基于机器学习的颅脑创伤后垂体功能减退症预测建模
本研究分析620例颅脑创伤(TBI)病例,按70%训练集和30%测试集拆分并采用5折交叉验证,评估10种机器学习模型对创伤后垂体功能减退症(PTHP)的预测能力。逻辑回归在训练集和测试集的AUC分别为0.905和0.887,表现最优,中线移位≥5 mm是最强预测因子。
要点
- 逻辑回归预测PTHP的AUC在训练集为0.905、测试集为0.887,敏感性与特异性均衡。
- 与PTHP显著相关的变量包括高血压、ICU入住、GCS≤8、弥漫性脑水肿等(P<0.05)。
- 中线移位≥5 mm是最强预测因子,模型经Hosmer-Lemeshow检验校准良好。
核心结论
- 逻辑回归凭借较高的AUC、敏感性、特异性和校准度,成为TBI后PTHP预测的较优模型。
- 本研究可能有助于TBI患者PTHP的早期诊断和治疗。
研究设计与样本
- 研究分析620例病例,采用逻辑自变量事件数方法。
- 数据按70%训练集和30%测试集拆分,并应用5折交叉验证。
- 评估10种机器学习模型:逻辑回归、决策树、随机森林、LightGBM、XGBoost、CatBoost、朴素贝叶斯、支持向量机、k近邻和多层感知机。
- 性能指标包括准确率、敏感性、特异性、阳性预测值、阴性预测值和受试者工作特征曲线下面积。
主要发现
- 与PTHP显著相关的变量(P<0.05)包括高血压、ICU入住、格拉斯哥昏迷评分≤8、弥漫性脑水肿、脑疝、中线移位≥5 mm、颅内压升高、开颅术、颅底骨折和住院时长。
- 逻辑回归在训练集和测试集的AUC分别为0.905和0.887,敏感性与特异性均衡。
- 朴素贝叶斯以及LightGBM和CatBoost等集成方法显示出有竞争力的AUC。
- 中线移位≥5 mm是最强预测因子。
- 模型经Hosmer-Lemeshow检验显示校准良好。
临床怎么用
- 原文未据此给出推荐。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: Traumatic brain injury (TBI) remains a leading cause of disability and death globally, and post-traumatic hypopituitarism (PTHP) is a common complication for which there is a lack of accurate predictive models. This study aimed to develop a machine learning model to assess the risk of hypopituitarism after TBI. METHODS: A sample of 620 cases was analyzed using the logistic independent variable event count method. Data was split into a training set (70%) and a test set (30%), with 5-fold cross-validation applied. Ten machine learning models were evaluated: Logistic Regression, Decision Tree, Random Forest, LightGBM, XGBoost, CatBoost, Naive Bayes, Support Vector Machine, k-nearest neighbors, and Multi-Layer Perceptron. Performance metrics included accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and area under the receiver operating characteristic curve. The following variables were significantly associated with PTHP (P < 0.05): hypertension, intensive care unit admission, Glasgow Coma Scale ≤8, diffuse cerebral edema, cerebral herniation, midline shift ≥5 mm, elevated intracranial pressure, craniotomy, skull base fracture, and length of stay. These variables may contribute to a predictive model for PTHP risk. RESULTS: Logistic regression excelled in training and test sets, boasting an area under the curve (AUC) of 0.905 and 0.887, respectively, with balanced sensitivity and specificity. Naive Bayes and ensemble methods like LightGBM and CatBoost showed competitive AUCs. Midline shift ≥5 mm was the strongest predictor. The model demonstrated excellent calibration per the Hosmer-Lemeshow test. CONCLUSIONS: Logistic regression, with its strong AUC, sensitivity, specificity, and calibration, emerged as the superior model for PTHP prediction post-TBI, This study could facilitate earlier diagnosis and treatment of PTHP in TBI patients.
引用
- Chen A, Zhong H, Peng J, et al. Machine Learning-Based Modeling for Predicting Hypopituitarism After Cranial Trauma. World Neurosurg. 2025. PMID: 40368087. doi:10.1016/j.wneu.2025.124018
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。