机器学习算法改进中重度创伤性脑损伤住院结局预测:一项中国回顾性队列研究
该单中心回顾性队列纳入2011至2020年住院的482例中重度TBI成人患者(GCS≤12),比较逻辑回归与三种机器学习算法预测住院死亡和GOS功能结局的性能。所有机器学习模型均优于逻辑回归,其中lightGBM表现最佳,预测死亡AUC为0.953(19个特征)和0.935(10个非实验室特征),预测GOS功能结局AUC为0.913和0.889。SHAP方法识别了关键贡献特征,整合不同预测目的的lightGBM模型可提供精细预后信息,尤其对存活患者。
要点
- 在482例中重度TBI患者中,住院死亡率为11.0%,出院时GOS≥4(功能良好)者占23.0%。
- lightGBM预测住院死亡AUC为0.953(19个特征)和0.935(10个非实验室特征),均高于逻辑回归的0.813和0.803。
- lightGBM预测GOS功能结局AUC为0.913(19个特征)和0.889(非实验室数据),均高于逻辑回归的0.832和0.818。
核心结论
- 研究支持机器学习在预测中重度TBI预后方面优于逻辑回归,并强调其临床应用的潜力。
研究设计与样本
- 单中心回顾性队列研究,纳入2011至2020年住院的482例中重度TBI成人患者(GCS≤12)。
- 使用全部19个临床和实验室特征或10个非实验室特征构建逻辑回归和三种机器学习模型(XGboost、lightGBM、FT-transformer),预测住院死亡和GOS功能结局。
- 采用Shapley(SHAP)值进行模型解释。
主要发现
- 总体住院死亡率为11.0%,出院时GOS≥4(功能良好)者占23.0%。
- 所有机器学习模型在预测TBI住院预后方面均优于逻辑回归模型。
- lightGBM预测死亡:使用全部19个特征时AUC为0.953(逻辑回归为0.813),使用10个非实验室特征时AUC为0.935(逻辑回归为0.803)。
- lightGBM预测GOS功能结局:使用全部19个特征时AUC为0.913(逻辑回归为0.832),使用非实验室数据时AUC为0.889(逻辑回归为0.818)。
- SHAP方法识别了lightGBM模型的关键贡献特征,整合不同预测目的的lightGBM模型可提供精细预后信息,尤其对存活的中重度TBI患者。
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
AIM: Controversy remains high over the superiority of advanced machine learning (ML) algorithms to conventional logistic regression (LR) in the prediction of prognosis after traumatic brain injury (TBI). This study aimed to compare the performance of ML and LR models in predicting in-hospital prognosis after TBI. METHOD: In a single-center retrospective cohort of adult patients hospitalized for moderate-to-severe TBI (Glasgow coma score ≤12) in our hospital from 2011 to 2020, LR and three ML algorithms (XGboost, lightGBM, and FT-transformer) were run to build prediction models for in-hospital mortality and the Glasgow Outcome Scale (GOS) functional outcomes using either all 19 clinical and laboratory features or the 10 non-laboratory ones collected at admission to the neurological intensive care unit. The Shapley (SHAP) value was used for model interpretation. RESULT: In total, 482 patients had an in-hospital mortality rate of 11.0%. A total of 23.0% of the patients had good functional scores (GOS ≥ 4) at discharge. All ML models performed better than the LR model in predicting in-hospital prognosis after TBI, among which the lightGBM model showed the best performance: When predicting mortality, the lightGBM model yielded an area under the curve (AUC) of 0.953 using all 19 features (the LR model: 0.813) and an AUC of 0.935 using 10 non-laboratory features (the LR model: 0.803); when predicting GOS functional outcomes, it yielded an AUC of 0.913 using all 19 features (the LR model: 0.832) and an AUC of 0.889 using non-laboratory data (the LR model: 0.818). The SHAP method identified key contributors to explain the lightGBM models. Finally, the integration of the lightGBM models with different prediction purposes was found to provide refined prognostic information, particularly for patients who survived moderate-to-severe TBI. CONCLUSION: The study supported the superiority of ML to LR in predicting prognosis after moderate-to-severe TBI and highlighted its potential use for clinical application.
引用
- Zhang Z, Wang SJ, Chen K, et al. Machine learning algorithms for improved prediction of in-hospital outcomes after moderate-to-severe traumatic brain injury: a Chinese retrospective cohort study. Acta Neurochir (Wien). 2023. PMID: 37382689. doi:10.1007/s00701-023-05647-x
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。