Machine learning model predicts postpartum depression risk using health record data

Postpartum depression (PPD) affects up to 15% of individuals after childbirth. Early identification of patients at risk of PPD could improve proactive mental health support. Mass General Brigham researchers developed a machine learning model that can evaluate patients' PPD risk using readily accessible clinical and demographic factors. Findings demonstrating the model's promising predictive capabilities are published in the American Journal of Psychiatry.

"Postpartum depression is one of the biggest challenges that some parents may experience in the period after childbirth—a time when many cope with sleep deprivation, new stresses, and significant life changes," said lead author Mark Clapp, MD, MPH, of the Department of Obstetrics and Gynecology at Massachusetts General Hospital, a founding member of the Mass General Brigham health care system.

"Persistent feelings of sadness, depression, or anxiety can be more common than many people realize. Our team, under the leadership of Dr. Roy Perlis, undertook this work to better understand which patients may be at higher risk of PPD to help us facilitate strategies and solutions to either prevent PPD or reduce its severity."

Typically, PPD symptoms are evaluated at postpartum visits, which occur six to eight weeks post-delivery. As a result, many parents may struggle for several weeks before receiving mental health support. To help deliver earlier PPD care, the researchers designed a model that requires only information readily available in the electronic health record (EHR) at the time of delivery, including data on demographics, medical conditions, and visit history. This model weighs and integrates these complex variables to more accurately evaluate PPD risk.

To develop and validate the model, the authors used information from 29,168 pregnant patients who delivered at two academic medical centers and six community-based hospitals in the Mass General Brigham system between 2017 and 2022. In this cohort, 9% of patients met the study's criteria for PPD in the six months following delivery.

The researchers used health record data from approximately one-half of the patients to train the model to identify PPD. They then tested the model by asking it to predict PPD in the other half of the patients. The researchers found that the model was effective in ruling out PPD in 90% of cases.

The model showed promise in predicting PPD: nearly 30% of those predicted to be high risk developed PPD within the six months after delivery. The model was about two to three times better at predicting PPD than estimating based on the general population risk.

In further analyses, the researchers showed that the model performed similarly regardless of race, ethnicity, and age at delivery. The study included only those without a previous psychiatric diagnosis to determine if the model can predict PPD even among low-risk patients and to better understand the risk factors that influence PPD outside of prior psychiatric diagnoses.

Notably, scores on the Edinburgh Postnatal Depression Scale acquired in the prenatal period improved the predictive capabilities of the model, highlighting that this existing tool may be useful both pre- and post-delivery.

The researchers are prospectively testing the model's accuracy, an essential step toward real-world use, and working with patients, clinicians and stakeholders to determine how information derived from the model might best be incorporated into clinical practice.

"This is exciting progress toward developing a predictive tool that, paired with clinicians' expertise, could help improve maternal mental health," Clapp said. "With further validation, and in collaboration with clinicians and patients, we hope to achieve earlier identification and ultimately improved mental health outcomes for postpartum patients."