Development and internal validation of a prediction model for postpartum haemorrhage following caesarean delivery in a resource-limited setting in Ethiopia

Introduction Postpartum haemorrhage remains the leading cause of maternal mortality globally and continues to disproportionately affect women in low- and middle-income countries including Ethiopia. Women undergoing caesarean delivery are at increased risk of haemorrhagic complications, yet context-specific prediction tools to guide early risk stratification are limited. Objective To develop and internally validate a parsimonious risk prediction model for postpartum haemorrhage (PPH) in Northeast Ethiopia. Design A multicentre cross-sectional study was conducted and data were collected using KoboCollect and analysed in R V.4.5.1. The model was developed using logistic regression, with variable selection performed via Least Absolute Shrinkage and Selection Operator (LASSO) to ensure a robust set of predictors. Performance of the model was evaluated through discrimination and calibration plots. Internal validation was conducted using 1000 bootstrap resamples. Clinical utility was assessed using decision curve analysis, and findings were translated into a simplified risk score and nomogram. Setting The study was conducted in three public hospitals, Northeast Ethiopia. Participants A total of 1556 study participants were included in the study Results From the total women who underwent cesarean delivery, 23.9% women developed postpartum haemorrhage. The resulting parsimonious model demonstrated excellent discriminative ability with an AUC of 0.823 (95% CI 0.788 to 0.858). The model demonstrated a sensitivity of 95.16%, specificity of 45.6%, positive predictive value of 84.72% and negative predictive value of 74.7%. Calibration plots showed strong agreement between predicted and observed PPH probabilities. The over-optimism coefficient was 0.012, indicating a low risk of overfitting. Decision curve analysis confirmed that the model provides a higher net benefit compared with ‘intervention-for-all’ or ‘no-intervention’ strategies across a wide range of threshold probabilities. Conclusion This risk prediction model demonstrates robust predictive performance and holds significant potential for guiding clinical decision-making, informing preventive strategies and optimising resource allocation, thereby advancing precision in risk management. A key limitation of this study is the absence of external validation in geographically diverse populations, which restricts the generalisability of the findings.