Failure Modes of MedGemma-1.5-4B in Safety-Critical Obstetric Triage: Catastrophic Forgetting, Magnitude Hallucination, and the Limits of General Medical Pretraining

Abstract Maternal mortality in rural Pakistan remains critically high, with 44% of deaths attributable to delayed or incorrect first response rather than distance alone. Existing medical AI systems require cloud connectivity and are designed for tertiary care settings, rendering them ineffective in the low-resource environments where intervention is most urgently needed. This paper presents BirthRight AI, an offline-first maternal emergency triage system built on fine-tuned MedGemma 1.5 4B, and introduces the Maternal Emergency Response Chain (MERC) — a rigid 8-field structured output framework designed to enforce clinical protocol adherence in small language models. Baseline evaluation of MedGemma 1.5 4B without fine-tuning revealed critical unsuitability for emergency triage — producing unstructured, conversational outputs incompatible with time-critical clinical decision-making. Through 30 iterative training runs on 1,118 manually curated cases spanning 17 obstetric emergency categories, we document the failure modes of fine-tuning small medical models for safety-critical deployment, including catastrophic forgetting under full fine-tuning, magnitude hallucination under 4-bit quantization, and the dissociation between emergency detection accuracy and correct clinical intervention. The system achieved 96% emergency detection with dual-mode architecture serving both clinical workers and untrained family members. Model weights are intentionally withheld pending clinical validation. We propose that safe edge deployment of medical AI requires either larger unquantized models or narrow single-domain specialist architectures, and that general-purpose medical pretraining introduces distributional interference incompatible with emergency triage specificity.