MACHINE LEARNING PREDICTIVE MODELS FOR POSTPARTUM HEMORRHAGE: A SYSTEMATIC REVIEW AND META-ANALYSIS

MACHINE LEARNING PREDICTIVE MODELS FOR POSTPARTUM HEMORRHAGE: A SYSTEMATIC REVIEW AND META-ANALYSIS | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 13 October 2025 V1 Latest version Share on MACHINE LEARNING PREDICTIVE MODELS FOR POSTPARTUM HEMORRHAGE: A SYSTEMATIC REVIEW AND META-ANALYSIS Authors : Carlos R. Mustre-Juarez , Andrea Godinez-Medina 0009-0007-9035-357X , Diana L. Brandt-Perez , Mariana M. Carachure-Rendon , Clara E. Gutierrez-Simpson , Briana M. Rodriguez-Paniagua , Olivia Vazquez-Hernandez 0000-0002-0663-304X , … Show All … , Paul Bain , Sandra Acevedo , Jose A. Ramirez-Calvo , Mario Rodriguez-Bosch , Maria Jose Rodriguez-Sibaja , and Mario I. Lumbreras-Marquez [email protected] Show Fewer Authors Info & Affiliations https://doi.org/10.22541/au.176037002.25876260/v1 395 views 187 downloads Contents Abstract Supplementary Material Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Background Postpartum hemorrhage (PPH) is a leading cause of maternal mortality worldwide. Early risk stratification may enable the implementation of preventive measures and facilitate timely management. Machine learning (ML) models offer potential for predicting PPH by capturing complex risk patterns. Objectives To assess the predictive performance of ML models applied at admission or before delivery for predicting PPH in singleton or twin pregnancies. Search Strategy We systematically searched PubMed, Embase, Web of Science, and Google Scholar without date restrictions. Only English-language studies were considered. Selection Criteria Eligible studies included observational designs or clinical trials that developed or validated supervised ML models to predict PPH. Exclusion criteria were conference abstracts and studies without original data or performance metrics. Data Collection and Analysis Two reviewers independently screened studies and extracted data using instruments based on the CHARMS and TRIPOD+AI frameworks. The risk of bias and applicability were assessed using the PROBAST tool. A random-effects meta-analysis was used to estimate the pooled area under the curve (AUC) with 95% confidence intervals (CIs). Heterogeneity was quantified using the I 2 statistic. Main Results Twenty-four studies met the inclusion criteria. The pooled AUC was 0.83 (95% CI: 0.78–0.88), indicating good discriminatory performance for PPH. However, 13 studies were rated as having a high risk of bias, and 7 raised significant concerns regarding applicability. Heterogeneity was substantial ( I 2 = 99.8%). Conclusions ML models show promise in predicting PPH but are limited by the methods employed, inconsistent outcome definitions, and limited validation. MACHINE LEARNING PREDICTIVE MODELS FOR POSTPARTUM HEMORRHAGE: A SYSTEMATIC REVIEW AND META-ANALYSIS Carlos R. Mustre-Juarez, MD a* , Andrea Godinez-Medina, MD a* , Dana L. Brandt-Perez, MD b , Mariana M. Carachure-Rendon, MD b , Clara E. Gutierrez-Simpson, MD b , Briana M. Rodriguez-Paniagua , MD b , C. Olivia Vazquez-Hernandez, BS b , Paul A. Bain, PhD c , Sandra Acevedo-Gallegos, MD, MS, PhD a , Jose A. Ramirez-Calvo, MD, MS a , Mario R. Rodriguez-Bosch, MD a , Maria J. Rodriguez-Sibaja, MD a , Mario I. Lumbreras-Marquez, MD, MMSc a,b,d** a Maternal-Fetal Medicine Department, Instituto Nacional de Perinatologia, Mexico City, Mexico b Department of Epidemiology and Public Health, Universidad Panamericana School of Medicine, Mexico City, Mexico c Countway Library of Medicine, Harvard Medical School, Boston, United States d Bioinformatics and Biostatistics Department, Instituto Nacional de Perinatologia, Mexico City, Mexico * Both authors contributed equally to this work. **Corresponding Author: Mario I. Lumbreras-Marquez, MD, MMSc Departments of Bioinformatics-Biostatistics and Maternal-Fetal Medicine Instituto Nacional de Perinatologia Mexico City, Mexico E-mail address: [email protected] Machine Learning Predictive Models for Postpartum Hemorrhage: A Systematic Review and Meta-Analysis SHORT TITLE Machine Learning Models for Postpartum Hemorrhage Prediction Background Postpartum hemorrhage (PPH) is a leading cause of maternal mortality worldwide. Early risk stratification may enable the implementation of preventive measures and facilitate timely management. Machine learning (ML) models offer potential for predicting PPH by capturing complex risk patterns. Objectives To assess the predictive performance of ML models applied at admission or before delivery for predicting PPH in singleton or twin pregnancies. Search Strategy We systematically searched PubMed, Embase, Web of Science, and Google Scholar without date restrictions. Only English-language studies were considered. Selection Criteria Eligible studies included observational designs or clinical trials that developed or validated supervised ML models to predict PPH. Exclusion criteria were conference abstracts and studies without original data or performance metrics. Data Collection and Analysis Two reviewers independently screened studies and extracted data using instruments based on the CHARMS and TRIPOD+AI frameworks. The risk of bias and applicability were assessed using the PROBAST tool. A random-effects meta-analysis was used to estimate the pooled area under the curve (AUC) with 95% confidence intervals (CIs). Heterogeneity was quantified using the I ² statistic. Main Results Twenty-four studies met the inclusion criteria. The pooled AUC was 0.83 (95% CI: 0.78–0.88), indicating good discriminatory performance for PPH. However, 13 studies were rated as having a high risk of bias, and 7 raised significant concerns regarding applicability. Heterogeneity was substantial ( I ² = 99.8%). Conclusions ML models show promise in predicting PPH but are limited by the methods employed, inconsistent outcome definitions, and limited validation. FUNDING This research received no specific funding from any public, commercial, or not-for-profit agency. KEYWORDS Postpartum hemorrhage, machine learning, prediction models, systematic review, meta-analysis. INTRODUCTION Postpartum hemorrhage (PPH) remains one of the leading causes of maternal morbidity and mortality worldwide, contributing to nearly a quarter of maternal deaths during pregnancy, childbirth, and the puerperium 1,2,4 . The American College of Obstetricians and Gynecologists (ACOG) defines PPH as blood loss ≥1,000 mL, or any blood loss accompanied by signs or symptoms of hypovolemia within 24 hours of birth 1 . Despite advances in obstetric care, approximately 300,000 women die each year from pregnancy- and childbirth-related complications 2 . The incidence of PPH has remained stable, and in some regions has even increased. Most maternal deaths occur within hours of onset, underscoring the urgent need for early detection and timely intervention. The burden is particularly pronounced in low- and middle-income countries, where limited resources and delayed treatment lead to disproportionately high rates of PPH-related mortality 5 . Several clinical tools have been developed to estimate PPH risk, but their predictive performance is modest and inconsistent across populations and healthcare systems 6 . Reliance on clinical judgment alone is often inadequate, particularly in low-resource settings where unexpected hemorrhagic events may be missed 7 . These shortcomings have prompted growing interest in machine learning (ML) as a means of improving predictive accuracy and supporting clinical decision-making. ML is a branch of artificial intelligence (AI) that encompasses computational methods capable of learning relationships from data, rather than being explicitly programmed to follow predetermined rules 8 . Unlike traditional statistical approaches, which typically assume linearity and require predefined model structures, ML algorithms can capture complex, nonlinear interactions and high-dimensional relationships among variables. They iteratively optimize predictive performance through exposure to data, using techniques such as supervised learning, where models are trained on labeled outcomes, and unsupervised learning, where patterns are inferred without prior outcome labels. In clinical contexts, ML has shown promise in analyzing diverse, large-scale datasets—ranging from electronic health records to imaging and biomarker data—to improve risk prediction and inform decision-making. In obstetrics, ML approaches have increasingly been applied to anticipate complications such as PPH 3 . Supervised algorithms, including decision trees, random forests, gradient boosting, and support vector machines, as well as deep learning methods, have demonstrated potential to identify women at risk of hemorrhage at admission or intrapartum 9–11 . This systematic review and meta-analysis evaluate the performance of supervised ML models in predicting PPH among singleton and twin pregnancies. We examine their discriminative capacity and calibration, assess methodological quality, and provide recommendations for future research to enhance reproducibility, generalizability, and clinical applicability. METHODS This systematic review and meta-analysis was conducted and reported in accordance with the PRISMA 2020 guidelines 13,14 . The review protocol was prospectively registered in PROSPERO (CRD42024571111). Search strategy and selection criteria A comprehensive search was performed in PubMed, Embase, Web of Science, and Google Scholar from inception to October 31, 2024. The strategy was developed in collaboration with experienced medical librarians. Reference lists of included studies were also hand-searched to identify additional eligible reports. Only peer-reviewed articles published in English were considered, with no restrictions on date. The full search strategy is provided in Appendix S1. We included observational studies and clinical trials that developed or validated supervised ML models to predict PPH in singleton or twin pregnancies. Studies were eligible irrespective of the PPH definition used, provided they reported at least a measure of discrimination (e.g., area under the receiver operating characteristic curve [AUC]). Exclusion criteria were: lack of reporting of original findings, absence of performance metrics, focus solely on algorithm development without clinical application, inconclusive results, or publication as a conference abstract only. Study screening was performed independently by two reviewers using Covidence software (Veritas Health Innovation, Melbourne, Australia). Disagreements were resolved by consensus with input from a third reviewer. Data extraction was conducted in duplicate using a standardized electronic form designed for this review, based on the CHARMS and TRIPOD+AI frameworks 14,15 . Extracted variables included study design, country, year of publication, sample size, timing of model application, PPH definition, algorithm type, and reported predictive performance metrics. Risk of bias assessment Two reviewers independently assessed the risk of bias and applicability of each included study using the Prediction model Risk Of Bias Assessment Tool for AI (PROBAST-AI) 14 . The tool evaluates four domains—participants, predictors, outcome, and analysis—with judgments classified as low, high, or unclear risk of bias. Applicability was assessed using the same domains. Discrepancies were resolved by consensus with the assistance of a third reviewer. Results were summarized qualitatively and presented graphically. Data analysis A meta-analysis was performed using reported AUC values to evaluate the discriminatory performance of ML models in predicting PPH. When standard errors for AUC values were not reported, they were derived using a validated approximation formula based on the number of positive and negative cases. Analyses were conducted in R (R Foundation for Statistical Computing) using the ‘meta’ package. A random-effects model with the Restricted Maximum Likelihood (REML) estimator was applied to account for methodological and clinical variability across studies. Ninety-five percent confidence intervals (95% CI) were calculated, and adjustment with the Hartung–Knapp method was used, given the anticipated heterogeneity. Statistical heterogeneity was quantified using the I ² statistic, and between-study variance was estimated using τ ². Forest plots were generated to visualize individual and pooled AUCs, while funnel plots were constructed to assess potential publication bias. The primary analysis included all eligible studies regardless of methodological quality. Sensitivity analyses were then conducted, excluding studies judged to be at high risk of bias according to the CHARMS and PROBAST-AI tools, to evaluate the robustness of the findings. Prespecified subgroup analyses were conducted to investigate potential sources of heterogeneity. These included: (i) type of predictive model (traditional ML models vs. deep learning models such as convolutional or recurrent neural networks), (ii) delivery mode (vaginal vs. caesarean delivery), and (iii) operational definition of PPH (≥500 mL vs. ≥1000 mL blood loss). Within each subgroup, pooled AUCs were estimated using random-effects models, and subgroup differences were formally tested using the chi-square test for interaction (Q statistic and Chi-squared test). All analyses were two-sided, with statistical significance defined as P <0.05. The review adhered to the PRISMA 2020 reporting standards (Supplementary Appendix) 12 . Core outcome sets and patient involvement No core outcome set specific to ML models for predicting PPH was identified in existing registries or literature. As this was a secondary analysis of previously published studies, patient or public involvement was not incorporated. RESULTS Characteristics of included studies The systematic search identified 1,496 records across four databases. After removal of 496 duplicates in EndNote and three in Covidence, 568 unique articles were screened by title and abstract. Of these, 45 full-text articles were assessed for eligibility, and 24 met the inclusion criteria (Figure 1) 3,9,10,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,32,33,34,35,36,37 . Reasons for exclusion are detailed in Figure 1. Manual screening of bibliographies yielded 23 additional references, though none were eligible for inclusion. The included studies were published between 2019 and 2024 and conducted across multiple regions, most frequently the United States, China, and Japan. Twenty were retrospective in design, and sample sizes ranged widely from fewer than 1,000 to more than 1.6 million deliveries (Table 1). Definitions of PPH were heterogeneous: 13 studies defined PPH as blood loss ≥1000 mL, eight as ≥500 mL, and three as ≥2000 mL. Populations varied by mode of delivery, with some studies restricted to caesarean births, others to vaginal deliveries, and several including both. With respect to modeling approaches, 15 studies implemented traditional ML algorithms (e.g., random forests, support vector machines), while 5 used deep learning architectures. Only eight studies reported external validation, and 11 provided at least one calibration measure. The AUC was the most frequently reported metric of predictive performance (Table 1). Predictive performance (AUC) The primary meta-analysis yielded a pooled AUC of 0.83 (95% CI: 0.78–0.88), indicating good overall discriminative performance of ML models for PPH prediction (Figure 2). An AUC greater than 0.80 generally reflects strong separation between affected and unaffected patients, indicating potential clinical utility. However, heterogeneity across studies was significant ( I ² = 99.8%), with between-study variance τ ² = 0.014, suggesting substantial methodological and clinical differences. Individual study results ranged from modest discrimination (AUC ~0.59) to near-perfect performance (AUC 0.98). High-performing studies, such as Westcott 2022 23 (AUC: 0.98 [0.97–0.99]) and Gong 2022 26 (AUC: 0.96 [0.95–0.97]), often used large datasets and ensemble approaches. Lower-performing models, frequently developed in smaller cohorts or with limited predictor sets, reported AUCs ranging from 0.65 to 0.75. Despite promising overall discrimination, only 11 studies reported calibration metrics, and only 8 performed external validation, which limits confidence in the generalizability of the findings. Together, these results highlight both the potential and the current limitations of ML models for PPH prediction. Sensitivity analysis for studies at low risk of bias A sensitivity analysis restricted to five studies judged at low risk of bias by PROBAST-AI (Shazly 2022, Zhang 2024, Zheng 2024, Zheutlin 2021, Zhou 2022) 32,35,34,20,18 yielded a pooled AUC of 0.84 (95% CI: 0.74–0.94) (Supplementary Appendix). While confirming good discrimination, heterogeneity remained high ( I ² = 97.5%). Individual AUCs ranged from 0.71 (Zheutlin 2021) 20 to 0.90 (Zhang 2024) 35 . The relatively lower AUC reported by Zheutlin 2021 20 likely contributed substantially to the heterogeneity. These findings support the internal validity of the pooled results and indicate that well-conducted studies still demonstrate strong predictive capacity. Subgroup analysis: Machine learning vs. deep learning Among traditional ML models (n = 19) 16,37,17,26,30,29,21,27,32,22,9,24,36,23,35,34,20,18,19 , the pooled AUC was 0.83 (95% CI: 0.78–0.89), with AUCs ranging from 0.59 to 0.98. Deep learning models (n = 5) 25,10,33,28,32 achieved a pooled AUC of 0.82 (95% CI: 0.72–0.93), with AUCs between 0.71 and 0.98. No statistically significant difference was observed between model types ( P = 0.897) (Supplementary Appendix). Subgroup analysis: Vaginal birth vs. caesarean section In the vaginal birth subgroup (n= 3) 10,30,37 , the pooled AUC was 0.80 (95% CI: 0.70–0.90; I ² = 91.9%). In the caesarean birth subgroup (n= 8) 25,26,33,28,3,32,19,18 , pooled AUC was 0.89 (95% CI: 0.84–0.95; I ² = 91.5%). The difference between subgroups was not statistically significant ( P = 0.128) (Supplementary Appendix). Subgroup analysis: PPH definition (≥500 mL vs. ≥1000 mL) Studies defining PPH as ≥1000 mL (n = 13) 16,10,25,37,26,27,32,3,9,23,19,18,34 reported a pooled AUC of 0.84 (95% CI: 0.78–0.90; I ² = 99.7%). Those using ≥500 mL (n = 8) 17,33,29,21,28,22,36,25 reported a slightly higher pooled AUC of 0.89 (95% CI: 0.82–0.95; I ² = 98.6%). Several studies in this subgroup, including Krishnamoorthy 2022 33 , Mehrounsh (2023), and Susanu 2024 22 , reported AUCs of ≥0.98. Subgroup differences were not statistically significant ( P = 0.338) (Supplementary Appendix). Quality assessment Risk of bias and applicability were assessed using PROBAST-AI (Figure 3). Most studies were judged to be of low risk for predictors (n = 12, 50%) and outcomes (n = 15, 62.5%). A higher risk was observed for participants (n = 2, 8.3%) and analysis (n = 6, 25.0%), often due to the handling of missing data or the absence of external validation. Regarding applicability, most studies had low concerns (participants: n = 19, 79.1%; predictors: n = 18, 75%; outcomes: n = 19, 79.1%). However, concerns were noted in studies with narrowly selected populations or non-standardized definitions. Overall, 5 studies (20.8%) were rated as having a low risk of bias, 14 (58.3%) were rated as having a high risk, and the remainder were rated as having an unclear risk. Applicability was judged as of low concern in 11 studies (75.0%) and high concern in 7 (29.1%). Assessment of publication bias Visual inspection of the funnel plot suggested asymmetry, with underrepresentation of studies reporting lower AUC values. This pattern indicates possible publication bias, as studies with stronger or more favorable results appeared more likely to be published (Supplementary Appendix). DISCUSSION Main findings This systematic review and meta-analysis is one of the most comprehensive syntheses to date evaluating the predictive performance of supervised ML models for PPH. We identified 24 eligible studies that applied various ML algorithms either at admission or before delivery to predict hemorrhage in singleton and twin pregnancies. The pooled area under the receiver operating characteristic curve (AUC) was 0.83 (95% CI: 0.78–0.88), suggesting that ML-based prediction models demonstrate good overall discriminatory performance. This level of discrimination is clinically meaningful and compares favorably with many currently available obstetric risk stratification tools. Importantly, sensitivity analyses restricted to studies at low risk of bias yielded nearly identical pooled estimates (AUC 0.84), reinforcing the robustness of these findings. Subgroup analyses suggested slightly better performance in studies focused on cesarean births (AUC 0.89) compared with vaginal births (AUC 0.80), though differences were not statistically significant. Similarly, models using deep learning architectures performed comparably to those employing more traditional ML methods, with no clear superiority demonstrated. These findings indicate that supervised ML models have the potential to contribute meaningfully to individualized risk stratification and decision support for PPH. However, their clinical adoption will require careful attention to issues of generalizability, interpretability, and integration within obstetric workflows. Strengths and limitations A comprehensive search was conducted across four databases, complemented by manual screening. Study selection, data extraction, and bias assessment were performed independently by two reviewers using standardized frameworks (CHARMS, TRIPOD+AI) to ensure consistency. The risk of bias was formally evaluated using PROBAST-AI, and a meta-analysis was carried out using random-effects models with a Restricted Maximum Likelihood (REML) estimator and the Hartung–Knapp adjustment, providing conservative estimates that account for extreme heterogeneity. Despite its strengths, several significant limitations should be acknowledged. Heterogeneity across studies was substantial ( I ² = 99.8%), due to differences in study design, data sources, sample sizes, predictors, algorithms, and outcome definitions, underscoring the need for greater methodological standardization. Second, 14 studies were judged to be at high risk of bias, particularly in the domains of analysis and outcome measurement. Common issues included inadequate handling of missing data and inappropriate selection of predictors. Such limitations may compromise internal validity and inflate reported predictive performance. Third, external validation was uncommon, reported in only eight studies. Without rigorous testing in independent datasets, it is difficult to assess whether model performance will generalize beyond the original development setting. Similarly, calibration metrics—essential for evaluating the agreement between predicted and observed probabilities—were reported in less than half of the studies. The absence of these assessments undermines confidence in the clinical applicability of many models. Fourth, the geographic distribution of studies was skewed toward high-income countries. Few studies originate from low- and middle-income countries, where the burden of PPH is most significant and where predictive tools may have the greatest impact. Finally, most included studies were retrospective, relying on secondary data sources that are susceptible to missing information, variable data quality, and inconsistent definitions. While large datasets offer statistical power, they may also embed systematic biases that limit external validity. Interpretation Our findings align with the broader literature on ML in obstetrics, where predictive models have been explored for conditions such as preeclampsia, preterm birth, and stillbirth. Several systematic reviews have shown that ML models often outperform traditional statistical methods in identifying high-risk patients 9,10,3 . For PPH specifically, individual studies included in our analysis frequently reported AUCs exceeding 0.85, particularly when using ensemble models such as random forests, gradient boosting machines, or extreme gradient boosting (XGBoost). These algorithms leverage multiple weak learners to enhance predictive accuracy and often demonstrate resilience in high-dimensional clinical datasets. Nevertheless, comparisons between ML and conventional regression models remain limited. When conducted, results are mixed, with some studies demonstrating marginal gains from ML and others finding little difference once models are carefully specified. This reflects an ongoing debate in the field: ML models are powerful in capturing nonlinear relationships and complex interactions, but they do not consistently outperform well-specified regression approaches, especially in smaller datasets or when predictor selection is limited 38 . Subgroup analyses from this review provide further nuance. Slightly higher performance in cesarean cohorts may reflect more homogeneous patient characteristics and more predictable hemorrhagic etiologies. In contrast, vaginal births encompass a broader spectrum of risk factors and clinical scenarios, which may reduce the predictive precision. Our analysis also found no significant advantage of deep learning over traditional ML approaches. This may be explained by the data structures available: deep learning excels in unstructured inputs such as imaging or free-text clinical notes, whereas most included studies relied on structured tabular datasets (e.g., demographic, clinical, and laboratory variables). Without high-dimensional, unstructured inputs, the full potential of deep learning may not be realized. A significant limitation is the presence of publication bias, as indicated by funnel plot asymmetry, which suggests underreporting of studies with lower predictive performance. This may inflate pooled results and overestimate the effectiveness of ML models, highlighting the importance of transparent reporting of both positive and negative findings. Finally, it is essential to emphasize that high discrimination alone is insufficient for clinical implementation. Calibration, interpretability, and clinical utility must also be taken into account. A model that predicts risk accurately in one dataset but poorly in another, or that provides risk estimates misaligned with observed outcomes, may have limited practical clinical value 40 . Implications for practice and research ML models for PPH prediction are promising but not yet ready for widespread clinical deployment. Their implementation should be preceded by prospective validation in diverse populations, particularly in low- and middle-income countries where the burden of PPH is most significant. Future research should prioritize transparent reporting and methodological rigor. Adherence to TRIPOD+AI and PROBAST-AI guidance will enhance reproducibility, comparability, and critical appraisal. The development of interpretable ML models is urgently needed. Clinicians are unlikely to adopt tools that cannot explain their predictions in a meaningful way. Techniques such as feature importance rankings and SHAP values can enhance transparency and support informed clinical decision-making 41 . Integration into clinical workflows should be a focus of implementation research. Prediction alone does not improve outcomes unless it is linked to actionable interventions. For example, ML-driven alerts for PPH risk could trigger targeted prophylactic measures (e.g., uterotonic preparation, anesthesia consultation, blood product availability) and structured team responses 10 . Collaboration across institutions and countries is needed to develop large, representative, and harmonized datasets for model training and validation. Standardized definitions of PPH could be adopted internationally to minimize heterogeneity and enable valid comparisons. Finally, evaluation should extend beyond accuracy to include health system outcomes. Prospective studies should assess whether ML models reduce time to intervention, optimize resource allocation, and ultimately improve maternal outcomes such as morbidity, transfusion rates, and mortality 23 . CONCLUSION This systematic review and meta-analysis show that supervised ML models demonstrate good discrimination for predicting postpartum hemorrhage. These models could improve risk stratification and prevention in obstetric care; however, their clinical use is limited by methodological heterogeneity, a lack of consistent calibration reporting, few external validations, and a concentration in high-income settings. Future research should prioritize model interpretability, standardization, and validation in real-world, diverse populations. Transparent reporting and integration into clinical workflows are necessary to translate these tools into practice and help reduce the global impact of PPH. ACKNOWLEDGMENTS The authors gratefully acknowledge the Instituto Nacional de Perinatologia for its financial support in covering the publication fees associated with this manuscript. DISCLOSURE OF INTERESTS The authors declare no conflicts of interest related to this work. CONTRIBUTION TO AUTHORSHIP CR Mustre-Juarez and A Godinez-Medina performed the screening, data collection, and drafting of the manuscript. DL Brandt-Perez, MM Carachure-Rendon, CE Gutierrez-Simpson, and BM Rodriguez-Paniagua conducted literature searches, contributed to data interpretation, and assisted with manuscript review and editing. CO Vazquez-Hernandez and PA Bain contributed to the development of the search strategy and provided critical revisions of the manuscript. S Acevedo-Gallegos, JA Ramirez-Calvo, and MR Rodriguez-Bosch contributed to the interpretation of findings and provided substantive intellectual input during the drafting and revision process. MJ Rodriguez-Sibaja contributed to data interpretation and critically revised the manuscript for important intellectual content. MI Lumbreras-Marquez supervised the project, contributed to the analysis and interpretation of the work, revised the manuscript critically for important intellectual content, and approved the final version of the manuscript. All authors reviewed and approved the final version of the manuscript. DETAILS OF ETHICS APPROVAL Ethics approval was not required for this study, as it is a systematic review and meta-analysis of previously published data. REFERENCES 1. Postpartum Hemorrhage ACOG PRACTICE BULLET IN Clinical Management Guidelines for Obstetrician-Gynecologists [Internet]. 2017. Available from: http://www1.health.nsw.gov.au/pds/ActivePDSDocuments/ 2. Lancet T. Reducing postpartum haemorrhage. The Lancet [Internet]. 2025;405:1119. Available from: https://www.cochranelibrary. 3. Shazly SA, Hortu I, Shih JC, Melekoglu R, Fan S, Ahmed F ul A, et al. Prediction of clinical outcomes in women with placenta accreta spectrum using machine learning models: an international multicenter study. Journal of Maternal-Fetal and Neonatal Medicine. 2022;35(25):6644–53. 4. Felis S, Tirloni V, Bertoni M. Postpartum Hemorrhage. Journal of Critical Care & Emergency Medicine [Internet]. 2023 Aug 31;1–16. Available from: https://www.onlinescientificresearch.com/articles/postpartum-hemorrhage.pdf 5. Makwe CC, Okunade KS. Conservative approaches to postpartum haemorrhage. Vol. 95, Best Practice and Research: Clinical Obstetrics and Gynaecology. Bailliere Tindall Ltd; 2024. 6. Rath WH. Postpartum hemorrhage - Update on problems of definitions and diagnosis. Acta Obstet Gynecol Scand. 2011 May;90(5):421–8. 7. Escobar MF, Nassar AH, Theron G, Barnea ER, Nicholson W, Ramasauskaite D, et al. FIGO recommendations on the management of postpartum hemorrhage 2022. International Journal of Gynecology and Obstetrics. 2022 Mar 1;157(S1):3–50. 8. Deo RC. Machine learning in medicine. Circulation. 2015 Nov 17;132(20):1920–30. 9. Venkatesh KK, Strauss RA, Grotegut CA, Heine RP, Chescheir NC, Stringer JSA, et al. Machine Learning and Statistical Models to Predict Postpartum Hemorrhage. Obstetrics and Gynecology. 2020 Apr 1;135(4):935–44. 10. Akazawa M, Hashimoto K, Katsuhiko N, Kaname Y. Machine learning approach for the prediction of postpartum hemorrhage in vaginal birth. Sci Rep. 2021 Dec 1;11(1). 11. Wen T. Interpretable machine learning predicts postpartum hemorrhage at time of admission. Vol. 6, American Journal of Obstetrics and Gynecology MFM. Elsevier Inc.; 2024. 12. Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. PLOS Medicine [Internet]. 2021 Mar;18(3):e1003583. Available from: https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1003583 13. Muka T, Glisic M, Milic J, Verhoog S, Bohlius J, Bramer W, et al. A 24-step guide on how to design, conduct, and successfully publish a systematic review and meta-analysis in medical research. Eur J Epidemiol. 2020 Jan 1;35(1):49–60. 14. Collins GS, Dhiman P, Andaur Navarro CL, Ma J, Hooft L, Reitsma JB, et al. Protocol for development of a reporting guideline (TRIPOD-AI) and risk of bias tool (PROBAST-AI) for diagnostic and prognostic prediction model studies based on artificial intelligence. BMJ Open. 2021 Jul 9;11(7). 15. Moons KGM, de Groot JAH, Bouwmeester W, Vergouwe Y, Mallett S, Altman DG, et al. Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies: The CHARMS Checklist. PLoS Med. 2014;11(10). 16. Ahmadzia HK, Dzienny AC, Bopf M, Phillips JM, Federspiel JJ, Amdur R, et al. Machine Learning Models for Prediction of Maternal Hemorrhage and Transfusion: Model Development Study. JMIR Bioinform Biotech. 2024;5(1). 17. Gallo DM, Romero R, Bosco M, Chaiworapongsa T, Gomez-Lopez N, Arenas-Hernandez M, et al. Maternal plasma cytokines and the subsequent risk of uterine atony and postpartum hemorrhage. J Perinat Med. 2023 Feb 1;51(2):219–32. 18. Zhou Y, Song Z, Wang X, Zhang M, Chen X, Zhang D. Ultrasound-based nomogram for postpartum hemorrhage prediction in pernicious placenta previa. Front Physiol. 2022 Aug 22;13. 19. Xu L, Liu Z, Ma N, Chen J, Shen J, Chen X, et al. Development and validation of an artificial neural network prediction model for postpartum hemorrhage with placenta previa. Minerva Anestesiol. 2023 Nov 1;89(11):977–85. 20. Zheutlin AB, Vieira L, Shewcraft RA, Li S, Wang Z, Schadt E, et al. Improving postpartum hemorrhage risk prediction using longitudinal electronic medical records. Journal of the American Medical Informatics Association. 2022 Feb 1;29(2):296–305. 21. Mehrnoush V, Ranjbar A, Farashah MV, Darsareh F, Shekari M, Jahromi MS. Prediction of postpartum hemorrhage using traditional statistical analysis and a machine learning approach. AJOG Global Reports. 2023 May 1;3(2). 22. Susanu C, Hărăbor A, Vasilache IA, Harabor V, Călin AM. Predicting Intra- and Postpartum Hemorrhage through Artificial Intelligence. Medicina (Lithuania). 2024 Oct 1;60(10). 23. Westcott JM, Hughes F, Liu W, Grivainis M, Hoskins I, Fenyo D. Prediction of Maternal Hemorrhage Using Machine Learning: Retrospective Cohort Study. J Med Internet Res. 2022 Jul 1;24(7). 24. Wang C, Johansson ALV, Nyberg C, Pareek A, Almqvist C, Hernandez-Diaz S, et al. Prediction of pregnancy-related complications in women undergoing assisted reproduction, using machine learning methods. Fertil Steril. 2024 Jul 1;122(1):95–105. 25. Akazawa M, Hashimoto K. A multimodal deep learning model for predicting severe hemorrhage in placenta previa. Sci Rep. 2023 Dec 1;13(1). 26. Gong J, Chen Z, Zhang Y, Liu Y yun, Pu J cai, Xiong C yan, et al. Risk-factor model for postpartum hemorrhage after cesarean delivery: a retrospective study based on 3498 patients. Sci Rep. 2022 Dec 1;12(1). 27. Meyer SR, Carver A, Joo H, Venkatesh KK, Jelovsek JE, Klumpner TT, et al. External Validation of Postpartum Hemorrhage Prediction Models Using Electronic Health Record Data. Am J Perinatol. 2021 Jul 14;41(5):598–605. 28. Shah SY, Saxena S, Rani SP, Nelaturi N, Gill S, Tippett Barr B, et al. Prediction of postpartum hemorrhage (PPH) using machine learning algorithms in a Kenyan population. Front Glob Womens Health. 2023;4. 29. Liu J, Wang C, Yan R, Lu Y, Bai J, Wang H, et al. Machine learning-based prediction of postpartum hemorrhage after vaginal delivery: combining bleeding high risk factors and uterine contraction curve. Arch Gynecol Obstet. 2022 Oct 1;306(4):1015–25. 30. Lengerich BJ, Caruana R, Painter I, Weeks WB, Sitcov K, Souter V. Interpretable machine learning predicts postpartum hemorrhage with severe maternal morbidity in a lower-risk laboring obstetric population. Am J Obstet Gynecol MFM. 2024 Aug 1;6(8). 31. Platt RW, Grandi SM. Machine learning for the prediction of postpartum complications is promising, but needs rigorous evaluation. Vol. 126, BJOG: An International Journal of Obstetrics and Gynaecology. Blackwell Publishing Ltd; 2019. p. 710. 32. Shazly SA, Anan MA, Makukhina TB, Melekoglu R, Ahmed F ul A, Pinto P V., et al. Placenta accreta risk—antepartum score in predicting clinical outcomes of placenta accreta spectrum: A multicenter validation study. International Journal of Gynecology and Obstetrics. 2022 Aug 1;158(2):424–31. 33. Krishnamoorthy S, Liu Y, Liu K. A novel oppositional binary crow search algorithm with optimal machine learning based postpartum hemorrhage prediction model. BMC Pregnancy Childbirth. 2022 Dec 1;22(1). 34. Zheng C, Yue P, Cao K, Wang Y, Zhang C, Zhong J, et al. Predicting intraoperative blood loss during cesarean sections based on multi-modal information: a two-center study. Abdominal Radiology. 2024 Jul 1;49(7):2325–39. 35. Zhang P, Jia Y, Song H, Fan Y, Lv Y, Geng H, et al. Novel biomarkers for prediction of atonic postpartum hemorrhage among ‘low-risk’ women in labor. Front Immunol. 2024;15. 36. Wang W, Liao C, Zhang H, Hu Y. Postpartum Haemorrhage Risk Prediction Model Developed by Machine Learning Algorithms: A Single-Centre Retrospective Analysis of Clinical Data. Clin Exp Obstet Gynecol. 2024 Mar 1;51(3). 37. Betts KS, Kisely S, Alati R. Predicting common maternal postpartum complications: leveraging health administrative data and machine learning. BJOG. 2019;126(6):702–9. 38. Christodoulou, E., Ma, J., Collins, G. S., Steyerberg, E. W., Verbakel, J. Y., & Van Calster, B. (2019). A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. Journal of clinical epidemiology , 110 , 12–22. 39. Obando-Ceron JS, Romero-Cano V, Mayor Toro WM. Exploiting the potential of deep reinforcement learning for classification tasks in high-dimensional and unstructured data. arXiv preprint arXiv:1912.09595; 2019. 40. Singh H, Mhasawade V, Chunara R. Generalizability challenges of mortality risk prediction models: A retrospective analysis on a multi-center database. PLOS Digit Health. 2021;1:e0000023. 41. Shobeiri S. Enhancing transparency in healthcare machine learning models using SHAP and DeepLIFT: a methodological approach. Iraqi J Inf Commun Technol. 2024;7(2):1–12. Figure legends: Figure 1. PRISMA 2020 flow diagram showing the study selection process Figure 2. Forest plot for the pooled odds ratio of postpartum hemorrhage, main analysis Figure 3. Risk of bias assessment of included studies Table legends: Table 1. Characteristics of included studies Supplementary Figures Figure S1. Funnel plot for publication bias Figure S2. Forest plot for sensitivity analysis. Low risk of bias Figure S3. Forest plot for subgroup analysis: machine learning vs deep learning Figure S4. Forest plot for subgroup analysis: vaginal birth vs cesarean delivery Figure S5: Forest plot for PPH subgroup analysis: PPH definition Supplementary Tables Supplementary Table S1. PRISMA 2020 checklist Supplementary Material File (2 table 1- characteristics of included studies.docx) Download 16.65 KB Information & Authors Information Version history V1 Version 1 13 October 2025 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords general obstetrics maternal mortality meta-analysis obstetric haemorrhage puerperium systematic reviews translational research Authors Affiliations Carlos R. Mustre-Juarez Instituto Nacional de Perinatologia View all articles by this author Andrea Godinez-Medina 0009-0007-9035-357X Instituto Nacional de Perinatologia View all articles by this author Diana L. Brandt-Perez Universidad Panamericana View all articles by this author Mariana M. Carachure-Rendon Universidad Panamericana View all articles by this author Clara E. Gutierrez-Simpson Universidad Panamericana View all articles by this author Briana M. Rodriguez-Paniagua Universidad Panamericana View all articles by this author Olivia Vazquez-Hernandez 0000-0002-0663-304X Universidad Panamericana View all articles by this author Paul Bain Harvard Medical School View all articles by this author Sandra Acevedo Instituto Nacional de Perinatologia View all articles by this author Jose A. Ramirez-Calvo Instituto Nacional de Perinatologia View all articles by this author Mario Rodriguez-Bosch Instituto Nacional de Perinatologia View all articles by this author Maria Jose Rodriguez-Sibaja Instituto Nacional de Perinatologia View all articles by this author Mario I. Lumbreras-Marquez [email protected] Instituto Nacional de Perinatologia View all articles by this author Metrics & Citations Metrics Article Usage 395 views 187 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Carlos R. Mustre-Juarez, Andrea Godinez-Medina, Diana L. Brandt-Perez, et al. MACHINE LEARNING PREDICTIVE MODELS FOR POSTPARTUM HEMORRHAGE: A SYSTEMATIC REVIEW AND META-ANALYSIS. Authorea . 13 October 2025. DOI: https://doi.org/10.22541/au.176037002.25876260/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.176037002.25876260/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ff0be037c0a4193',t:'MTc3OTMzNjc0MA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();