Development and validation of a machine learning model based on laboratory parameters for preoperative prediction of Ki-67 expression in gliomas

OBJECTIVE Glioma is the most common form of brain tumor and has high mortality. The Ki-67 proliferation index, a vital marker of cell proliferation, has been demonstrated to predict tumor classification and prognosis. The aim of this study was to develop and validate a noninvasive model based on machine learning (ML) and routine laboratory parameters to preoperatively predict the level of Ki-67 in gliomas. METHODS A total of 506 patients with pathological confirmation of glioma from 2 medical centers (January 2020 to December 2023) were retrospectively enrolled and divided into training (n = 352), internal validation (n = 88), and external validation (n = 66) cohorts. According to the Ki-67 proliferation index, patients were classified into low Ki-67 (index < 10%) and high Ki-67 (index ≥ 10%) groups. Laboratory parameters were obtained within 1 week before surgery from the Laboratory Information System. The potential features associated with Ki-67 levels were screened using extreme gradient boosting (XGBoost), support vector machine (SVM), and least absolute shrinkage and selection operator (LASSO). Then, 10 ML classifiers, including SVM, XGBoost, logistic regression (LR), random forest, adaptive boosting (AdaBoost), gradient boosting machine, partitioning around medoids, naive Bayes, neural network, and bagged classification and regression trees (CART), were trained. The performance of these models was evaluated on internal and external validation sets using the area under the receiver operating characteristic curve (AUC). Calibration curve, decision curve, and clinical impact curve analyses were used for validation. RESULTS Fifteen laboratory parameters that met the requirements of XGBoost, SVM, and LASSO were selected. Among all tested ML models, the LR model had superior performance with relatively high AUC, accuracy, sensitivity, and specificity. The LR model achieved AUCs of 0.838 in the training set, 0.800 (with the highest accuracy [0.782] and optimal sensitivity [0.845]) in the internal validation set, and 0.757 in the external validation set. Finally, the LR model was visualized as a nomogram based on the top 6 laboratory parameters (age, anion gap, apolipoprotein A-1, apolipoprotein B, calcium, creatinine) to individually predict the Ki-67 proliferation index in patients with gliomas. CONCLUSIONS The authors successfully constructed an LR model based on routine laboratory parameters, with relatively high sensitivity and specificity, to preoperatively predict the level of Ki-67 in patients with gliomas, which might be helpful for prognostic evaluation and clinical decision-making.