Machine Learning Models for Predicting Liver Metastasis at Diagnosis and Overall Survival in Ovarian Cancer: A SEER-Based Study.

Purpose: Ovarian cancer (OC) is one of the most common gynecological tumors, and liver metastases are the most common metastatic sites of OC. However, there is a lack of machine learning (ML) models that can predict the diagnosis and prognosis of liver metastases in OC patients. Therefore, this study aims to develop an effective predictive model. Methods: This was a retrospective study based on the Surveillance, Epidemiology, and End Results (SEER) database. Patients with OC diagnosed from 2010 to 2020 were extracted from the SEER database. The dataset was partitioned into a training cohort (60%) and a validation cohort (40%). The primary endpoints of this study were the liver metastases at diagnosis in OC patients and the 12-, 36-, and 60-month overall survival (OS) of OC liver metastases patients. After feature selection via Boruta algorithm, 9 ML diagnostic models and 5 prognostic models were constructed. In the diagnostic model, area under the curve (AUC), accuracy, kappa, sensitivity, specificity, positive predictive value, negative predictive value evaluated model reliability. AUC and Brier score were used for prognostic models. Results: Of 27,065 OC patients, 1053 had liver metastases at diagnosis. It was observed that histological type, T stage, grade, age, N stage, CA125, laterality, and race were associated with liver metastasis at diagnosis in OC patients. Histological type, chemotherapy, surgery, radiotherapy, lung metastasis, bone metastasis, age, tumor grade, and marital status were associated with OS in patients with OC liver metastasis. In the training cohort, KNN had the highest AUC value (0.863). In the validation cohort, Ridge had the highest AUC (0.758). Among the 9 diagnostic models, Ridge regression stands out the most, with AUC (0.758), sensitivity (0.929), and negative predictive value (0.994) were the highest. For 12-month OS, RSF model had highest AUC (0.876) in both training and validation groups. Among 5 prognostic models, RSF showed best comprehensive performance. The stage T was the most discriminative features for diagnosing liver metastasis. For OS at 12, 36, and 60 months, the most discriminative prognostic features were chemotherapy, histological type, and age, respectively. Surgery and chemotherapy were associated with improved OS. Conclusion: The Ridge regression and RSF had favorable predictive performance in the diagnostic and prognostic models, respectively, compared with the other tested models; this may further help clinicians identify patients with liver metastasis at the time of OC diagnosis and select appropriate treatment options.