Moderate evidence
More
Machine learning did not beat a simple screening score at predicting postpartum depression, study of 1,634 women finds
Study title: Dichotomisation and limited antenatal signal constrain prediction of early postpartum depressive symptoms: A secondary analysis of 1634 women
- Evidence type
- Observational study
- Sample
- n = 1,634
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2951cite or check it
Study title: Dichotomisation and limited antenatal signal constrain prediction of early postpartum depressive symptoms: A secondary analysis of 1634 women
A secondary analysis of 1,634 women found that machine learning did not improve on the standard antenatal Edinburgh Postnatal Depression Scale (EPDS) for predicting depressive symptoms within a week of delivery. An 18-variable model scored a mean AUC of 0.708, identical to the raw score, and did not differ from it in any of 20 repeats (DeLong p=0.53 to 0.97).
Read the study abstract
Antenatal screening with the Edinburgh Postnatal Depression Scale (EPDS) guides postnatal follow-up. We quantified its performance against early postpartum depressive symptoms and asked whether psychological measures or machine learning recover information discarded at the cutoff. Secondary analysis of 1634 women assessed with the EPDS and attachment and coping measures in the third trimester, and with the EPDS again within one week of delivery. Antenatal screening (≥ 12) was evaluated against early postpartum s…

The problem
In a secondary analysis of 1,634 women, an 18-variable machine learning model (AUC 0.708) did no better than the raw antenatal Edinburgh score (0.708) at predicting early postpartum depressive symptoms in any of 20 repeats.
In brief
The standard cutoff of 12 or more caught only 33.4% of early postpartum cases, with 90.9% specificity. Of 287 cases, 191 screened negative. Risk still climbed among screen-negative women, from 4.5% at scores of 0-2 to 29.2% at 9-11.
The authors conclude that no threshold gives a useful screening programme. A threshold of 5 captured 81.7% of cases but flagged 54.4% of screen-negative women at 20.3% precision, which approaches universal follow-up.
Main points
- Women completed the EPDS plus attachment and coping measures in the third trimester and the EPDS again within one week of delivery.
- The antenatal cutoff of 12 or more had 33.4% sensitivity, 90.9% specificity, 44.0% positive predictive value and 86.5% negative predictive value.
- 191 of 287 early postpartum cases screened negative on the antenatal screen.
- Among screen-negative women, risk rose from 4.5% (scores 0-2) to 29.2% (scores 9-11).
- Ten algorithms were compared with the antenatal score alone; an 18-variable model matched it at AUC 0.708.
- Without the antenatal score, 17 variables reached an AUC of 0.651.
- A threshold of 5 flagged 54.4% of screen-negative women and captured 81.7% of cases, at 20.3% precision.
The full story
Researchers wanted to know how well antenatal EPDS screening, which guides postnatal follow-up, performs against early postpartum depressive symptoms. They also asked whether psychological measures or machine learning could recover information discarded when the score is dichotomised at a cutoff.
Using data from 1,634 women, they evaluated the antenatal cutoff of 12 or more against a postpartum threshold of 9 or more, characterised risk below the cutoff, and compared ten algorithms with the antenatal score alone. The 18-variable model matched the raw score, and a model without the score fell to 0.651.
The findings suggest risk varies continuously with the antenatal score, but a dichotomised screen identifies only a minority of cases. Capturing most of the remainder would require following up a large share of women.
Sources
1 on file- Journal of Affective Disordersdoi:10.1016/j.jad.2026.122731
Talk it over
The 18-variable machine learning model scored a mean AUC of 0.708, identical to the raw antenatal Edinburgh score. What does that tie suggest about where the useful information sits, and what would you try next?
Machine learning did not beat a simple screening score at predicting postpartum depression, study of 1,634 women findshttps://truvace.com/item/dichotomisation-and-limited-antenatal-signal-constrain-prediction-of-early-postp
The 18-variable machine learning model scored a mean AUC of 0.708, identical to the raw antenatal Edinburgh score. What does that tie suggest about where the useful information sits, and what would you try next?
Answer in the threadA threshold of 5 captured 81.7% of cases but flagged 54.4% of screen-negative women at 20.3% precision. How would you weigh catching more cases against following up so many women who did not develop symptoms?
Answer in the threadOf 287 early postpartum cases, 191 screened negative at the cutoff of 12 or more. If you were one of those women, what would you want your care team to do differently?
Answer in the thread
If you have more time
Look up how antenatal screening is followed up where you live. Which women would a threshold of 5 pull into follow-up, and who would have to decide whether that is workable?
Answer in the threadThe outcome was measured within one week of delivery and overlaps with postpartum blues. What kind of later measurement of these 1,634 women would settle whether the screen misses lasting depression?
Answer in the thread
ace
What readers say