Study: Predictive Screening Models Less Likely to Widen Existing Health Disparities in PMAD Diagnosis

Perinatal mood and anxiety disorders (PMADs) include a range of mental health disorders—including major depressive disorder and various anxiety disorders—that occur during pregnancy or up to a year postpartum that affect up to 20% of birthing parents. Postpartum depression (PPD) is the most common PMAD, affecting approximately 13% of birthing parents in high-income countries and 20% of birthing parents in low-income and middle-income countries. Nearly half of PMADs are never identified, despite recommendations and guidelines currently in place. Authors of a study published in JAMA Network Open expand on available literature by evaluating potential bias in predictive models of PMADs trained on commonly available electronic health records (EHRs).

Image credit: grooveriderz | stock.adobe.com

For this study, the investigators included screening data from individuals who gave birth and were admitted to the postpartum unit or maternal-fetal care unit following delivery. Edinburgh Postnatal Depression Scale (EPDS), which was originally developed to screen for perinatal depression, was used to assess PMADs. This scoring method also assesses further for anxiety symptoms. Patients indicated on a 4-point Likert scale the frequency with which they experienced symptoms (eg, “I have blamed myself unnecessarily when things go wrong”) within the past 7 days. Total scores ranged from 0 to 30, with higher scores indicating greater symptoms of PPD.

Additionally, race and ethnicity were provided by the patient during intake or later through their patient portal. Patients selected from applicable categories which included the following (patients were categorized as multiple races if they identified as such): African American or Black, American Indian or Alaskan Native, Asian or Native Hawaiian or Other Pacific Islander, White, and other (not further specified). For ethnicity, patients selected from either Hispanic or non-Hispanic. Additionally, none of the enrolled patients identified as American Indian or Alaskan Native. If a patient declined to answer, they were categorized as “unknown race” which did not exclude their data from analysis.

The target variable was patients’ screening results either from the EPDS or Patient Health Questionnaire 9 (PHQ-9), who were dichotomized into low risk (negative) or moderate to high risk (positive) based on the clinical cutoffs used. Additionally, both measures screened for suicidal ideation, which were determined with a PHQ-9 score off 5 or higher or a EPDS score of 8 or higher.

The following 3 supervised classification algorithms were fitted to the data: logistic regression, random forest, and extreme gradient boosting. Data were pulled from 19,790 EHRs, and patients were screened for depression no more than 9 days following delivery.

A total of 19,430 unique patients were enrolled in the study. Participants had a median age of 34.1 years (range: 14-59 years) at the time of first delivery. Among this population, approximately 56% (n = 10,942) identified as non-Hispanic White, 12% (n = 2371) as Asian American and Pacific Islander, 11% (n = 2146) as other or not further specified, 10% (n = 1842) as Hispanic White, 7% (n = 1402) as African American or Black, 3% (n = 606) as multiple races, and less than 1% (n = 121) did not provide this information. According to the investigators’ sample, racial and ethnic minority patients were more likely than non-Hispanic White patients to screen positive on both the PHQ-9 (odds ratio, 1.47 [95% CI, 1.23-1.77]) and EPDS (odds ratio, 1.38 [95% CI, 1.20-1.57]) scales.

The present models were comparable with or outperformed other studies that utilized similar target and predictor features (mean area under the receiver operating curve [AUROCs] between 0.602 and 0.635). The mean test AUROCs of each model with (0.610 to 0.635) or without (0.602 to 0.622) reweighing the samples in the training set. Without reweighing, the models had predicted greater positive rates for racial and ethnic minor patients compared with non-Hispanic White patients (mean DP, 0.238 [95% CI, 0.231-0.244]; P < .001). Additionally, they also displayed lower false negatives for racial and ethnic minorities (mean difference, −0.184; [95% CI, −0.195 to −0.174]; P < .001).

The investigators noted that the primary limitation was the models may not have been accurate across all patient demographics. Additionally, reweighing data based on expected and observed frequencies in the data may potentially re-bias against certain subgroups within the study.

REFERENCE

Wong EF, Saini AK, Accortt EE, Wong MS, Moore JH, Bright TJ. Evaluating Bias-Mitigated Predictive Models of Perinatal Mood and Anxiety Disorders. JAMA Netw Open. 2024;7(12):e2438152. doi:10.1001/jamanetworkopen.2024.38152