The study presents the Mother‑Child AI Agent (MoChiAgent), an LLM‑based clinical assistant that coordinates multiple tools to integrate sequential electronic health record (EHR) data, including routine laboratory tests, for forecasting maternal and infant diseases. The predictive core, MoChiFormer, was developed and internally evaluated using 4,401,599 longitudinal clinical visits and externally validated on independent maternal and infant cohorts with 263,452 and 23,192 visits, respectively. MoChiFormer performs reconstruction of missing laboratory values, reduces batch effects and learns EHR representations that support gestational, fetal and infant age estimation, health‑trajectory modelling and stratification of current and future disease risk. A companion Knowledge Search Tool uses the model forecasts to retrieve evidence‑based intervention and treatment recommendations from curated medical literature and authoritative guidelines. Key reported performance metrics include AUROCs of 0.89 for placental abruption, 0.89 for premature rupture of membranes and 0.91 for preterm labour. Paired mother–infant analyses demonstrated transgenerational risk associations; infants born to mothers in certain clusters had elevated risks of neonatal jaundice (HR 2.81, 95% CI 2.60–3.03) and haematological diseases (HR 2.83, 95% CI 2.62–3.05). Integration of maternal gestational EHRs with infant records improved prediction of infant conditions such as chromosomal abnormalities and respiratory disorders.
The system architecture centers on MoChiFormer, a transformer‑based predictive engine embedded within the broader MoChiAgent framework. MoChiAgent functions as an LLM‑based clinical assistant that orchestrates multiple tools: the predictive engine for EHR representation and forecasting, a data processing pipeline to handle missing and heterogeneous laboratory data, and a Knowledge Search Tool that maps model outputs to intervention and treatment recommendations sourced from curated literature and guidelines. The source describes MoChiFormer as responsible for reconstructing missing laboratory values, reducing batch effects and learning representations suitable for multiple downstream clinical tasks.
Model development used a large internal set comprising 4,401,599 longitudinal clinical visits. External validation employed independent maternal and infant cohorts totaling 263,452 and 23,192 visits, respectively. The source specifies these visit counts but does not provide further demographic breakdowns, geographic origin or detailed inclusion and exclusion criteria in the summarized content.
MoChiFormer includes methods to reconstruct missing laboratory values and to mitigate batch effects that arise from heterogenous data collection. These preprocessing steps enable the model to learn stable EHR representations across sequential visits and variable data completeness. The reconstruction and batch‑correction approach supports downstream tasks such as accurate age estimation and trajectory modelling for gestational, fetal and infant states.
Using the learned EHR representations, the system was evaluated on multiple maternal prediction tasks. Reported discriminative performance for key gestational conditions includes AUROCs of 0.89 for placental abruption, 0.89 for premature rupture of membranes, and 0.91 for preterm labour. The source frames these results as evidence that MoChiFormer can accurately identify important gestational conditions from routine EHR data and reconstructed laboratory information. Additional performance metrics or calibration details are not provided in the summary text.
The study analyzed paired mother–infant records to examine transgenerational risk patterns. Clustering of maternal gestational EHR trajectories identified groups of mothers whose infants experienced substantially elevated risks for specific neonatal conditions. Reported hazard ratios include an increased risk of neonatal jaundice (HR 2.81, 95% CI 2.60–3.03) and haematological diseases (HR 2.83, 95% CI 2.62–3.05) among infants born to mothers in particular clusters. These findings indicate that maternal EHR patterns during pregnancy can convey information about infant risk beyond individual, cross‑sectional indicators.
MoChiAgent evaluated the benefit of combining maternal gestational EHRs with infant records. The integrated approach improved prediction of infant conditions, including chromosomal abnormalities and respiratory disorders, compared with models using infant data alone. The source emphasizes that leveraging longitudinal maternal data adds predictive value for early infant health outcomes. Details on magnitude of improvement, statistical tests or ablation analyses are not included in the provided summary.
In addition to prognostic modelling, MoChiAgent incorporates a Knowledge Search Tool that uses model forecasts to retrieve intervention and treatment recommendations from curated medical literature and authoritative guidelines. This component is intended to translate risk forecasts into actionable, evidence‑based decision support. The summary notes the existence of this retrieval tool but does not report quantitative evaluation of its accuracy, relevance ranking or clinician usability testing.
The authors suggest that MoChiAgent can deliver clinically relevant, actionable decision‑support to facilitate risk‑stratified care for mothers and infants by identifying high‑risk gestational conditions, revealing transgenerational risk patterns and enhancing infant outcome prediction through integrated maternal–infant EHRs. The presented results include strong discrimination for several maternal outcomes and notable hazard ratios linking maternal clusters to infant morbidity. The source text does not report prospective clinical trials, real‑world deployment outcomes, user‑acceptance studies, regulatory pathways or cost‑effectiveness analyses. Those implementation and validation steps appear to be beyond the scope of the summarized content and were not reported in the source.