Post-traumatic stress after childbirth: promising predictive models, but still limited validation

Models to identify risk after childbirth

The researchers examined studies developing or validating prediction models for post-traumatic stress disorder within 12 months after childbirth. Their search covered ten databases up to August 12, 2025. Risk of bias and applicability were assessed using the PROBAST tool.

These models combine several characteristics to estimate risk. Frequently identified predictors include mode of delivery, social support, pregnancy complications and fear of childbirth. Their inclusion in the models does not demonstrate that each causes the disorder, or that addressing any one of them would be enough to prevent it.

Encouraging performance, not evidence of prevention

The meta-analysis reports an area under the curve, or AUC, of 0.84, with a 95% confidence interval of 0.80 to 0.88. This metric measures the models’ ability to distinguish participants with the disorder from those without it. It does not mean that 84% of individual predictions are correct.

Pooled sensitivity reaches 0.77 and specificity 0.82. These measures describe the ability to detect cases and identify people without the disorder, respectively. On their own, they do not provide the probability that a person flagged as at risk will actually develop the disorder.

The finding therefore concerns predictive performance. The supplied abstract does not report evidence that using these models reduces symptoms, prevents the disorder or improves children’s health.

Validation remains the weak point

Heterogeneity between studies is substantial, with an I² of 89.7%. The average performance therefore masks considerable variability and should not be equated with that of a specific tool in a particular maternity unit.

Calibration and external validation are insufficiently reported. Calibration checks whether predicted risks match the frequencies actually observed. External validation tests whether a model maintains its performance in a population other than the one used to develop it.

The authors call for multicentre validation in large samples, along with better assessment of calibration and clinical usefulness. The aim is to verify that a score can actually guide care, rather than simply distinguish groups within the data studied.

What we can take away

These models offer a potential way to better identify the risk of post-traumatic stress after childbirth. However, this review is not enough to establish that they are ready to guide care or that they prevent the disorder.