RonaldKato/MaHealthBiasAudit-v2

MaHealthBiasAudit v2 is a multilingual AI auditing framework that detects and explains language bias in maternal health QA datasets across English, Luganda, Runyankore, and Swahili. It combines statistical, linguistic, and model-level analysis to improve fairness in African health AI. # MaHealthBiasAudit v2 **Multilingual Bias Detection for Maternal Health AI in East Africa** ## Overview MaHealthBiasAudit v2 is the first comprehensive AI auditing framework designed to detect, quantify, and explain language bias in maternal health question-answering systems across four East African languages: English, Luganda, Runyankore, and Swahili. ## Pipeline Stages 1. **INPUT**: Multilingual Maternal Health Dataset (4 languages, 8 categories, ~4,500 answers) 2. **PREPROCESSING**: Normalisation → Tokenisation 3. **MULTI-LAYER BIAS DETECTION**: - Stratum I: Rule-Based Detection (Lexicons, Heuristics, Linguistic Patterns) - Stratum II: Classifier-Based Detection (mBERT, XLM-R, RoBERTa) - Stratum III: Python-Based Detection (Pandas, Scikit-learn, Sentence-Transformers) 4. **CROSS-LINGUAL ANALYSIS**: Bias Comparison, Cultural Adaptation, Consistency Mapping 5. **OUTPUT**: Bias Detection Reports, High-Risk Flagging, Recommendations ## Quick Start **Multilingual Bias Detection for Maternal Health AI in East Africa** **Overview** MaHealthBiasAudit v2 is the first comprehensive AI auditing framework designed to detect, quantify, and explain language bias in maternal health question-answering systems across four East African languages: English, Luganda, Runyankore, and Swahili. This toolkit addresses a critical gap in AI fairness evaluation for low-resource African languages in healthcare contexts. **Dataset** dataverse.harvard.edu **Why This Matters** Over 150 million people across East Africa speak Bantu languages, yet most health AI systems are trained primarily on English data. This creates significant disparities in: 1)Health information quality across languages 2)Cultural appropriateness of responses 3)Trustworthiness of AI-driven health advice 4)Patient outcomes in multilingual communities **What Makes This Unique** 1. Linguistically-Informed Metrics Unlike generic bias tools, our framework accounts for Bantu langua …