Machine Learning Models to Predict Survival Outcomes According to the Surgical Approach of Primary Radical Hysterectomy in Patients with Early Cervical Cancer

Abstract

Simple SummaryAn increased risk of relapse and death from minimally invasive radical hysterectomy has been reported in some patients with early cervical cancer. Thus, the development of an intuitive and precise decision-aid tool, which estimates recurrence and mortality rates by surgical approach, is necessary. To develop models predicting survival outcomes according to the surgical approach, we collected clinicopathologic and survival data of patients with 2009 FIGO stage IB cervical cancer who underwent a radical hysterectomy. Using only variables that could be obtained preoperatively, we developed various models predicting the probability of 5-year progression-free survival and overall survival. Among them, hybrid ensemble models, combined with logistic regression and multiple machine learning models, achieved the best predictive performance. The developed models are expected to help physicians’ and patients’ decision making related to the surgical approach for primary radical hysterectomy.We purposed to develop machine learning models predicting survival outcomes according to the surgical approach for radical hysterectomy (RH) in early cervical cancer. In total, 1056 patients with 2009 FIGO stage IB cervical cancer who underwent primary type C RH by either open or laparoscopic surgery were included in this multicenter retrospective study. The whole dataset consisting of patients’ clinicopathologic data was split into training and test sets with a 4:1 ratio. Using the training set, we developed models predicting the probability of 5-year progression-free survival (PFS) and overall survival (OS) with tenfold cross validation. The developed models were validated in the test set. In terms of predictive performance, we measured the area under the receiver operating characteristic curve (AUC) values. The logistic regression models comprised of preoperative variables yielded AUCs of 0.679 and 0.715 for predicting 5-year PFS and OS rates, respectively. Combining both logistic regression and multiple machine learning models, we constructed hybrid ensemble models, and these models showed much improved predictive performance, with 0.741 and 0.759 AUCs for predicting 5-year PFS and OS rates, respectively. We successfully developed models predicting disease recurrence and mortality after primary RH in patients with early cervical cancer. As the predicted value is calculated based on the preoperative factors, such as the surgical approach, these ensemble models would be useful for making decisions when choosing between open or laparoscopic RH.