Development and validation of an interpretable prediction model for the risk of unplanned reoperation in patients underwent intracranial tumor surgery.

Brain and central nervous system (CNS) malignancies represent a substantial burden on healthcare systems worldwide, and unplanned reoperations following initial surgery are critical events influencing clinical prognosis. Current predictive tools for such reoperations remain limited in their ability to synthesize multifaceted clinical data into accurate risk assessments. This study sought to develop and validate interpretable machine learning algorithms designed to predict the likelihood of unplanned reoperations in patients underwent intracranial tumor surgery. We collected data on patients underwent intracranial tumor surgery who were admitted the First Affiliated Hospital of Xi'an Jiaotong University between January 2023 and January 2024. Patients were additionally partitioned into a training cohort and a validation cohort at a 7:3 proportion. We used least absolute shrinkage and selection operator regression to efficiently screen feature variables associated with CNS cancers postoperative unplanned reoperation. Five machine learning models were employed to predict postoperative unplanned reoperation. The predictive performance of these models was compared by utilizing evaluation metrics, including the area under the receiver operating characteristic curve (AUC). Moreover, the SHapley Additive exPlanation (SHAP) approach was adopted to rank the feature importance and interpret the final model. 11 independent key variables were ultimately chosen to build the model. Among these five machine learning models, the logistic regression (LR) model demonstrated the highest performance. The LR model effectively predicted the risk of unplanned reoperation in patients who underwent intracranial tumor surgery, achieving strong results in both the training set (AUC: 0.836, 95% CI 0.806-0.863) and the internal test set (AUC: 0.769, 95% CI 0.652-0.814). The calibration curve and brier score indicated a close alignment between the predicted and the actual observed risks in the internal test set. Analysis using SHAP identified the duration of surgery, tumor location, modified Frailty Index-5, and tumor type as the most significant predictive factors. To support the practical application of this ML model in a clinical environment, a web-based application was developed for easy access (https://unplanned-reoperation-risk-predicting.streamlit.app/). We developed and internally validated an explainable ML model for predicting the risk of unplanned reoperation in patients underwent intracranial tumor surgery. In this single-center cohort, this model shows promise for assisting healthcare professionals in the early identification of patients at elevated risk, thereby providing a potential basis for exploring personalized treatment strategies tailored to each patient's specific needs.