Bridging the Trust Gap: Clinician-Validated Hybrid Explainable AI for Maternal Health Risk Assessment in Bangladesh

Hybrid explanations outperform single approaches: The study found that combining two explanation methods—fuzzy logic rules and SHAP feature importance—was preferred by 71.4% of clinicians across clinical cases, suggesting that different explanation types serve complementary purposes in building trust for high-stakes medical decisions.

Trust remains conditional even with explainability: While the hybrid XAI framework achieved strong technical performance (88.67% accuracy), only 54.8% of healthcare professionals expressed trust sufficient for clinical use, highlighting that explainability alone is insufficient—clinicians require additional validation and integration with their existing knowledge.

Healthcare access emerged as the strongest predictor: SHAP analysis revealed that access to healthcare services was the primary factor driving risk predictions, not clinical measurements alone, suggesting that AI systems for maternal health in resource-constrained settings must account for structural barriers alongside medical factors.

Clinician feedback identified critical data gaps: Healthcare professionals identified missing variables—particularly obstetric history, gestational age, and connectivity barriers—that the model couldn't capture, demonstrating the importance of involving end-users in AI development to ensure systems address real-world clinical needs.

Local validation is essential for adoption: The study's systematic engagement with 14 Bangladeshi clinicians revealed context-specific concerns (connectivity barriers) and preferences that might not emerge from technical validation alone, suggesting responsible AI deployment requires localized testing with actual end-users.