Leveraging a network of ehr-linked biobanks to investigate clinical and genetic risk for perinatal depression

Perinatal depression (PND) is a depressive illness occurring during pregnancy and up to one year following childbirth. PND affects 10-20% of mothers and can contribute to maternal mortality and morbidity, as well as adverse outcomes in children. Despite increasing interest in risk factor detection, clinical prediction of PND using electronic health record (EHR)-based models and evaluation of model fairness across sociodemographic strata remains understudied. Most existing work has focused on postpartum depression (PPD), though upwards of 50% of PND onset occurs antepartum. Additionally, although PND shows higher heritability (40-55%) than non-PND major depressive disorder (MDD; 28-44%), its genetic architecture, particularly relative to MDD, is poorly understood. In this work, we present a multifaceted approach to clinical and genetic risk prediction for PND. Using EHR data from the University of California Los Angeles Health system, we built predictive models to forecast PND at a mother’s first prenatal visit. PND was defined as a diagnosis of depression or pregnancy-related mood disturbance/disorder, or a prescription for antidepressant medications in the perinatal period. We found that random forest models robustly predicted PND (Area under the receiver operating characteristic curve [AUROC] = 0.75; 95% confidence interval [CI] = [0.66, 0.84]), highlighting prior psychopathology, social determinants of health (SDoH), and vitals and lab values as important predictive features. We further evaluated model fairness across self-reported ethnoracial groups and SDoH strata, noting similar predictive performance across the board (AUROC: 0.70-0.74) while still observing heterogeneity in feature importances. In the broader PsycheMERGE network which includes genetic data linked to EHR for over 1.5 million individuals, we analyzed PND phenotypes across 8 sites (n=8,163 PND cases; 81,780 non-PND pregnancies; 78,425 non-PND MDD cases). We used PRS-CS to develop a panel of 19 polygenic scores (PGS) from large, publicly-available genome-wide association study summary statistics, spanning psychiatric, hormonal, and pregnancy-related traits. Comparing PND pregnancies to non-PND pregnancies in a random effect meta-analysis, we observed a significant increase in odds of developing PND with increased genetic liability for MDD, bipolar disorder, substance use disorder and other psychiatric traits. We further investigated whether PND and non-PND MDD were distinguishable using PGS, finding a nominally significant negative association between MDD PGS and PND (relative to MDD); strikingly, this association was found in EHR-based cohorts only, but not when including volunteer-based cohorts. When stratifying PND cases by time-of-onset, distinguishing antepartum depression (APD) from PPD, we found that the lower MDD PGS in PND relative to non-PND MDD was driven by PPD, whereas APD was not distinguishable from non-PND MDD. Here, we present multimodal explorations of risk factors governing PND using large and diverse datasets. We demonstrate notable effects of prior psychopathology, clinical risk factors, and SDoH on PND risk. In parallel, we highlight patterns of genetic risk for PND, and how time-of-onset might distinguish PND severity and etiology in relation to the spectrum of mood disorders. Integration of clinical predictive models and genetic risk factors can facilitate a personalized medicine approach for better diagnosis and treatment of PND.