In This article

• AI and machine learning models can use screening tools and EHR data to detect postpartum depression with good accuracy, though they currently mirror rather than surpass traditional tools.
• A recent systematic review shows many AI models achieve AUROC values around 0.80–0.90 for identifying postpartum depression, using EPDS, PHQ-9, ICD codes, or diagnoses as reference standards.
• A large EHR-based study of more than 29,000 births demonstrated that a machine-learning model could stratify postpartum depression risk at discharge with AUC 0.721 and strong calibration.
• These models could be especially useful for flagging “low-risk” women who might otherwise be overlooked and for differentiating who needs routine follow-up versus targeted preventive interventions.
  
• More research is needed to externally validate models, assess clinical impact, and determine whether AI-guided risk stratification improves treatment access and outcomes beyond existing screening practices.

Even though ChatGPT hit the market only a few years ago, it has been widely adopted by the American public. According to data from the Pew Research Center (2025), 58% of adults under 30 and 41% of adults between the ages of 30 and 49 have used ChatGPT for any purpose.

While ChatGPT is used for many different purposes, survey data from OpenAI indicate that 60% of U.S. adults reported using AI tools for healthcare questions in the past 3 months, asking ChatGPT about symptoms (55%), medical terms (48%), and treatment options (44%). To be more specific, 40 million Americans turn to ChatGPT for answers to their healthcare questions each day.

Understandably, many health care providers, as well as those in the general public, have raised concerns regarding the use of ChatGPT to provide health information and to provide diagnoses and treatment recommendations. Where does the medical data come from? Will users receive accurate, evidence-based information about their medical conditions? Will they delay seeking evidence-based treatments? Will they be targeted by companies and providers of alternative, unproven interventions? Particularly potent are recent stories of individuals seeking emotional support for mental health challenges, where the use of AI platforms has led to delays in receiving effective treatment, worsening symptoms, and, at the worst, self-harm and suicide.

While there may be risks to the use of unsupervised, untested, or “free-range” AI platforms, there is also immense potential for innovative machine learning models to sort through large quantities of data and to identify complex and not readily apparent trends that could be medically useful. Recent studies suggest that AI and machine learning models may help to improve the identification of individuals with perinatal mood and anxiety disorders. Despite efforts encouraging universal screening of pregnant and postpartum women over a period of many decades, postpartum depression and other psychiatric challenges associated with pregnancy and childbirth frequently go unrecognized and untreated.

AI for the Detection of Postpartum Depression

A recent systematic review explored the emerging role of AI in the detection of postpartum depression, asking whether this technology could be used to overcome the limitations of conventional screening methods. Currently, screening tools such as the Edinburgh Postnatal Depression Scale (EPDS) facilitate the process of early identification of women with postpartum depression and anxiety. Building on this information, researchers have been able to develop more accurate prediction models by integrating information from the electronic health record (EHR).

The review and meta-analysis concluded that many AI and machine learning models exhibited high predictive performance for PPD, with several studies showing area under the receiver operating characteristic curve (AUROC) values in the 0.80–0.90 range (indicating that the model was able to identify about 80% to 90% of cases). However, these models often used EPDS or PHQ-9 scores, ICD codes, or clinician diagnoses as the reference standard, meaning that AI was usually trained to reproduce what routine assessments would yield rather than to outperform conventional screening tools or an independent gold standard.

Several studies also used data from social media platforms, like Facebook, gathering information from posts written by new mothers to improve the identification of individuals with postpartum depression. While text describing mental health struggles can be used to flag individuals with depressive symptoms and anxiety, researchers note that more nuanced and complex linguistic changes may also be used to identify depression and anxiety.

Only a handful of the studies included in the meta-analysis compared AI model metrics (sensitivity, specificity, AUROC) head-to-head with the EPDS and/or PHQ; AI models were at least non-inferior and sometimes modestly superior in terms of AUROC or sensitivity.

Can AI Be Used to Prevent Postpartum Depression?

In a similar way, machine learning models, supplemented with data from the EHR, could be trained to identify women at increased risk for PPD. Armed with information collected from the medical record, sociodemographic information, obstetric history, previous mental health diagnoses, and data from screening tools, AI algorithms could be trained to predict risk of psychiatric symptoms across pregnancy and the postpartum period rather than at a single visit.

One such example is a retrospective cohort study in which Clapp and colleagues used routinely collected EHR data from over 29,000 deliveries across eight Boston-area hospitals (2017–2022) to train and externally validate a machine-learning model for predicting postpartum depression among individuals without known high-risk psychiatric histories. The model drew on sociodemographic variables, medical history, and prenatal depression screening results available at the time of discharge and defined postpartum depression as a new mood disorder diagnosis, antidepressant prescription, or EPDS score >13 within six months.

Trained on roughly half the cohort and tested on the remaining 14,150 patients, the model showed reasonable discrimination (AUC 0.721) and good calibration (Brier score 0.087), indicating that predicted risks aligned well with observed outcomes. Its positive predictive value was 28.8% (about one in three labeled high-risk went on to develop postpartum depression), while the negative predictive value was 92.2%, suggesting it was particularly useful for ruling out low-risk patients.

Overall, the findings support the feasibility of leveraging EHR-based AI models to identify previously unflagged individuals at elevated risk for postpartum depression before they leave the hospital. In this way, these models could allow us to stratify risk so that we could distinguish women who need routine monitoring from those who might benefit from a targeted intervention to reduce the risk of PPD.

What’s Next?

The findings of the current systematic review and pilot studies indicate that AI and machine learning models could be used to reliably identify women with postpartum depression. There is no indication at this point that AI and machine learning can outperform conventional screening tools; however, the studies are preliminary, and the models thus far have been trained to match, rather than outperform, current screening tools. Further evaluation would be required to confirm diagnosis in order to make appropriate treatment recommendations.

There are also emerging data to support the use of AI to facilitate the early identification of women at increased risk of PPD, with markedly improved predictive performance compared to traditional risk checklists. At the present time, we use checklists and clinical judgment to predict risk for postpartum psychiatric illness after delivery; however, our ability to accurately predict risk is very limited and largely unstudied.

While previous studies have documented that the women at highest risk for PPD are those who experience depressive symptoms during pregnancy—about 10% to 15% of women—we have not been as successful at defining risk in the remaining 85% of women. This type of model would help us to better define risk in populations that would normally be considered to be at low risk. Thus, AI-based modeling could be especially valuable for predicting risk in women with no history of psychiatric illness.

Ideally, we would like to be able to identify women at risk for postpartum depression before it occurs. This would not only allow us to increase monitoring when needed and to treat early if PPD emerges, but it may also provide an opportunity to initiate preventative 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 (for example, age and BMI) to improve our ability to predict and quantify risk.

—Ruta Nonacs, MD PhD