMs), such as CLIP and OpenCLIP, can encode and reflect stereotypical associations between medical professions and demographic attributes learned from web-scale data. We present an evaluation protocol for healthcare settings that quantifies associated biases and assesses their operational risk. Our methodology (i) defines a taxonomy spanning clinicians and allied healthcare roles (e.g., surgeon, cardiologist, dentist, nurse, pharmacist, technician), (ii) curates a profession-aware prompt suite to probe model behavior, and (iii) benchmarks demographic skew against a balanced face corpus. Empirically, we observe consistent demographic biases across multiple roles and vision models. Our work highlights the importance of bias identification in critical domains such as healthcare as AI-enabled hiring and workforce analytics can have downstream implications for equity, compliance, and patient trust.

Surgeons Are Indian Males and Speech Therapists Are White Females: Auditing Biases in Vision-Language Models for Healthcare Professionals

TL;DR

This study audits demographic biases in vision-language models for healthcare professions using a profession-aware taxonomy, a balanced FairFace dataset, and four off-the-shelf VLMs (CLIP and OpenCLIP). It employs neutral prompts and top- retrieval alongside JS-divergence-based bias scores to quantify gender, race, and age associations across 33 healthcare roles. The findings show systematic but model-dependent biases, with age bias being especially dominant and notable intersectional patterns that could impact hiring and workforce analytics. The work highlights the need for model-specific fairness audits, intersectional benchmarks, and governance to ensure equitable and trustworthy AI deployment in healthcare.

Abstract

Vision language models (VLMs), such as CLIP and OpenCLIP, can encode and reflect stereotypical associations between medical professions and demographic attributes learned from web-scale data. We present an evaluation protocol for healthcare settings that quantifies associated biases and assesses their operational risk. Our methodology (i) defines a taxonomy spanning clinicians and allied healthcare roles (e.g., surgeon, cardiologist, dentist, nurse, pharmacist, technician), (ii) curates a profession-aware prompt suite to probe model behavior, and (iii) benchmarks demographic skew against a balanced face corpus. Empirically, we observe consistent demographic biases across multiple roles and vision models. Our work highlights the importance of bias identification in critical domains such as healthcare as AI-enabled hiring and workforce analytics can have downstream implications for equity, compliance, and patient trust.

Paper Structure

This paper contains 21 sections, 8 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. Taxonomy
  5. Dataset
  6. VLMs used
  7. Prompt Design
  8. Bias identification
  9. Performance measures
  10. Findings
  11. VLMs do show bias
  12. Bias analysis
  13. Gender
  14. Race
  15. Age
  16. ...and 6 more sections

Figures (6)

  • Figure 1: Gender and race bias in vision language model
  • Figure 2: Overall pipeline of our adopted methodology
  • Figure 3: Representative samples from the FairFace dataset across varying demographic attributes including gender, race and age.
  • Figure 4: The top-10 roles in order where the gender disagreement between VLMs was highest.
  • Figure 5: Bias scores for demographic attributes for all VLMs.
  • ...and 1 more figures