Designing Around Stigma: Human-Centered LLMs for Menstrual Health

Paper Title

Designing Around Stigma: Human-Centered LLMs for Menstrual Health

Publication Info

• Topic area: Human-centered design of conversational AI for menstrual health education in stigmatized, low-resource contexts.
• Keywords: Menstrual health, conversational AI, stigma-aware design, large language models, Retrieval-Augmented Generation, cultural sensitivity, Roman Urdu, Pakistan, health education, privacy.

Background and Problem

• Problem / challenge: Menstrual health education in Pakistan is constrained by cultural taboos, misinformation, and the absence of formal curricula, leaving young women with limited trusted resources. Existing AI tools often fail to align with local cultural and linguistic contexts, producing misaligned or harmful outputs.
• Significance: Addressing menstrual health gaps is critical for improving women’s health, dignity, and autonomy in conservative, low-resource settings. Digital tools offer potential for private, stigma-sensitive education.
• Motivation and related work: Prior HCI work has explored digital tools for reproductive health, but these often target Western or secular contexts, neglecting the cultural and religious norms of Muslim-majority countries like Pakistan. Existing chatbots lack localization for Roman Urdu and fail to address cultural myths, limiting their effectiveness in taboo-laden domains.

Solution

• Proposed approach: A WhatsApp-based chatbot powered by a large language model (LLM) and Retrieval-Augmented Generation (RAG), co-designed with Pakistani college women to support menstrual health education.
• Novelty:
  1. Development of a stigma-aware design framework for conversational AI in culturally sensitive contexts.
  2. Localization for Roman Urdu and integration of a gynecologist-curated knowledge base to ensure cultural and medical relevance.
  3. Empirical insights from in-the-wild deployment, highlighting user interaction patterns and trust-building mechanisms.
• Procedure and key techniques:
  1. Conducted six co-design workshops (N=30) to identify design requirements, including language preferences, platform accessibility, and trust barriers.
  2. Built a WhatsApp-based chatbot using a RAG framework, integrating a locally validated knowledge base and language classification for Roman Urdu.
  3. Deployed the chatbot with 13 participants over two weeks, collecting 403 messages and conducting semi-structured interviews to evaluate usability, cultural fit, and trust.

Results

• Concrete findings:
  • Participants exchanged 403 messages, with queries covering menstruation physiology, pain, hygiene, myths, and reproductive health.
  • The chatbot reframed "normal suffering" as legitimate health concerns, introducing medical terminology and validating user experiences.
  • Users engaged in iterative questioning, building health literacy through follow-up queries.
• Advantage over baselines:
  • Compared to generic LLMs, the chatbot provided culturally aligned, concise, and medically validated responses.
  • Users preferred its private, judgment-free space over public or family discussions and other digital sources like Google or Snapchat’s ‘My AI’.
• Experiments / evaluation:
  • Conducted a two-week deployment with 13 participants (ages 18–22) from public-sector women’s colleges in Lahore, Pakistan.
  • Evaluated through interaction logs and interviews, focusing on accuracy, cultural relevance, and trust.
• Limitations and future work:
  • Small, demographically narrow sample limits generalizability.
  • Short deployment duration may not capture long-term engagement.
  • Continuous human oversight for safety and cultural alignment limits scalability.
  • Future work should explore broader populations, longer deployments, and additional health domains like fertility and menopause.

Summary

This study presents a culturally localized, LLM-powered chatbot for menstrual health education in Pakistan, addressing gaps in trusted, stigma-sensitive resources. By integrating Roman Urdu support, a gynecologist-curated knowledge base, and a WhatsApp deployment, the chatbot provided private, judgment-free learning spaces for young women. Empirical findings highlight how the chatbot reframed myths, validated health concerns, and fostered trust through cultural alignment. While limited by sample size and scalability challenges, this work contributes a stigma-aware design framework for conversational AI in patriarchal, low-resource contexts, with potential applications in broader reproductive health domains.