We read with great interest the comprehensive study by Wang et al (2025) that established an integrated prognostic classification system for cesarean scar pregnancy (CSP) by combining machine learning (ML) and traditional linear scoring models, with ultrasound and clinical features as core predictors.1 CSP, a life-threatening ectopic pregnancy subtype, has seen a rising incidence alongside the increasing cesarean section rate globally, and its individualized prognostic stratification and precise treatment decision-making have long been clinical challenges.2 Wang et al's work innovatively addresses this gap by constructing a dual-model system that balances predictive performance and clinical interpretability, and the findings provide valuable practical guidance for CSP management. Herein, we discuss the strengths of this study and offer several perspectives for further exploration and validation. A major advantage of this study is its rigorous design and multidimensional feature mining: the study included 26 ultrasound and 8 clinical features, systematically screened prognostic variables through univariate analysis and SHapley Additive exPlanations-based feature importance ranking, and identified Implantation Range of Gestational Sac Gestational Sac Uterine Cavity Protrusion Level, and Residual Myometrial Thickness as the most critical predictors of poor CSP prognosis.1 This finding is highly consistent with recent clinical evidence that the implantation range of gestational sac, degree of uterine cavity protrusion, and residual uterine muscle layer thickness are key factors determining the risk of uterine rupture and severe bleeding in CSP patients.3 It is worth noting that this study innovatively developed a simplified linear scoring model for poor prognosis (Group C) with an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.939, which is comparable to the Gradient Boosting Decision Tree (GBDT) ML model (p = .814).1 This linear model with only 3 quantitative indicators has high inter operator consistency and is easy to promote in primary healthcare institutions with limited ML technology, which is an important practical contribution of this work. Another highlight is the study's stratified modeling strategy for different prognostic subgroups: for the easily distinguishable poor prognosis group (Group C), a concise linear model was adopted, while for the favorable prognosis group (Group A) with overlapping feature distributions with the moderate group (Group B), a GBDT model with 13 variables was constructed, achieving an AUC of 0.917.1 This targeted modeling fully leverages the advantages of linear models in interpretability and ML models in capturing complex non-linear feature correlations, which is a significant improvement over previous CSP prediction models that used a single algorithm.4 In addition, the study used Principal Component Analysis for patient distribution visualization and Synthetic Minority Over-sampling Technique (SMOTE) for sample imbalance correction, and multiple robustness verifications ensured the reliability of the model results, reflecting the high methodological rigor of the research. Although significant progress has been made in this study, there are several aspects that require further in-depth investigation and external validation. Firstly, this study is a single center retrospective study with a relatively small sample size in the poor prognosis group (n = 31).1 Although SMOTE is used to mitigate the impact of sample imbalance, the generalizability of this model needs to be validated through multicenter, large sample prospective studies, particularly in populations with different races, regions, and clinical practice characteristics. Secondly, the study did not include dynamic clinical indicators such as the decrease rate of serum β-hCG after initial treatment and changes in ultrasound characteristics during follow-up.1 Recent studies have shown that the dynamic changes in β-hCG within 48 hours after treatment are an independent predictor of successful CSP treatment, and combining static baseline features with dynamic follow-up indicators can further improve prognostic accuracy.5 Thirdly, the current model only divides CSP into 3 prognostic groups based on treatment methods and short-term outcomes (such as intraoperative blood loss and surgical transition), without including long-term prognostic indicators such as uterine scar healing quality, secondary infertility, and recurrent CSP risk. As more and more patients of childbearing age suffer from CSP, long-term reproductive prognosis is an important clinical issue, and incorporating it into the prognostic system will make the model more comprehensive.6 In conclusion, Wang et al's study is a pioneering work in CSP prognostic stratification that successfully integrates ML and traditional linear models to achieve a balance between predictive accuracy and clinical applicability.1 The identified core ultrasound predictors and simplified scoring model provide a feasible tool for clinical rapid risk assessment of CSP, and the stratified modeling strategy offers a valuable methodological reference for the prognosis research of other gynecological acute diseases. We believe that with multi-center external validation, the incorporation of dynamic and long-term prognostic indicators, and the optimization of model features, this prognostic classification system will be further improved and widely applied in clinical practice, ultimately contributing to the precise and individualized management of CSP patients and the reduction of severe clinical complications. Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.
Liu et al. (Wed,) studied this question.