martinfrasch/martinfrasch.github.io: Demographic and socioeconomic factors predict maternal postpartum rehospitalization and depression: a retrospective nuMoM2b dataset study

The advent of artificial intelligence/machine learning (AI/ML) in medicine has opened opportunities for harnessing the power of ML to predict patient outcomes based on various data contained in the electronic medical records (EMRs) such as patient demographics, basic health, and physiological characteristics. Taken by themselves such singular features are not clearly predictive, but combined in an ML modeling framework such so-called "weak learners" have the power to yield highly predictive models of health outcomes. I hypothesized that sociodemographic and basic, easily obtainable health characteristics of pregnant mothers predict strongly the risk of rehospitalization and depression postpartum. Using the nuMoM2b dataset from NICHD and reproducible, open access AI/ML techniques I tested and validated these hypotheses. Among a large and geographically diverse cohort of nulliparous women with singleton gestations, the present findings show that while some of these characteristics highlight the known disparities in health outcomes (race/ethnicity), most of them are modifiable through social policies, psychological and health counseling, and support. These include access to and support during education, good health insurance, alleviating coming out of poverty, smoking, and unhealthy lifestyle habits as well as early recognition, counseling, and treatment of depression and PTSD. Across the eight ML models deployed, the results are mutually reinforcing yielding a clear cohesive picture. This solution can be used to develop a prospective risk prediction for these important health outcomes in pregnant mothers. The strengths of the presented solution are its focus on preventable maternal morbidities, demonstration of the impact on adversely affected populations, and reproducible step-by-step executable, documented open-access code with relatively low computational requirements. This enables easy adoption and further development on this and other similar datasets to study the contributions and predictive power of various factors to maternal morbidities.