Development and validation of machine learning-based predictive models for the risk of medication-related osteonecrosis of the jaw and surgical outcomes
- VernacularTitle:基于机器学习的药物相关性颌骨坏死发病风险和手术疗效的预测模型构建及验证
- Author:
Mengmeng WU
1
;
Changfei MAO
2
;
Jing ZHANG
1
;
Yuan ZHANG
1
;
Xiaolin LIU
3
;
Yingqiu PU
1
Author Information
1. Dept. of Pharmacy，Nanjing Stomatological Hospital，Affiliated Hospital of Medical School，Institute of Stomatology，Nanjing University，Nanjing 210008，China
2. Dept. of Breast Surgery，Jiangsu Cancer Hospital/The Affiliated Cancer Hospital of Nanjing Medical University/Jiangsu Institution of Cancer Research，Nanjing 210009，China
3. Dept. of Pharmacy，Jiangsu Cancer Hospital/The Affiliated Cancer Hospital of Nanjing Medical University/ Jiangsu Institution of Cancer Research，Nanjing 210009，China
- Publication Type:Journal Article
- Keywords:
medication-related osteonecrosis of the jaw;
risk of onset;
surgical outcomes;
machine learning;
drug risk management
- From:
China Pharmacy
2026;37(16):2187-2194
- CountryChina
- Language:Chinese
-
Abstract:
OBJECTIVE To develop and validate interpretable machine learning models for predicting the risk of medication-related osteonecrosis of the jaw （MRONJ） and surgical outcomes， thereby providing quantitative support for MRONJ risk assessment and treatment decision-making.METHODS Two study cohorts were established. Cohort A adopted a nested case-control study design to develop a predictive model for the risk of MRONJ onset. Cohort B comprised patients with MRONJ from Cohort A who underwent surgical treatment and met the follow-up eligibility criteria， and employed a prospective follow-up cohort study design to develop a predictive model for the efficacy of MRONJ surgery. After missing-data processing， variable standardization， and encoding， predictive features were selected using the least absolute shrinkage and selection operator （LASSO） regression， and collinearity was assessed using the variance inflation factor （VIF）. Random Forest （RF）， eXtreme Gradient Boosting （XGB）， Multi-layer Perceptron （MLP）， Support Vector Machine （SVM）， and Gaussian Naive Bayes （NB） models were subsequently developed using the selected features. Hyperparameter optimization and internal performance evaluation were performed using two-level nested cross-validation repeated 20 times. In Cohort B， the synthetic minority over-sampling technique （SMOTE） was applied within the inner training folds to address class imbalance. Model performance was comprehensively evaluated using the receiver operating characteristic （ROC） curve， the precision-recall （PR） curve， calibration curve， and decision curve analysis （DCA）. Shapley additive explanations （SHAP） were used to quantify the relative contributions of individual clinical features to model predictions.RESULTS Seven clinical features， including alkaline phosphatase and tooth extraction， were retained for Cohort A； whereas eight clinical features， including age and duration of medication use， were retained for Cohort B. In Cohort A， the area under the ROC curve， the area under the PR curve， Brier score， sensitivity， and specificity of the RF model were 0.94， 0.92， 0.09， 0.92， and 0.85， respectively； the corresponding values in Cohort B were 0.93， 0.93， 0.12， 0.83， and 0.93， respectively. The results of calibration curves and DCA indicated that the RF models achieved favorable calibration and clinical net benefit， with VIF values for all variables below 5. SHAP visualization results showed that alkaline phosphatase， tooth extraction， and cumulative drug dose made substantial contributions to MRONJ risk prediction in Cohort A， whereas age， duration of medication use， and MRONJ stage contributed substantially to surgical outcome prediction in Cohort B.CONCLUSIONS The RF-based models for predicting MRONJ risk and surgical outcomes demonstrated good discrimination， accuracy of probabilistic predictions， and potential clinical utility during internal validation. These models may provide decision support for comprehensive MRONJ risk management and individualized clinical decision-making. However， external validation in independent， multicenter， and geographically diverse cohorts is still required.
