(209) Preconception Prediction of Placenta Accreta Spectrum Using Machine Learning and EMR Data

To develop a machine learning (ML) model using only preconception electronic medical record (EMR) data to identify patients at risk for Placenta Accreta Spectrum (PAS) and evaluate its predictive performance.


Study Design:

We conducted a retrospective case-control study using EMR data from 118,890 deliveries (2013–2023) within a single academic health system. PAS was confirmed in 274 cases (0.23%) via manual review of delivery and pathology records. Pre-pregnancy data included demographics, obstetric and surgical history, vitals, labs, billing codes, and provider notes. Natural language processing (NLP) extracted structured features from free-text notes. ML models tested included logistic regression, decision trees, random forest, XGBoost, and multilayer perceptron. Model performance was evaluated using AUC, sensitivity, specificity, and SHAP (Shapley Additive Explanations) for interpretability.


Results:

XGBoost achieved the highest area under the receiver operating curve (AUROC) (0.86) and outperformed logistic regression, AUROC 0.76 [Figure 1]. Random forest demonstrated the highest sensitivity (91%), while logistic regression had the highest specificity (91%) [Figure 1]. SHAP analysis revealed that traditional risk factors—including advanced maternal age, prior cesarean delivery, dilation and curettage (D&C), gynecologic surgery, and history of obstetric complications were among the top predictors. Pre-pregnancy anemia emerged as a previously unrecognized risk factor [Figure 2].


Conclusion:

Preconception ML models using EMR data can identify patients at elevated risk for PAS with high sensitivity and specificity. This early screening approach may facilitate risk counseling, referral, and personalized care planning before conception—especially in patients without access to expert imaging. The identification of anemia as a novel predictor highlights the potential role of modifiable risk factors and warrants further investigation.