Individualized survival prediction and risk stratification using machine learning for patients with malignant struma ovarii: a population-based study with external validation.

Background: Malignant struma ovarii (MSO) is a rare thyroid-type cancer originating in ovarian teratoma. Prognosis of MSO is less studied without unanimous staging or stratification system. This study aimed to developed and validated a machine learning (ML)-based model to predict overall survival (OS) for patients with MSO and to risk-stratify them.

Methods: Patients with histologically confirmed MSO diagnosed in 1975-2021 from the Surveillance, Epidemiology, and End Results (SEER) program were identified as the training cohort. Patients in a systematic literature review were collected as the testing cohort. OS was selected as the outcome, while demographic, clinicopathological and therapeutic information were used as features. Following data encoding, imputing and scaling, univariate feature selection was performed. Cox proportional hazard (CoxPH), Cox with elastic net penalty (CoxNet), random survival forest (RSF), gradient boosting machine (GBM), and survival tree (ST) models were trained and tuned. Each model was evaluated on its c-index, time-dependent area under the curve (AUC), time-dependent Brier score (BS) and stratification ability in the training and the testing cohort respectively. The algorithm that performed the best in the testing cohort was finally chosen for SHapley Additive exPlanations (SHAP) interpretation and Streamlit web application deployment.

Results: The study included 120 and 194 patients in the training and testing cohort respectively. At the end of follow-up (median time 115.5 and 32.5 months respectively), 101 (84.2%) and 181 patients (93.3%) survived respectively. RSF had the best performance in the testing cohort, possessing the highest c-index (0.841, 95% confidence interval: 0.732-0.916), the highest mean AUC (0.852), the lowest integrated BS (0.042), and the smallest P value (<0.001) on log-rank test comparing the stratified groups. According to SHAP, older age, hysterectomy, larger tumor size and more advanced American Joint Committee on Cancer stage had the strongest predictive power for worse OS among all 13 features. An interactive application (https://mso-surv.streamlit.app/) was then implemented which can display the predicted Kaplan-Meier curve, survival probability, risk stratification and the contributions of features for the output.

Conclusions: We reported the first externally tested time-to-event prognostic prediction model for MSO. ML algorithms enabled precise individual-patient prediction and stratification, and can potentially assist patient counselling and decision-making for treatment and surveillance.