Artificial intelligence for postpartum hemorrhage: a systematic review.

Background: Postpartum hemorrhage (PPH) remains a leading cause of maternal morbidity and mortality, and conventional risk tools often rely on static factors that may miss rapidly evolving intrapartum events. Artificial intelligence (AI) offers data-driven, potentially dynamic approaches for PPH prediction, yet prior syntheses have provided limited coverage of recent methods and validation practices. Objective: To systematically synthesize the evidence on AI models for PPH prediction, including clinical applications, data sources, prediction targets, validation strategies, and study quality. Methods: We conducted a PRISMA-guided systematic review of studies published between 2015 and 2025 across MEDLINE, Embase, Scopus, IEEE Xplore, and Google Scholar. Two reviewers independently screened records, extracted data using a prespecified form, and resolved disagreements by consensus. Risk of bias and applicability were appraised using a modified QUADAS-2 tool tailored to AI-based PPH prediction. Findings were synthesized narratively across study design, prediction context, inputs, modeling approaches, and validation strategies. Results: In total, 33 studies met the eligibility criteria. Publications were concentrated in 2023-2025 (79%) and were predominantly retrospective (91%) and single-site (55%). Most models targeted anticipatory risk stratification (79%), with fewer addressing intrapartum/immediate postpartum early warning or severity escalation (24%). Outcomes primarily modeled PPH occurrence (79%) as binary classification (91%). All studies used closed datasets. Classical machine learning dominated (85%), while deep learning (33%) and LLM-based approaches (6%) were less frequent. Validation was mainly internal (76%), with limited external validation. Participant selection was the main quality concern (high risk of bias: 48%), while index test and reference standard were largely low risk. Conclusions: AI-based PPH prediction research is rapidly expanding but remains constrained by retrospective, closed datasets, heterogeneous outcome definitions, and limited external/temporal validation. Progress toward clinical readiness requires harmonized labeling, multicenter datasets, deployment-oriented evaluation, and prospective implementation studies.