|
| 机器学习方法对绝经后女性患骨质疏松的风险预测 |
| Risk prediction of osteoporosis in postmenopausal women by machine learning methods |
| |
| DOI：10.3969/j.issn.1006-7108.2026.03.011 |
| 中文关键词: 骨质疏松 绝经后女性 预测模型 |
| 英文关键词:osteoporosis postmenopausal women prediction model |
| 基金项目: |
|
| 摘要点击次数: 27 |
| 全文下载次数: 0 |
| 中文摘要: |
| 目的 基于不同机器学习（ML）算法构建绝经后女性患骨质疏松的风险预测模型。方法 收集2019年1月至2024年12月期间来自南京鼓楼医院老年科就诊的495名绝经后女性患者的临床资料，比较骨质疏松组和非骨质疏松组的基线资料差异，再根据7∶3随机分成训练组和测试组。在训练集中通过LASSO回归筛选预测因子，使用Bayes-net、XGBoost、Adaboost、RF、SVM算法建立ML模型，使用ROC曲线、AUC、准确度、灵敏度、特异性、阳性预测值等指标进行评价。超参数调优模型。在测试集中比较各机器学习模型效能。采用SHAP分析各特征因子对预测结果的相对贡献。结果 495例患者中，178例为骨质疏松患者（患病率为35.95 %）。骨质疏松组和非骨质疏松组的age、BMI、ALB、TBIL、DBIL、BUN、TC、Scr、TG、Apo-B、P、eGFR、FSH、PRGE、DHEAS、SHBG、是否高血压、是否冠心病等临床资料差异存在统计学意义（P＜0.05）。训练组（347例）和测试组（148例）的各项临床资料差异均不存在统计学意义（P＞0.05）。Lasso回归筛选得到预测因子：age、BMI、ALB、TBIL、AKP、SHBG。在训练组中RF模型AUC（0.871，95 %CI：0.831~0.910），模型超参数调优后，在测试组中得到优化RF模型的AUC（0.824，95 %CI：0.759~0.890）。结论 特征因子按重要性排序得到age、BMI、albumin、TBIL、SHBG、AKP。基于机器学习方法建立的模型能有效预测绝经后女性患骨质疏松风险。 |
| 英文摘要: |
| Objective To construct risk prediction models for osteoporosis in postmenopausal women based on different machine learning (ML) algorithms. Methods Clinical data were collected from 495 postmenopausal patients attending the Geriatrics Department of Nanjing Drum Tower Hospital from January 2019 to December 2024. The differences in baseline data between the osteoporosis and non-osteoporosis groups were compared. Patients were randomly divided into a training group and a testing group based on 7:3. In the training set, predictors were screened with LASSO regression. The ML models were built using Bayes-net, XGBoost, Adaboost, Random forest, and SVM algorithms. Calibration was performed by using ROC curve, AUC, accuracy, sensitivity, and specificity. The positive predictive value were used for evaluation. The models were adjusted with hyperparameters. The effectiveness between each machine learning model were compared. Results Of the 495 patients, 178 were osteoporotic (prevalence 35.95%). The age, BMI, ALB, TBIL, DBIL, BUN, TC, Scr, TG, Apo-B, P, eGFR, FSH, PRGE, DHEAS, SHBG, and rate of hypertension and coronary artery disease were statistically different between the two groups (P<0.05). There was no statistical difference (P>0.05) between the training group (347 cases) and testing group (148 cases) in all clinical information. The factors predicting osteoporosis were obtained with Lasso regression screening: including age, BMI, ALB, TBIL, AKP, and SHBG. AUC of RF model was 0.871 (95% CI: 0.831-0.910) in the training group. After model hyperparameter tuning, the AUC of optimized RF model was obtained in the test group (0.824, 95% CI: 0.759-0.890). Conclusion Characterization factors obtained in the order of importance are AGE, BMI, albumin, TBIL, SHBG, and AKP. The established model based on machine learning approach is effective in predicting the risk of osteoporosis in postmenopausal women. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|