Development and validation of a machine learning model based on interpretable clinical characteristics for preoperative prediction of Ki-67 expression in pituitary adenomas.

BACKGROUND

Elevated levels of Ki-67 are commonly linked to more aggressive tumors and worse clinical prognosis. Thus, this study was performed with the aim of developing and validating a machine learning (ML) model to preoperatively identify the expression of Ki-67 within PAs.

METHODS

Clinical information was collected from individuals with PAs treated at Weifang People's Hospital between 2019 and 2023. The dataset was split randomly into training and validation sets, utilizing a 3:1 allocation. In the training set, variables markedly correlated with high Ki-67 expression were selected through the LASSO regression analysis. Based on these variables, Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), and XGBoost ML models were then constructed. The area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, and specificity were examined to explore the predictive accuracy of the constructed models. The decision curve analysis (DCA) was conducted to assess the clinical applicability.

RESULTS

Altogether 227 patients with PAs were included. LR analysis identified age, changes in visual acuity, prothrombin time, tumor location within the sella turcica, and functionality as significant predictors of high expression of Ki-67 in PAs. The ML models built with these variables, involving SVM and LR, demonstrated favorable predictive performance and exhibited high clinical applicability.

CONCLUSIONS

Future studies should focus on identifying predictors with stronger associations with Ki-67 expression and developing preoperative ML models to facilitate early and accurate assessment of Ki-67 expression.