Related Experiment Videos
Development and Validation of an Interpretable Machine Learning Model for Prediction of the Need for Surgical
Yusheng Li1, Wenjun Chen2, Mingxiao Wen3
1School of Information Engineering, Hangzhou Medical College, Hangzhou, Zhejiang, People's Republic of China.
Background:
International guidelines lack standardized quantitative criteria for surgical evacuation in incomplete abortion, leading to heterogeneous clinical practice, which may result in unnecessary interventions and increased clinical and economic burdens on women and healthcare resources.
Methods:
This retrospective study included clinical data from 411 patients with medical abortion at Hangzhou Women's Hospital (2023-2025). A stacking ensemble model with 5-fold cross-validation was constructed, and its performance was evaluated using AUC, with 95% CI, accuracy, sensitivity, and specificity, followed by internal hold-out validation. Feature importance was determined using a weighted average of Information Gain (IG) and SHapley Additive exPlanations (SHAP) values. SHAP analyses were further used to quantify the effects of key features on surgical intervention risk and identify critical risk thresholds.
Results:
The stacking model exhibited the optimal performance, with an internal validation AUC of 0.821 (95% CI: 0.795-0.846), accuracy 0.794, sensitivity 0.757 and specificity 0.806, and a hold‑out validation AUC of 0.801 (95% CI: 0.700-0.892), outperforming traditional and single ensemble models. Blood β-human chorionic gonadotropin (β-hCG), body mass index (BMI), uterine residual tissue blood flow resistance index (RI), and residual tissue size at first post-medical abortion review (RTS‑FPAR) were identified as core predictive factors. SHAP analysis confirmed that β-hCG ≥ 824.90 mIU/mL, BMI ≤ 18.91 or ≥ 23.42, RI ≤ 0.51, RTS‑FPAR ≥ 3.30 cm significantly elevated surgical intervention risk.
Conclusion:
The stacking model showed promising moderate performance in predicting the need for surgical evacuation in our study. Our findings suggest that existing criteria for surgical intervention may cover an overly broad range of cases, and more selective thresholds may support improved risk stratification to potentially reduce unnecessary interventions. These findings offer preliminary research insights for auxiliary risk assessment of incomplete abortion, and multi-centre external validation is required before clinical use.