Accuracy of machine learning and traditional statistical models in the prediction of postpartum haemorrhage: a systematic review.

OBJECTIVES

To evaluate whether postpartum haemorrhage (PPH) can be predicted using both machine learning (ML) and traditional statistical models.

DESIGN

Diagnostic systematic review and meta-analysis of observational and clinical studies, prospectively registered on PROSPERO, performed accordingly to the Preferred Reporting Items for Systematic Reviews and Meta-analysis and Prediction model risk of bias assessment tool for studies developing, validating or updating prediction models, with the use of an independent analysis by a large language model (GPT-4 Open AI).

DATA SOURCES: MEDLINE/PubMed, LILACS-BVS, Cochrane Library, Scopus-Elsevier, Embase-Elsevier and Web of Science.

ELIGIBILITY CRITERIA FOR SELECTED STUDIES: The literature search was conducted on 4 January 2024 and included observational studies and clinical trials published in the past 10 years that assessed early PPH and PPH prediction and that applied accuracy metrics for outcomes evaluation. We excluded studies that did not define PPH or had exclusive PPH subgroups evaluation.

PRIMARY AND SECONDARY OUTCOME MEASURES: The primary outcome is the accuracy of PPH prediction using both ML and conventional statistical models. A secondary outcome is to describe the strongest risk factors of PPH identified by ML and traditional statistical models.

RESULTS

Of 551 citations screened, 35 studies were eligible for inclusion. The synthesis gathered 383 648 patients in 24 studies conducted with conventional statistics (CS), 9 studies using ML models and 2 studies using both methods. Multivariate regression was a preferred modelling approach to predict PPH in CS studies, while ML approaches used multiple models and a myriad of features. ML comparison to CS was only performed in two studies, and ML models demonstrated a 95% higher likelihood of PPH prediction compared with CS when applied to the same dataset (OR 1.95, 95% CI 1.88 to 2.01, p<0.001). The I² had a value of 54%, p=0.14, indicating moderate heterogeneity between the studies.

CONCLUSIONS

ML models are promising for predicting PPH. Nevertheless, they often require a large number of predictors, which may limit their applicability or necessitate automation through digital systems. This poses challenges in resource-scarce settings where the majority of PPH complications occur.

PROSPERO REGISTRATION NUMBER: CRD42024521059.