Does a random forest model integrating clinical and MRI features improve diagnostic performance for placenta accreta spectrum in pregnant women with non-previa placenta?
Integrating clinical features such as hormone replacement cycle-frozen embryo transfer and MRI findings into a random forest model improves the diagnostic accuracy for placenta accreta spectrum in non-previa placenta.
OBJECTIVES To identify clinical and MRI features useful for diagnosing placenta accreta spectrum (PAS) in non-previa placenta and to develop diagnostic models integrating these features. METHODS This retrospective study included 101 pregnant women with non-previa placenta who underwent MRI between January 2022 and June 2024. Nineteen were confirmed as PAS. Clinical variables and 11 MRI findings were evaluated using intraoperative or pathological results as the reference standard. Diagnostic performance was assessed using univariable analysis and repeated cross-validation of a random forest (RF) model, with ROC analysis used to assess discriminative performance. RESULTS Hormone replacement cycle-frozen embryo transfer (HRC-FET) (sensitivity 0.89, specificity 0.63) and abnormal placental bed vascularization (sensitivity 0.63, specificity 0.90) showed the strongest univariable performance. The RF model using six variables with acceptable interobserver agreement achieved an AUC of 0.88, sensitivity 0.92, specificity 0.79, demonstrating higher discriminative performance than individual predictors. Feature importance analysis highlighted HRC-FET and abnormal placental bed vascularization as the most influential factors. CONCLUSIONS Integrating clinical and MRI features improves PAS diagnosis in non-previa placenta. The RF model demonstrated a more balanced diagnostic profile than individual predictors in this exploratory cohort and may aid preoperative risk assessment. HRC-FET and abnormal placental bed vascularization were key contributors, supporting their relevance for risk stratification.
Ueno et al. (Fri,) studied this question.