Modeling Multiple Adverse Pregnancy Outcomes: Learning from Diverse Data Sources
摘要
We consider the problem of modeling adverse pregnancy outcomes (APOs) from diverse data sets and aim to understand what is common between them and what is unique for each of these data sets. To this effect, we consider three different data sets (a clinical study from the US, EHRs from a US hospital, and a clinical study in India) and model three specific APOs - preterm birth, new hypertension, and preeclampsia. Since LLMs can efficiently summarize the scientific literature, we use them to generate initial hypotheses and use the different data sets to refine the hypotheses to create joint probabilistic models (as Bayesian networks). Our analyses show that there are eight relationships between risk factors common to all three populations and some unique relationships for specific populations.