Objective To develop preoperative machine learning models predicting survival outcomes according to the surgical approach in early-stage cervical cancer.

Methods We retrospectively identified patients with FIGO stage IB cervical cancer who underwent either primary open RH or laparoscopic RH at three tertiary institutional hospitals between 2000 and 2018. Patients’ clinicopathologic, image, and survival data were collected. The whole dataset was separated into training and test sets with a 4:1 ratio. Combining both statistical and deep neural network models, we constructed hybrid ensemble predictive models for 5-year PFS and OS rates. Only the variables that could be obtained before surgery were used. Model development was conducted in the training set with ten-fold cross-validation, and the developed models were validated in the test set.

Results In total, 1,141 patients were included; 578 and 563 received open RH and laparoscopic RH, respectively. The median length of observation was 57.6 months during which 157 patients (13.8%) experienced disease recurrence and 86 patients (7.5%) died. In terms of preoperative prediction, while the logistic regression model showed AUCs of 0.68 and 0.71 for 5-year PFS and OS rates, respectively, the ensemble model showed better performance: AUCs, 0.71 and 0.78. These models commonly included the surgical approach as the main prognostic factor.

Conclusion We developed preoperative models predicting survival outcomes according to the surgical approach in early-stage cervical cancer. These models will be useful for making decisions in choosing open RH or laparoscopic RH in individualized counseling practices.