Prediction models for postpartum post-traumatic stress disorder: a systematic review and meta-analysis

Abstract

Background

Despite the increasing number of studies on prediction models for identifying the risk of postpartum post-traumatic stress disorder (PP-PTSD), the quality and clinical applicability of these models have not been clarified yet.

Objectives

To systematically review and appraise the prediction models for PP-PTSD.

Methods

From inception to March 30, 2026, Web of Science, PubMed, the Cochrane Library, Embase, Scopus, CNKI, WANFANG and VIP were searched for primary studies involving development or validation of prediction models. The Prediction Model Risk of Bias Assessment Tool (PROBAST) was used to evaluate the quality of the literature for issues of bias and applicability. The area under the curve (AUC) of externally validated models was extracted for meta-analysis. The protocol for this study is registered with PROSPERO (registration number: CRD42025630468).

Results

Sixteen studies comprising 29 models met the inclusion criteria. Thirteen studies were from China, and 13 models applied machine learning. All studies were judged to have high risk of bias, primarily due to data source quality and inadequate reporting of outcomes and analysis. Fourteen studies(87.5%) were rated as having low concern regarding applicability. Seven studies (43.8%) performed external validation, including only two studies validating for ML models. The pooled AUC of these validated logistic regression models was 0.86 (95% CI: 0.79–0.90). Moderate heterogeneity was observed across these studies (I² = 70.8%).

Conclusion

Current PP-PTSD prediction models show preliminary predictive potential but suffer from significant methodological limitations. Future research needs to follow the TRIPOD guidelines and obtain reliable models through rigorous study design and external validation. This will facilitate external validation of the models in the next step, so that it can be promoted to diverse populations.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12888-026-08247-y.

Introduction

Post-traumatic stress disorder (PTSD) is a psychological condition that arises after trauma, affecting emotions, thoughts, and behavior in daily life and social interactions [1]. Although childbirth is comonly viewed as a positive experience, it can also be a highly stressful event with physical and psychological risks, including PTSD [2]. Studies showed that approximately one-third of women perceived childbirth as traumatic, and about a quarter experienced significant PTSD symptoms after delivery [3, 4]. Postpartum post-traumatic stress disorder (PP-PTSD) is a specific form of PTSD triggered by childbirth, characterized by persistent and repetitive re-experiencing of traumatic events, avoidance and emotional numbing, and persistent increased vigilance [5].

Recent studies revealed that PP-PTSD affected a significant portion of new mothers, with an estimated prevalence of 10% in the general perinatal population [6]. Among high-risk pregnancies, the rates were even higher, ranging from 16% to 41.2% [7, 8]. These findings highlighted PP-PTSD as a critical public health issue, particularly due to its associations with adverse maternal and infant outcomes, including preterm birth, gestational diabetes, pre-eclampsia, and increased mortality risks [9]. Beyond immediate obstetric complications, PP-PTSD negatively impacted child development. When mothers experienced trauma during the perinatal period, it disrupted the infant’s stress response system, weakened mother-infant bonding, and contributed to insecure attachment patterns, ultimately affecting the child’s emotional, social, and cognitive development [10–12]. Even at sub-clinical levels, PP-PTSD symptoms persisted, with detectable psychological distress lasting up to two years postpartum [13].

The significant prevalence and clinical impact of PP-PTSD highlight the urgent need for effective screening tools. These tools could help find out vulnerable women, reduce exposure to known risk factors, and facilitate timely interventions [14]. According to the Diagnostic and Statistical Manual of Mental Disorders(DSM) published by the American Psychiatric Association [1], the diagnostic criteria for PP-PTSD require that the patient has experienced core symptoms for more than one month. Physicians diagnose the condition using high-standard structured clinical diagnostic interviews or patient self-report questionnaires [15, 16]. In clinical practice, the management of PP-PTSD primarily focuses on secondary prevention, due to low public awareness and understanding of PP-PTSD. Most postpartum women do not receive timely screening and assessment for PP-PTSD.

Furthermore, evidence remained limited regarding early risk detection and primary prevention [17]. As a trauma and stress disorder with delayed psycho-pathological reactions, PP-PTSD is influenced by multiple factors, including three categories: prenatal susceptibility, delivery trauma, and postpartum maintenance [18]. Several studies have developed risk prediction models using logistic regression by incorporating subjective and objective factors related to these three categories, in order to achieve early identification of women at high risk of PP-PTSD [19, 20]. Moreover, with the application of artificial intelligence and machine learning (ML) algorithms to the diagnosis and prediction of diseases in recent years, opportunities have opened up for dealing with health problems such as PP-PTSD [21]. However, these studies may have some sources of bias in data collection, study design, and analysis that would reduce the validity of these research findings. Several studies [20, 22, 23] developed models based on data from retrospective cohort, leading to recall bias and lack of blinding in predictor assessment [24]. Rousseau et al. [25] and Bartal et al. [26] used machine learning methods such as decision tree and natural language processing to develop models for predicting PP-PTSD, but did not perform external validation. The predictive ability of these models in different populations remained unknown. Concerns have been raised about their applicability in clinical practice [27]. Considering the rising interest in studies developing PP-PTSD risk prediction models, the growing concerns suggest the importance of critically examining potential biases in these studies.

The purpose of this review is to systematically evaluate the development, performance, and validation of the existing prediction models for PP-PTSD, providing theoretical references for future model improvement and clinical practice.

Methods

The transparent reporting of multivariable prediction models for individual prognosis or diagnosis: checklist for systematic reviews and meta-analyses (TRIPOD-SRMA) [28] was followed in the reporting of this study.

Search strategy

The databases Web of Science, PubMed, the Cochrane Library, Embase, Scopus, CNKI, WANFANG and VIP were searched. Searches were conducted from the inception of the databases until March 30, 2026. The keywords were “postpartum OR postnatal OR perinatal OR maternal OR pregnan* OR prenatal OR antenatal”, “post-traumatic stress disorder OR PTSD OR stress disorder OR trauma after childbirth”, and “prediction model OR risk model OR predictive model OR prognostic model OR machine learning”. The search strategies were detailed in the Appendix 1. In addition, the reference lists of retrieved studies and review articles were screened to identify relevant studies.

As the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies (CHARMS) recommended [29], the PICOTS system [27] were utilized to frame the aim, search strategy, study inclusion and exclusion criteria. Key items of this review were outlined below:

• P (Population): Women during pregnancy or postpartum.

• I (Intervention model): Prognostic prediction models developed or validated for PP-PTSD (predictors ≥ 2).

• C (Comparator): No competing model.

• O (Outcome): The occurrence of PP-PTSD.

• T (Timing): Within one year after childbirth.

• S (Setting): Different settings including hospital, communities, et al.

Inclusion and exclusion criteria

The inclusion criteria: (1) studies involving development or validation of prognostic models with multivariable (≥ 2 predictors); (2) participants comprising women diagnosed with PP-PTSD within one year after childbirth; (3) primary outcome explicitly reporting the occurrence or risk of PP-PTSD; (4) studies providing detailed statistical models, instruments, or predictive indicators for estimating individual risk of PP-PTSD; (5) tested individual level predictive performance (e.g., discrimination, calibration).

The exclusion criteria consisted of studies that: (1) focused only on diagnostic models rather than prognostic models; (2) evaluated multiple predictive variables without reporting subgroup analysis or specific evaluation outcomes; (3) focused on PTSD occurring prior to delivery rather than postpartum; (4) were published in languages other than English or Chinese; (5) reported findings from duplicate cohorts; in such cases, the study with the largest sample size will be prioritized. If sample sizes are equal, the earliest published study will be selected; (6) consisted of secondary research formats, such as literature reviews, meta-analyses, protocols, quality improvement reports, editorials, commentaries, letters to the editor, or theses that were purely based on secondary analysis without original empirical data; (7)were unavailable in full text.

Study selection

In accordance with the predefined selection criteria, one reviewer(S.Z.) evaluated the search strategy across all databases. The records were imported into EndNote X9 for finding duplicates and preliminary screening. Two reviewers(S.Z. and L.K.) independently screened titles and abstracts against the inclusion and exclusion criteria. Any discrepancies in article selection was resolved through deliberation. When consensus could not be reached, a third reviewer(N.M.S.) made the final determination.

Data extraction

Extraction of key items was performed by two reviewers (S.Z. and L.K.) independently using the electronic form based on CHARMS. These items were categorized into 11 dimensions, including data source, participants, outcomes to be predicted, candidate predictors, sample size, missing data, model development, model performance, model evaluation, results, interpretation and discussion. A third researcher (N.M.S.) randomly checked about 20% of the extracted articles, and reconciled extracted data by comparing independent extractions and mediating discrepancies through consensus discussions. Unresolved disagreements were escalated to the co-investigator (N.M.S.) for final arbitration.

Quality assessment

To assess potential bias risk in the studies, the reviewers used the PROBAST checklist [24]. Meanwhile, PROBAST + AI checklist [30] was applied to evaluate the studies involving machine learning algorithms. The assessment tools comprised four domains: participants, predictors, outcome, and analysis. For each domain, whether bias and applicability should be classified as low, high, or unclear was determined by answering signaling questions. For each question, possible answers included “Yes”, “Probably yes”, “No”, “Probably no”, or “No information”. A domain was considered to have a potential for bias, if there was any signaling question answered with “No” or “Probably no”. Applicability concerns were graded as high when the population, predictors, or outcomes of the included studies differed from those specified in the review question. Only when all domains exhibited low risk, the overall risk of bias could be deemed low. Two reviewers (S.Z. and L.K.) independently evaluated bias presence and applicability concerns. Three studies were randomly selected from those included, and re-assessed the risk of bias. When any conflict of opinion occurred between the two reviewers, it was resolved by the co-investigator (N.M.S.) after a rigorous discussion to reach a final agreement.

Data synthesis and statistical analysis

A meta-analysis was performed on the area under the curve (AUC) values derived from external-validated models, using Stata software (version 17.0; Stata Corporation, College Station, Texas, USA) with the metan module. The AUC and its 95% confidence interval were pre-logit-transformed to improve the normality of effect sizes. The standard errors were derived as (transformed upper confidence limit - transformed lower confidence limit) / 2*1.96 [31]. Given the presence of heterogeneity, a random effects model was applied, where the weights were based on the within-study error variance [27]. Heterogeneity was assessed through the I² index and Cochrane Q test. The I² index indicated heterogeneity, where values of 25%, 50%, and 75% represented low, moderate, and high heterogeneity, respectively [32]. To assess the robustness of the pooled AUC, the leave-one-out method was employed to conduct sensitivity analysis. When the number of included studies exceeds 10, Egger’s test is used to assess publication bias [33].

Results

Study selection

The systematic literature search initially identified 167 records from the selected databases. After removing 68 duplicates identified across all databases, the unique records underwent title and abstract screening based on predefined eligibility criteria. Of these, twenty-four full-text articles were retained for further evaluation. Following comprehensive evaluation, four studies were excluded because they reported multiple predictive variables without specific evaluation outcomes or complete models. Bartal et al. reported two studies from duplicate cohorts, so the one with a smaller sample size was excluded. Qi et al. and Chen et al. respectively had studies from duplicate cohorts with same sample, and the later published ones were both excluded. One study did not specify when PTSD was measured in the abstract. To further verify whether it was eligible, we reviewed the full text and found that PTSD assessments mostly occurred more than one year after childbirth. Therefore, this study was excluded. Ultimately, sixteen studies with 29 models were included in this review. For more details, please go to Fig. 1, which gives a Preferred reporting items for systematic review and meta-analysis (PRISMA) process of screening and selecting studies.

Fig. 1.

PRISMA 2020 search flow diagram

Study characteristics

Basic characteristics

The 29 included models were developed using 15 data sources, which encompassed data from 10,435 women (range 122 to 1,752) across four different countries. Among those studies, there were 13 studies from China [19, 22, 23, 34–43], one from USA [26], one form Israel [25], and one from Spain [20]. The publication years ranged from 2021 to 2025, with the majority published from 2023 to 2025 (n = 13). The study designs consisted of retrospective cohort studies (n = 8) and prospective cohort studies (n = 8). Regarding the study population, most of the studies recruited postpartum women (n = 13), while 3 studies recruited pregnant women in the last trimester of pregnancy and followed them up to postpartum. One study merely included women after caesarean section, three study focused on women with complicated pregnancy, while the rest studies paid attention on common population who were without severe medical issues or complicated obstetric issues. Table 1. summarizes the basic characteristics of the included studies.

Table 1.

Basic characteristics of included studies

No.	Author (Year)	Study design	Enrolment period	Country	Data source	Characteristics of participants	Outcome	Number of participants/ outcomes(%)
1	Hernández-Martínez, A. (2021)	Retrospective cohort	2018 to 2019	Spain	Not report	Postpartum women	PP-PTSD within 1 year postpartum	1752/248 (14.16%)
2	Rousseau, S. (2022)	Prospective cohort	2016 to 2017	Israel	One large public hospital	Women in the last trimester of pregnancy without premature birth, multiple pregnancy, or known severe medical issues	Stable high post-traumatic-stress-following-childbirth during 4 months postpartum	182/31 (17.03%)
3	Li, Xiaoyan (2023)	Retrospective cohort	January 2021 to December 2022	China	The postpartum examination outpatient department of a hospital	Postpartum women over 20 years old and without severe medical issues	PP-PTSD at 42 day postpartum	719/94 (13.07%)
4	Meng, Yixin (2023)	Retrospective cohort	March 2022 to July 2022	China	The maternity ward of a hospital	Postpartum women who underwent cesarean section	PP-PTSD at 42 day postpartum	122/24 (19.67%)
5	Qi, Junli (2024)	Retrospective cohort	August 2022 to May 2023	China	The maternity outpatient departments of two hospitals	Postpartum women over 20 years old and without serious medical issues	PP-PTSD at 42 day postpartum	682/54 (7.92%)
6	Wang, Xiaoting (2024)	Prospective cohort	November 2022 to October 2023	China	The obstetrics departments of Two tertiary-level hospitals	Pregnant women over 20 years old and without premature birth, multiple pregnancy, or severe medical issues	PP-PTSD at 42 day postpartum	420/46 (10.04%)
7	Peng, Rui (2024)	Retrospective cohort	June 2021 to June 2023	China	One hospital	Postpartum women over 18 years old and without severe physical or mental health issue	PP-PTSD within 1 year postpartum	458/55 (12.01%)
8	Bartal, A. (2024)	Retrospective cohort	November 2016 to April 2017, April 2020 to December 2020	United States	Hospital announcements, social media, and professional organizations	Women who gave birth to a live baby in the last six months and were at least 18 years old	PP-PTSD within 6 months postpartum	1295/190 (14.67%)
9	Chen, Jingfen (2024)	Prospective cohort	January 2022 to April 2023	China	The Postnatal ward and outpatient department of a tertiary maternity and child specialized hospital	Postpartum women over 18 years old and without multiple pregnancy, stillbirth, severe physical or mental health issue, or adverse newborn outcome	PP-PTSD at 42 day postpartum	900/99 (11.00%)
10	Li, Zhumei (2024)	Prospective cohort	March 2023 to December 2023	China	The obstetrics department of a tertiary hospital	High-risk pregnant women in the last trimester of pregnancy	PP-PTSD within 6–8 weeks postpartum	532/115 (21.62%)
11	Wang, Xiaoting (2025)	Prospective cohort	September 2023 to July 2024	China	The obstetrics departments of three tertiary-level hospitals	Pregnant women over 20 years old and without premature birth, multiple pregnancy, or severe medical issues	PP-PTSD at 42 day postpartum	718/87 (12.12%)
12	Guo, Yikun (2025)	Prospective cohort	Febrary 2023 to Octorber 2023	China	The obstetrics departments of three hospitals	Postpartum women over 18 years old, married and without serious medical issues	PP-PTSD within 6–8 weeks postpartum	1340/127 (9.48%)
13	Mao, Lichao (2025)	Prospective cohort	August 2023 to December 2024	China	The obstetrics department of a hospital	Advanced-age primiparous women without adverse life events or prenatal mental health conditions	PP-PTSD at 42 day postpartum	203/31 (15.27%)
14	Li, Xianan (2025)	Prospective cohort	December 2023 to August 2024	China	The obstetrics departments of two hospitals	Postpartum women over 18 years old, experiencing vaginal delivery and without severe physical or mental health issue	PP-PTSD at 1 month after the mother’s discharge from hospital	400/54 (13.50%)
15	Wang, Xinling (2025)	Retrospective cohort	January 2023 to December 2023	China	The obstetrics department of a hospital	Postpartum women over 18 years old and without severe mental illnesses or adverse pregnancy outcomes	PP-PTSD within 6–8 weeks postpartum	180/87 (48.33%)
16	Lu, Liuyi (2025)	Retrospective cohort	September 2022 to September 2023	China	The NICUs of three tertiary hospitals	Mothers of very low birth weight infants	PP-PTSD within 1 month postpartum	532/144 (27.07%)

Model development and validation

Logistic regression was the most commonly used modeling method (n = 13), while machine learning algorithms were also employed in four studies with 13 models. The selection of candidate predictors mainly relied on univariable analysis with significance thresholds varying between P < 0.05 and P < 0.25. Across the 16 studies included, a total of 45 predictors were were reported by the final models. The predictor most frequently included in the final model was mode of delivery (n = 7), followed by social support (n = 6), antenatal childbirth fear (n = 5), postpartum depression(n = 4) and quality of sleep in late pregnancy (n = 4). Pregnant complications, parity, postpartum anxiety, neonatal feeding practices, skin-to-skin contact, neuroticism and spousal support were reported as final predictors three times. In addition, educational qualifications, employment status, household income, pregnancy plan, adverse pregnancy history, pain on the third day postpartum, admission of newborns to care unit were mentioned in two studies each.

Among the included studies, validation strategies were limited, with internal validation methods such as random split data, bootstrapping and k-fold cross-validation reported in most studies (n = 9). External validation was conducted in 7 studies, including geographical (n = 4), and temporal (n = 2) validation. Only 2 studies contained both internal and external validation. Across 4 studies applying machine learning, half of them were validated externally, but one did not conduct any validation. Table 2. presents the detailed characteristics of the prediction models developed in the included studies.

Table 2.

Model development and validation information of the included studies

No.	Author (Year)	Modeling methods	Number of candidate predictors/predictors included in the final model	EPV	Selection of candidate predictors	Selection of final predictors	Handling of missing data	Type of validation	Performance measures	Model presentation format
1	Hernández-Martínez, A. (2021)	Logistic regression	11/6 (model A) 13/6 (model B)	22.5 (model A) 19.1 (model B)	Univariable analysis (P < 0.25)	Backward elimination	No information	Only external: Temporal	Cal: HL test Disc: AUC graph Ov: R-squared CU: Not evaluated	Logistic regression equation
2	Rousseau, S. (2022)	ML algorithms (Decision tree)	36/6	0.9	Not applicable	Not applicable	Complete-case analysis	Only internal: K-fold cross-validation	Cal: Not evaluated Disc: Not evaluated Ov: Accuracy CU: Not evaluated	Score chart
3	Li, Xiaoyan (2023)	Logistic regression	6/6	15.7	Univariable analysis	Unclear	No information	Only internal: Random split data	Cal: Calibration plot / Slope / HL test Disc: AUC graph Ov: Not evaluated CU: Not evaluated	Logistic regression equation, nomogram
4	Meng, Yixin (2023)	Logistic regression	5/3	4.8	Univariable analysis (P < 0.05)	Unclear	No information	None	Cal: HL test Disc: AUC graph Ov: Not evaluated CU: Not evaluated	Logistic regression equation
5	Qi, Junli (2024)	Logistic regression	9/5	9.2	Univariable analysis (P < 0.05)	Unclear	No information	Internal: Bootstrap, External: Geographical	Cal: Calibration plot / HL test Disc: AUC graph Ov: Not evaluated CU: DCA	Logistic regression equation
6	Wang, Xiaoting(2024)	Logistic regression	8/6	8.4	Univariable analysis (P < 0.05)	Unclear	Complete-case analysis	Only external: Geographical	Cal: Calibration plot / HL test Disc: AUC graph Ov: Not evaluated CU: DCA	Logistic regression equation, nomogram
7	Peng, Rui (2024)	Logistic regression	7/5	7.9	Univariable analysis (P < 0.05)	Unclear	No information	Only internal: Random split data	Cal: Calibration plot / HL test Disc: AUC graph Ov: Not evaluated CU: Not evaluated	Logistic regression equation, nomogram
8	Bartal, A. (2024)	ML algorithm based natural language processing	Not applicable	Not applicable	Not applicable	Not applicable	Single imputation	Only internal: Cross-validation	Cal: Not evaluated Disc: AUC graph Ov: Not evaluated CU: Not evaluated	Code
9	Chen, Jingfen (2024)	Logistic regression, ML algorithms (logistic regression, decision tree, random forest, support vector machine, and eXtreme Gradient Boosting)	27/7a 27/9b 27/9c 35/8d 35/7e 35/7f 35/6g 35/8h	3.7a, b, c 2.8d, e, f, g, h	Univariable analysis for logistic regression models (P < 0.05); Filter, Boruta and LASSO for ML models	Stepwise selectiona, Forward selectionb, Backward eliminationc, Based on AUC size and number of predictor variablesd, e, f, g, h	Multiple imputation	Only external: Temporal	Cal: Calibration plot Disc: AUC graph Ov: Brier score CU: DCA	Logistic regression equation, online calculator
10	Li, Zhumei (2024)	Logistic regression	16/6	7.2	Univariable analysis (P < 0.1)	Unclear	Complete-case analysis	Only internal: Bootstrap	Cal: Calibration plot / HL test Disc: AUC graph Ov: Not evaluated CU: DCA	Logistic regression equation, nomogram
11	Wang, Xiaoting (2025)	Logistic regression	5/5	17.4	Lasso regression	10-fold cross-validation	Complete-case analysis	Internal: Bootstrap External: Geographical	Cal: Calibration plot / HL test Disc: AUC graph Ov: Not evaluated CU: DCA	Logistic regression equation, nomogram
12	Guo, Yikun (2025)	Logistic regression	17/8	7.5	Univariable analysis (P < 0.05)	Unclear	No information	Only external: Geographical	Cal: Calibration plot / HL test Disc: AUC graph Ov: Not evaluated CU: DCA	Logistic regression equation, nomogram
13	Mao, Lichao (2025)	Logistic regression	9/5	3.4	Univariable analysis (P < 0.05)	Unclear	No information	None	Cal: Calibration plot / HL test Disc: C-Statistic/ AUC graph Ov: Not evaluated CU: Not evaluated	Nomogram
14	Li, Xianan (2025)	ML algorithms (random forest, support vector machine, logistic regression and K-nearest neighbor)	17/6	3.2	Univariable analysis (P < 0.05) and Lasso regression	Unclear	Single imputation	Only external: Geographical	Cal: Not evaluated Disc: AUC graph Ov: Not evaluated CU: Not evaluated	Logistic regression equation
15	Wang, Xinling (2025)	Logistic regression	12/7	7.3	Univariable analysis (P < 0.05)	Unclear	Complete-case analysis	Only internal: Bootstrap	Cal: Calibration plot Disc: AUC graph Ov: Not evaluated CU: DCA	Logistic regression equation, nomogram
16	Lu, Liuyi (2025)	Logistic regression	13/6	11.1	Univariable analysis (P < 0.05)	Unclear	No information	Only internal: Bootstrap	Cal: Calibration plot / HL test Disc: AUC graph Ov: Not evaluated CU: DCA	Logistic regression equation, nomogram

EPV, the number of events-per-variable. ML, machine learning; Cal, calibration; Disc, discrimination; Ov, overall fit; CU, clinical utility

^a tranditional logistic regression model 1; ^b tranditional logistic regression model 2; ^c tranditional logistic regression model 3; ^d machine learning logistic regression model; ^e Decision tree model; ^f Random forest model; ^g Support vector machine model; ^h eXtreme Gradient Boosting model

Model performance

Model discrimination was commonly evaluated via the area under the receiver operating characteristic curve (AUC), which is equivalent to the C-statistic for binary outcomes. AUC values of 29 developed models ranged from 0.6 to 0.969, which were reported in 15 studies except for Rousseau et al.’s. Table 3. lists the AUC values included in the model. For calibration measures, Hosmer-Lemeshow(HL) test was the most commonly method, with 10 studies reporting. Ten studies provided calibration plots, while 3 studies did not report information about calibration evaluation. Only 8 studies applied Decision Curve Analysis (DCA) to evaluate model clinical utility. Table 2 shows the specific content.

Table 3.

AUC reported by the included studies

No.	Model	Model development AUC(95%CI)	Internal validation AUC(95%CI)	External validation AUC(95%CI)
1	Hernández-Martínez, A., 2021 (model A)	0.70 (0.67–0.74)		0.69 (0.63–0.75)
2	Hernández-Martínez, A., 2021 (model B)	0.82 (0.79–0.85)		0.83 (0.78–0.87)
3	Rousseau, S., 2022
4	Li, Xiaoyan, 2023	0.930 (0.905–0.955)	0.900 (0.850–0.950)
5	Meng, Yixin, 2023	0.805
6	Qi, Junli, 2024	0.944 (0.911–0.976)		0.851 (0.761–0.941)
7	Wang, Xiaoting, 2024	0.928 (0.885–0.971)		0.907 (0.836–0.978)
8	Peng, Rui, 2024	0.865 (0.830–0.895)	0.838 (0.801–0.878)
9	Bartal, A., 2024 (model#1)	0.6
10	Bartal, A., 2024 (model#2)	0.6
11	Bartal, A., 2024 (model#3)	0.8
12	Chen, Jingfen, 2024 (traditional logistic regression model 1)	0.831 (0.785–0.876)		0.602 (0.439–0.771)
13	Chen, Jingfen, 2024 (traditional logistic regression model 2)	0.842 (0.799–0.885)		0.639 (0.495–0.802)
14	Chen, Jingfen, 2024 (traditional logistic regression model 3)	0.845 (0.801–0.888)		0.679 (0.541–0.825)
15	Chen, Jingfen, 2024 (Machine learning logistic regression model)	0.826 (0.807–0.846)		0.850 (0.776–0.923)
16	Chen, Jingfen, 2024 (Decision tree model)	0.803 (0.782–0.825)		0.787 (0.662–0.856)
17	Chen, Jingfen, 2024 (Random forest model)	0.844 (0.825–0.836)		0.768 (0.685–0.851)
18	Chen, Jingfen, 2024 (Support vector machine model)	0.789 (0.767–0.811)		0.811 (0.723–0.889)
19	Chen, Jingfen, 2024 (eXtreme Gradient Boosting model)	0.835 (0.816–0.854)		0.844 (0.766–0.912)
20	Li, Zhumei (2024)	0.827 (0.786–0.868)	0.78
21	Wang, Xiaoting, 2025	0.962(0.947–0.978)		0.934(0.898–0.971)
22	Guo, Yikun, 2025	0.812(0.770–0.854)		0.868(0.760–0.892) 0.785(0.696–0.874)
23	Mao, Lichao, 2025	0.903(0.851 ~ 0.955)
24	Li, Xianan, 2025 (Random forest model)	0.882(0.819–0.945)		0.842
25	Li, Xianan, 2025 (Support vector machine model)	0.807(0.718–0.896)		0.734
26	Li, Xianan, 2025 (K-nearest neighbor model)	0.821(0.773–0.909)		0.725
27	Li, Xianan, 2025 (Logistic regression model)	0.857(0.778–0.936)		0.767
28	Wang, Xinling, 2025)	0.800(0.624–0.901)
29	Lu, Liuyi, 2025)	0.969 (0.953–0.984)

Results of quality assessment

The risk of bias and applicability of the ten studies were as shown in Table 4. All studies were rated as high risk of bias, indicating methodological problems of the model development or validation. In the participant domain, eight studies were assessed as high risk of bias, mainly because of their retrospective study design [20, 22, 23, 26, 34, 35, 42, 43]. In the predictors domain, seven studies did not have relevant information for judging whether predictors were assessed without knowledge of the outcome data, so the risk of bias was assessed as unclear [20, 22, 23, 34, 40, 42, 43]. In the outcome domain, retrospective cohort studies are unable to make sure that researchers defined outcomes without knowledge of predictors, so eight other studies were also rated as high risk [20, 22, 23, 26, 34, 35, 42, 43]. Of those studies, two study did not detail the way in which the outcome was defined [35, 42], one study also did not report detailed methods and processes for outcome definition [35], two studies reported inappropriate time intervals between outcome assessment and predictor assessment [20, 26], and two studies did not provided explicit process of predictors and outcome assessment [42, 43]. Additionally, which could result in a high risk of outcome bias. In the analysis domain, eight studies had insufficient sample sizes [25, 34, 36, 37, 39–42]. Eight studies treated continuous variables inappropriately [19, 22, 35–39, 42]. Most studies did not report explicit information about missing data or handled it in an inappropriate way [20, 22, 23, 25, 26, 34, 35, 39–41, 43]. One study reported missing data of model development, but not did that in model validation [36]. Most of the studies based their initial screening of predictors only on univariable analysis (n = 11). Eventually, all the analysis domains of the included studies were rated as high risk of bias.

Table 4.

Risk of bias and applicability assessment results of the included studies

“+” indicates low ROB/low concern regarding applicability; “-” indicates high ROB/high concern regarding application; “?” indicates unclear ROB/unclear concern regarding applicability; “/” indicates no relevant results were reported. ^aThe updated PROBAST + AI checklist was used to assess the ROB and applicability of the studies using ML

Meta-analysis of validation models

The limited availability of externally validated machine learning models restricted our meta-analysis to logistic regression models from six studies [19, 20, 23, 36, 38, 39]. The model with the best effect size was selected, and the pooled AUC from the random effects model was 0.86 (95% CI: 0.79–0.90) (Fig. 2). The I² index was 70.8% (P = 0.004), indicating apparent heterogeneity between studies. Sensitivity analysis showed that the pooled AUC ranged from 0.83 (95%CI: 0.77–0.88) to 0.88 (95%CI: 0.83–0.91) after sequentially excluding each one of the six models. The pooled AUCs changed more when omitting the studies of Chen et al. and Wang et al. individually (Appendix 2). Due to the small number of studies included in the meta-analysis, our study did not assess publication bias.

Fig. 2.

Forest plot of the random effects meta-analysis of pooled AUC estimates for 6 validation models

Discussion

Main findings

Most of the studies’ publication years were concentrated in 2024 and 2025, which showed that the attention to PP-PTSD has increased dramatically in the last two years. Despite the increasing number of prediction models for PP-PTSD, most of them had significant methodological flaws, such as retrospective cohort study design, inadequate sample size, improper selection of predictors based on univariable analysis and inappropriate handling methods of continuous variables and missing values. Consequently, all these studies were assessed to have a high risk of bias, thereby restricting the applicability of these models in clinical settings. Moreover, the ratio of outcomes to participants in the included studies indicated a higher prevalence of PP-PTSD among women with high-risk pregnancy [34, 37, 40, 43]. Some studies also suggested that women with high-risk pregnancy are more susceptible to PP-PTSD [44]. But to date most models are designed for general population without complications. It is evident that existing studies on prediction models paid limited attention to women with high-risk pregnancy, potentially failing to meet their greater need for PP-PTSD prediction.

In terms of discrimination ability, most of the models reported AUC values greater than 0.75 in development and validation. Although the AUC values seemed to indicate good performance in discrimination, limitations including insufficient avoidance of overfitting and absence of external validation raised concerns regarding the real predictive capabilities of these models and their adaptability to diverse populations. In the meta-analysis, the pooled discrimination performance of validated logistic regression models should be interpreted with caution due to moderate heterogeneity. The heterogeneity was usually affected by differences in populations and study design [27]. The majority of the study population was women without high-risk pregnancy from mainland China and employed prospective cohorts. Meanwhile, differences in outcome assessment might contribute to heterogeneity. After excluding one study [20] that measured PP-PTSD within one year postpartum, the pooled AUC did not show a significant increase. But many studies that measured PP-PTSD at 42 days postpartum reported better discrimination performance. All studies used self-report scales [45–50] to measure PP-PTSD, including the Post-Traumatic Stress Disorder Checklist - Civilian version (PCL-C), Perinatal Post-Traumatic Stress Disorder Questionnaire (PPQ), and others. All studies applied criteria for PP-PTSD based on the Diagnostic and Statistical Manual of Mental Disorders, yet various assessment tools would yielded different ratings, which could also explain high heterogeneity. Among the included studies, the logistic regression model developed by Wang et al. [38] appeared most promising. The reported nomogram may be a helpful visual tool in facilitating clinical practice. However, traditional logistic regression models could not adequately capture complex nonlinear relationships or interactions among predictors, whereas ML excelled in this regard [51]. Externally validated ML models from Chen et al. [36] and Li et al. [41] performed comparably to the pooled AUC of logistic regression models, also indicating good development potential.

Interpretation

The included studies of this review provided some hints for the development and evaluation of future models. Prospective longitudinal cohort study design is the best option for developing prediction models, which can ensure complete measurement of relevant predictors and outcomes [52]. Inadequate sample sizes (EPV < 20) across the most included studies raised concerns about potential overfitting and limited generalizability of the developed models [53]. In fact, prediction models usually have more parameters than actual predictors, events per candidate predictor parameter(EPP) is more important for sample size determination [54]. More large-sample prospective cohort studies are encouraged to develop predictive models so that the models can be more reliable when applied to new individuals.

As calibration plot, the total number of observed and expected events, DCA were not reported in several studies, summarizing estimates of calibration performance and clinical utility was challenging in this review. Although HL test was reported most frequently, it is not ideal significant for small data sets and has limited applicability to the assessment of poor calibration [24]. Some studies [20, 34] reported only the HL test without calibration plots, so they could not provide effective information on the accuracy of the predicted risks. Future studies are recommended to report explicit calibration and clinical utility metrics, which are essential for clinical applicability.

ML algorithms was beginning to be introduced into the field of prediction models for PP-PTSD, showing considerable potential and innovation for identifying risk of PP-PTSD. The results of Chen et al. [36] showed that ML-based logistic regression model outperformed the traditional logistic regression model and other ML models. Rousseau et al. [25] proposed a short set of questions and prediction rules based on decision tree modeling, providing a concise tool for clinical assessment. As well, Bartal et al. [26] developed language models to assess mental health using text-based personal narratives, which could be a patient-friendly strategy to identify risk of PP-PTSD. But of those four studies involving ML algorithms, two studies lacked reporting on model performance, internal or external validation. So more rigorously reported studies are needed, and the generalized applicability of ML models remained to be confirmed by more external validation studies.

The existing models also provided some insights into the prediction of PP-PTSD. Hernández-Martínez et al. [20] found that the model B with psycho-social related predictors performed better than the model A which was based exclusively on clinical variables. Chen et al. [36] included general obstetric information, psycho-social factors, and biochemical test results as candidate predictors, and developed different models using a variety of methods. The results showed that only one indicator from the biochemical test was included into the final prediction models. In sensitivity analysis, Chen’s study [36] clearly reduced the pooled AUC. Since its population, design, and assessment tool were similar to most other studies, this may also be linked to the predictors used in the model. A systematic review of risk factors for PP-PTSD conducted by Ijlas et al. [55] also found that maternal psycho-social health and social support were particularly important in the occurrence of PP-PTSD. This suggested that psycho-social factors may play a more important role in predicting PP-PTSD.

In addition, antenatal childbirth fear and sleep quality in late pregnancy showed significant predictive roles for PP-PTSD in most models. Currently, most prospective cohort studies and retrospective cohort studies recruited participants after their delivery, ignoring various prenatal psycho-social factors. Therefore, in addition to maternal physiological and psychological conditions during childbirth and postpartum, researchers should pay more attention to the influence of prenatal psycho-social factors on the development of PP-PTSD. Future studies are recommended to track maternal information from the prenatal period to more accurately assess the predictive role of prenatal factors and provide early identification of PP-PTSD.

Limitations

The primary limitation of this review is that the majority of the populations included in the studies originated from mainland China. This may restrict the generalizability of the findings to other countries and necessitate adjustments when applying these models in diverse regions. Additionally, most studies lacked external validation and not enough studies reported detailed calibration of the validation set. The small number of validated models also limited the robustness of the meta-analysis, precluding further exploration of heterogeneity sources and a reliable assessment of publication bias. This highlights the need for future research to prioritize model validation and transparent reporting in accordance with TRIPOD guideline [52]. And the pooled AUC in this study reflected the average discrimination achieved across early postpartum settings with currently available tools, rather than representing a single, universally applicable predictive performance. Moreover, the scope of our review was confined to studies published exclusively in English or Chinese. This language restriction may cause this review to miss valuable insights and data from non-English and non-Chinese speaking regions, which could have contributed to more robust and globally representative analysis.

Conclusion

Of the 16 studies included in this review, fifteen performed either internal validation or external validation. The pooled AUC of 6 validated logistic regression models was 0.86 (95% confidence interval: 0.79–0.90). While some studies reported satisfactory model performance based on metrics such as AUC, the overall assessment of model performance must consider the limitations identified. Upon assessment using the PROBAST checklist, these studies were found to harbor high risk of bias, with 2 study deemed as high concern regarding application. No model can be recommended for direct application in clinical practice of risk prediction for PP-PTSD. In the future, there is a need for rigorously designed studies that develop robust models with more adequate sample size, multi-centre external validation, and transparent reporting according to the TRIPOD guideline. So that the models could be applied to pregnant women in different regions and settings, and provide healthcare professionals with reliable theoretical references to guide clinical practice.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

S.Z.: Writing – original draft, review & editing, Software, Visualization, Methodology, Conceptualization, Data curation. L.K.: Writing – review & editing, Software, Formal analysis, Data curation. S.Y.: Conceptualization, Formal analysis. X.S.: Software. R.L.: Methodology, Investigation. N.M.S.: Writing – review & editing, Supervision, Methodology. All authors reviewed the manuscript.

Funding

This research project was supported by General science and technology program of Quanzhou Medical College (Grant Number: XJK2303B) and Quanzhou guiding science and technology program (Grant Number: 2021N137S).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.