In the United States, a pregnant person will attend approximately 15 prenatal visits with a medical provider for monitoring an uncomplicated pregnancy. During these visits, a tremendous amount of demographic and clinical information is collected and entered into the electronic health record (EHR). Much of the information relates to pregnancy monitoring, such as weight gain, blood pressure, urinalysis; however, there is information in the medical record that can be used to predict the risk of perinatal depression. Several recent studies have sifted through this information from the electronic health record, using machine learning to generate algorithms that can be used to estimate the risk of postpartum depression.
Estimating the risk of PPD in new mothers
In a retrospective cohort study analyzing data from the NIH Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be, Wakefield and colleagues examined the medical records of 10,038 new mothers. They tested the performance of four different risk prediction models:
- Model 1 used only readily available sociodemographic data.
- Model 2 also included data on maternal mental health before pregnancy.
- Model 3 used recursive feature elimination to construct a parsimonious model.
- Model 4 further titrated the input data to simplify the pre-pregnancy mental health variables.
The analysis included 8,454 births; 338 (4%) received treatment for depression during the postpartum period (as documented in the EHR). In terms of predicting women who would later require treatment, model 3 performed best, with an area under the receiver operating characteristic curve of 0.91 (± 0.02). This means that this model would identify 91% of women who ultimately need treatment for depression.
The models identified nine variables that were the most robust predictors of postpartum depression treatment: history of maternal depression (highest), any current mental health problems, recent psychiatric medication use, BMI, income, age, history of anxiety, education level, and preparedness. for pregnancy (lowest).
Estimation of the risk for PPD with model including EPDS scores
In another study, Amit and colleagues analyzed EHR data from 266,544 women in the United Kingdom who had their first child between 2000 and 2017. A subgroup of 5959 women also had Edinburgh Postnatal Depression Scores (EPDS) scores recorded in the EHR. The researchers extracted multiple sociodemographic and medical variables and constructed a machine learning model to predict the risk of PPD during the year after giving birth.
In this cohort, the prevalence of PPD was 13.4%. PPD was defined based on the occurrence of any of the following documented in the EHR during the first year postpartum: (1) diagnosis of depression; (2) new antidepressant treatment; or (3) nonpharmacological treatment of depression.
In a model that used only EHR data, the area under the curve (AUC) of the prediction model ranged from 0.72 to 0.74. Interestingly, the model worked quite well when only pre-pregnancy data was used to predict risk; the EPD-based prediction model applied before pregnancy identified at least 70% of women who were later diagnosed with PPD.
When the model combined EHR-based data with EPDS scores, the area under the receiver-operator characteristic curve (AUC) increased from 0.805 to 0.844, with a sensitivity of 0.76 at a specificity of 0.80. In other words, the best predictive model would be able to identify at least 80% of women who would later be diagnosed with PPD.
The factors most strongly associated with risk of PPD included history of antidepressant use, history of depression, number of antidepressant prescriptions filled, younger age, BMI, smoking status, and deprivation index.
Can we use the electronic medical record to predict the risk of PPD?
The answer is yes. Both studies indicate that machine learning can be used to construct a model using data collected from the EHR that can be used to predict the risk of depression within the first year after giving birth. In these two studies, the predictive models were able to identify between 84% and 91% of women at risk of developing postpartum depression. Statisticians get quite excited about screening tools like the AUC is greater than 0.8.
Now for the caveats. Both studies document the diagnosis or treatment of PPD using documentation in the electronic health record: documentation of the diagnosis itself or the treatment (antidepressants or non-pharmacological treatment). While the prevalence of PPD in the US-based Wakefield study was 4%, the Amit study from Britain reported that the prevalence of PPD was 13.4%.
Based on previous epidemiological studies, the prevalence of PPD is typically around 15%. There are no studies suggesting that the prevalence of PPD is lower in the US than in Britain; It should be noted that the American study only examined the prevalence of therapy for PPD, while the British study looks at the prevalence of diagnosis and/or treatment. This discrepancy is consistent with previous studies indicating low treatment rates among women with PPD, with lower treatment rates reflecting underdiagnosis of perinatal mood and anxiety disorders, as well as barriers to accessing treatment in the US.
The models described in these studies are likely to identify individuals at risk more severe PPD and not women with less severe depressive symptoms. While identifying women with the most severe symptoms is essential to limit morbidity in both mother and child, we may be missing an opportunity to support other mothers who are also struggling during the postpartum period. Further studies are needed to test these predictive models in more diverse populations, including multiparous mothers, and to use a broader definition of perinatal depression.
Ideally, we would like to identify women at risk for postpartum depression before it occurs. This would not only allow us to increase monitoring when necessary and treat early if PPD occurs, but it may also provide an opportunity to initiate preventive interventions. Currently, our strongest predictors of risk include a history of depression prior to pregnancy and depressive symptoms during pregnancy. These models build on these robust risk factors and include other risk factors (such as age, BMI) to improve our ability to predict and quantify risk.
Ruta Nonacs, MD PhD
References
Wakefield C, FraschMG. Predicting patients requiring treatment for depression in the postpartum period using common electronic health record data available antepartum. AJPM focus. Apr 27, 2023;2(3):100100.
Amit G, Girshovitz I, Marcus K, Zhang Y, Pathak J, Bar V, Akiva P. Estimating the risk of postpartum depression from electronic health records using machine learning. BMC Pregnancy Childbirth. 2021 Sep 17;21(1):630.
