(862) Improving placenta accreta spectrum prediction: Integrating conventional statistics and machine learning

To evaluate the predictive performance of logistic regression and a machine learning (ML) algorithm in identifying placenta accreta spectrum (PAS) using ultrasonographic signs assessed before delivery. Early and accurate PAS prediction is critical for surgical planning and improving maternal outcomes. However, access to expert interpretation of PAS-related ultrasound findings is often limited, particularly in low-resource settings.

Study Design:

Retrospective study conducted at a quaternary referral center. We included patients with suspected PAS who underwent prenatal ultrasound assessment and had definitive intraoperative or histopathological confirmation. Ultrasonographic signs and clinical variables were extracted and labeled (yes/no) using the ISUOG pro forma by maternal-fetal medicine specialists. Grayscale ultrasound parameters included loss of the retroplacental hypoechoic zone, myometrial thinning, placental lacunae, bladder wall interruption, placental bulge, and focal exophytic mass. Color Doppler parameters included bridging vessels, uterovesical hypervascularity, subplacental hypervascularity, and turbulent lacunar flow. A multivariable logistic regression and a random forest model were developed. Model performance was evaluated for discrimination (using the area under the curve [AUC]) and calibration.

Results:

We included 233 patients, of whom 157 (67.4%) had confirmed PAS. The logistic regression model achieved an AUC of 0.89, while the ML model achieved an AUC of 0.88, indicating similar discrimination. Calibration showed good agreement between predicted and observed risks for both models (Hosmer-Lemeshow P=0.523; Brier Score=0.121 for random forest). Key predictors included loss of the “clear zone”, myometrial thinning, and bridging vessels.

Conclusion:

Both models demonstrated strong performance in predicting PAS. ML offered no improvement in discrimination. With further validation, such models could help standardize PAS risk assessment and guide referral or surgical planning, especially in settings with limited access to high-risk obstetric ultrasound expertise.