Abstract
Postpartum hemorrhage (PPH) remains a leading cause of preventable maternal mortality in the US. Our goal was to develop a novel risk assessment tool and compare its accuracy to those used in current practice. We used a PPH digital phenotype we developed and validated previously to identify 6,639 cases from our delivery cohort (N=70,948). Using a vast array of known and potential risk factors extracted from electronic medical records available prior to delivery, we trained a gradient boosting model in a subset of our cohort. In a held-out test sample, we compared performance of our model to three clinical risk tools and one previously published model. Our 24-feature model achieved an area under the curve (AUC) of 0.71 (95% confidence interval [CI], 0.69-0.72), higher than all other tools (research-based AUC: 0.67 [95%CI, 0.66-0.69], clinical AUCs: 0.55 [95%CI, 0.54-0.56] to 0.61 [95%CI, 0.59-0.62]). Five features were novel including red blood cell indices and infection markers measured standardly upon admission. Additionally, we identified inflection points for several vital signs and labs where risk rose substantially. Most notably, patients with median intrapartum systolic blood pressure above 132mmHg had an 11% [interquartile range, 4%] median increase in relative risk for PPH. We developed a novel approach for predicting PPH and identified clinical feature thresholds that can guide intrapartum monitoring for PPH risk. Our results suggest our model is an excellent candidate for prospective evaluation and could ultimately reduce PPH morbidity and mortality through early detection and prevention.
Introduction
Postpartum hemorrhage (PPH)—excessive bleeding following childbirth—is a leading cause of preventable maternal mortality and morbidity. Despite its severity, identifying which patients are at high risk remains a significant clinical challenge. Current risk assessment toolkits, such as those from the California Maternal Quality Care Collaborative (CMQCC) or the Association of Women’s Health, Obstetric and Neonatal Nurses (AWHONN), rely on static, expert-curated checklists of known risk factors.
However, these tools are often binary and imprecise. Research indicates that standard clinical tools can classify over 40% of PPH cases as "low-risk" simply because the patients do not fit a traditional profile. There is a critical need for more nuanced, data-driven tools that leverage the longitudinal data stored in modern electronic medical records (EMR) to identify risk before emergencies occur.
Methodology
The researchers developed an Integrated Machine Learning (IML) model using a gradient boosting framework (LightGBM). The study utilized a massive dataset from the Mount Sinai Health System (MSHS) in New York City, encompassing 70,948 deliveries from 57,151 mothers between 2011 and 2019. The approach was distinguished by several key innovations:
Digital Phenotyping
Rather than relying solely on estimated blood loss, which is notoriously inaccurate, the team used a validated "digital phenotype." This algorithm identifies PPH cases by combining blood loss data with sharp drops in hematocrit, the administration of "rescue" medications (like misoprostol), and surgical interventions.
Feature Engineering and Longitudinal Analysis
The model extracted thousands of potential predictors from EMR data, covering the period from eight months prior to pregnancy up to the moment of delivery. To handle time-series data, such as fluctuating blood pressure during labor, the researchers used Functional Principal Component Analysis (fPCA) to summarize longitudinal trends into usable model inputs.
Explainability
To ensure the model was not a "black box," the team used Shapley values (SHAP) to rank feature importance. This allowed them to identify specific physiological "inflection points" where a patient's risk begins to climb significantly.
Key Results and Insights
The 24-feature IML model outperformed all existing clinical and research benchmarks:
- Predictive Power: The IML model achieved an AUC of 0.71, significantly higher than widely used clinical toolkits (AUCs of 0.55 to 0.61) and a previously published research model (AUC 0.67).
- Precision and Alarm Fatigue: In the "high-risk" category, the IML model achieved a Positive Predictive Value (PPV) of 28%, nearly double that of some clinical toolkits (15-19%). This suggests that a high-risk flag from this model is much more likely to represent a true case, potentially reducing alarm fatigue for clinicians.
Physiological Inflection Points
A striking finding was the identification of risk thresholds within the "normal" clinical range. For example, the model found that a median intrapartum systolic blood pressure above 132 mmHg was associated with an 11% increase in relative risk for PPH. In standard obstetric practice, hypertension is typically defined as ≥140 mmHg, suggesting that subtle shifts within the conventional normal range are early warning signs.
Novel Risk Markers
The model highlighted several features not currently used in standard risk toolkits, including red blood cell indices (such as RDW) and infection markers (white blood cell and absolute neutrophil counts) measured upon hospital admission.
Discussion and Limitations
The researchers demonstrated that the definition of PPH matters; models may appear more accurate when tested against simple blood-loss metrics, but these results are often misleading. Using a robust, multi-factor digital phenotype provides a more honest and clinically useful assessment of performance.
However, the study has limitations. It was a retrospective analysis of historical data from a single health system. The authors emphasize that prospective, real-time evaluation in diverse hospital settings is the necessary next step. Furthermore, the current model does not utilize unstructured clinical notes, which may contain rich context that could further improve prediction.
Conclusion
This work moves PPH risk assessment from a static checklist to a dynamic, high-resolution monitoring strategy. By identifying novel markers and physiological inflection points, this model provides a data-driven path to earlier intervention, potentially saving lives through more accurate and timely risk detection.
Summary updated