MoChiAgent, short for Mother-Child AI Agent, is built around an LLM that orchestrates several tools over time-ordered EHR data, routine laboratory results included. Its outputs reach beyond risk scores, as a knowledge-search component fetches treatment guidance from peer-reviewed literature and clinical guidelines. The system is described in Nature Medicine, published September 4.
Its core engine, MoChiFormer, trained on 4.4 million longitudinal clinical visits and validated externally on 263,452 maternal and 23,192 infant visits. For gestational complications, the model reached AUROCs of 0.89 for placental abruption, 0.89 for premature rupture of membranes and 0.91 for preterm labor.
The team also found transgenerational signals. Infants born to mothers in high-risk clusters faced sharply elevated risk of neonatal jaundice (hazard ratio 2.81) and hematological diseases (hazard ratio 2.83). Adding maternal records improved prediction of infant conditions such as chromosomal abnormalities and respiratory disorders.
The work comes from researchers at Chongqing Medical University, Shanghai Jiao Tong University, Macau University of Science and Technology and collaborators in the US and Australia. It argues routine visit data can substitute for costly tests and imaging in pregnancy risk stratification. In physician evaluations, MoChiAgent’s case assessments were scored alongside ChatGPT and Gemini, part of an effort to show clinicians how the assistant reaches its forecasts.
