Machine learning helps close gaps in drug safety during pregnancy

JMIR Publications today published a report on evolving evidence gaps in drug safety during pregnancy in its News and Outlook section. In “How Machine Learning Can Help Fill the Evidence Gaps on Drug Safety in Pregnant Women,” health writer Michelle Falci interviews the principal investigators of two projects that use machine learning to analyze large datasets on drug exposures and outcomes to identify and evaluate possible associations.

Pregnant participants will be excluded from clinical trials

Falci reports that medical research has a serious problem of underestimation. Only 4% of clinical trials over the past decade included pregnant women as participants. This trend dates back to 1977. At the time, the U.S. Food and Drug Administration recommended that pregnant women or women who may become pregnant not participate in Phase 1 and 2 clinical trials. This has resulted in gaps in the evidence regarding the safety of medicines for pregnant women (and has contributed to the growing underrepresentation of female participants in research). Efforts have been made to determine the safety of medicines for pregnant and lactating women, but in practice they are inadequate.

Closing the gap with machine learning

Falci takes a closer look at two new efforts to fill this evidence gap. One is the BOOST-HP project, which uses a tree-based approach to data mining. The other is the BIONIC study, which combines causal inference and machine learning. Each approach uses machine learning to do the heavy lifting of analyzing large datasets, allowing researchers to monitor and infer potential causal relationships.

But this type of AI-assisted research would ideally benefit from more data and requires great care, according to BIONIC research leader Cristina Longoplas. Transparency is key, notes Almut G. Winterstein, principal investigator of the BOOST-HP project. She and her team use AI models that can track the decision-making path that leads to model evaluation. If they used “black box” models (systems whose inner workings are opaque or obfuscated), they would run the risk of missing important epidemiological errors. Still, more thoughtful designs of machine learning models and larger, more comprehensive datasets hold great promise in filling this evidence gap.