Machine learning-based prediction of long-term prognosis in cervical adenocarcinoma: a retrospective cohort study

BACKGROUND: Cervical adenocarcinoma continues to pose a major global health challenge, highlighting the importance of accurate survival prediction and timely clinical intervention. This study developed machine learning (ML)–based models that integrate key demographic and clinical variables to enable accurate and non-invasive identification of patients at high prognostic risk. METHODS: This retrospective study enrolled 192 patients with cervical adenocarcinoma from the First Affiliated Hospital of Gannan Medical University and included an external testing cohort of 144 patients from Ganzhou People’s Hospital. Ten ML algorithms—logistic regression, multilayer perceptron, extreme gradient boosting, bootstrap aggregating, decision tree, K-nearest neighbors, light gradient boosting machine, naive Bayes, and random forest (RF)—were used to construct prognostic risk prediction models. The primary endpoint was overall survival, with model performance evaluated at the 5-year time point, whereas time to pelvic recurrence and time to distant metastasis were assessed with a focus on a 2-year time horizon. RESULTS: The RF model demonstrated robust predictive performance, achieving F1 scores of 0.897 in the training cohort and 0.846 in the internal testing cohort, with corresponding areas under the receiver operating characteristic curve of 0.836 and 0.880, respectively. External testing in an independent cohort further confirmed the robustness and generalizability of the models, with the RF model achieving the best performance (AUC = 0.779). A higher FIGO stage, larger tumor diameter, and LNM were strongly associated with worse survival outcomes, whereas lower RBC counts and older age also contributed to elevated prognostic risk. Model-derived risk stratification revealed distinct recurrence and metastasis patterns, with the high-risk group showing significantly higher cumulative incidences of pelvic recurrence and distant metastasis compared with the low-risk group (P < 0.001). CONCLUSION: Our ML–based model accurately identified high-risk patients with poor prognostic outcomes, providing a clinically practical tool to support individualized prognostic assessment and treatment planning for cervical adenocarcinoma.