|
| 基于机器学习建立门诊无痛胃镜诊疗患者低氧血症预测模型 |
| Prediction model for hypoxemia based on machine learning in outpatients undergoing painless gastroscopy diagnosis and treatment |
| |
| DOI：10.12089/jca.2025.12.003 |
| 中文关键词: 机器学习 门诊患者 无痛胃镜 低氧血症 预测模型 |
| 英文关键词: Machine learning Outpatients Painless gastroscopy Hypoxemia Prediction model |
| 基金项目:贵州省高层次创新人才“千层次”人才项目（黔人领发〔2020〕4号） |
|
| 摘要点击次数: 1253 |
| 全文下载次数: 445 |
| 中文摘要: |
目的：基于机器学习(ML)建立门诊无痛胃镜诊疗患者发生低氧血症的风险预测模型。
方法：回顾性收集2023年3—8月陆军军医大学第一附属医院和遵义医科大学第三附属医院行门诊无痛胃镜患者823例，男418例，女405例，年龄60～85岁，ASA Ⅰ—Ⅲ级，收集患者人口学信息、既往病史以及临床相关资料。根据无痛胃镜诊疗过程中是否发生低氧血症(SpO2<90%)将患者分为两组：低氧血症组和非低氧血症组。建立逻辑回归（LR）、支持向量机（SVM）、随机森林（RF）、极端梯度提升（XGB）以及轻量梯度提升机（LightGBM）共5种ML模型，采用受试者工作特征曲线下面积（AUROC）、精确率-召回率曲线下面积（AUPRC）、准确性、敏感性、特异性、F1分数、Brier分数评估模型性能，采用夏普利加法解释（SHAP）对最佳模型进行解释分析。
结果：有110例（13.4%）发生低氧血症。在测试集中LR的AUROC为0.873(95%CI 0.861～0.877），SVM的AUROC为0.845（95%CI 0.804～0.855），RF的AUROC为0.881（95%CI 0.840～0.887），XGB的AUROC为0.880（95%CI 0.857～0.899），LightGBM的AUROC为0.891（95%CI 0.853～0.900）。进一步综合AUPRC、准确性、敏感性、特异性、F1分数、Brier分数的结果显示，基于LightGBM算法构建的预测模型效能最佳。在SHAP特征重要性解释方面，依托咪酯+丙泊酚(EP合剂)诱导用量、基础SpO2、胃镜诊疗时间、追加EP合剂、术中呛咳、反复咽部刺激和小下颌对模型具有较高的贡献度。
结论：基于LR、SVM、RF、XGB和LightGBM算法构建5种门诊无痛胃镜诊疗患者低氧血症预测模型，其中LightGBM的预测效能最佳。 |
| 英文摘要: |
Objective: To establish a risk prediction model for hypoxemia based on machine learning (ML) in outpatients undergoing painless gastroscopy diagnosis and treatment.
Methods: A retrospective cohort of 823 outpatients undergoing painless gastroscopy from March to August 2023 was established, 418 males and 405 females, aged 60-85 years, and ASA physical status Ⅰ-Ⅲ. Patient demographic information, past medical history, and relevant clinical data were collected. The patients were divided into two groups based on whether hypoxemia occurred (SpO2< 90%) during painless gastroscopy diagnosis and treatment: hypoxemia group and non-hypoxemia group. Five machine learning (ML) models were evaluated, including logistic regression (LR), support vector machine (SVM), random forest (RF), extreme gradient boosting (XGB), and light gradient boosting machine (LightGBM). Model performance was comprehensively assessed using the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), accuracy, sensitivity, specificity, F1-score, and Brier score. The optimal model was constructed and subsequently interpreted using Shapley additive explanations (SHAP).
Results: Hypoxemia occurred in 110 patients (13.4%). The AUROC values for the five ML models in the test set were as the following: LR (AUROC = 0.873, 95% CI 0.861-0.877), SVM (AUROC = 0.845, 95% CI 0.804-0.855), RF (AUROC = 0.881, 95% CI 0.840-0.887), XGB (AUROC = 0.880, 95% CI 0.857-0.899), and LightGBM (AUROC = 0.891, 95% CI 0.853-0.900). Further integration of the results of AUPRC, accuracy, sensitivity, specificity, F1 score, and Brier score showed that the prediction model constructed based on the LightGBM algorithm has the best performance. SHAP analysis identified the following features as having the highest contribution to the optimal model: induction dose of the combination of propofol and etomidate (EP compound agent), baseline SpO2 upon entering the procedure room, gastroscopy procedure duration, administration of additional EP compound agent, occurrence of cough, repeated pharyngeal stimulation, and micrognathia.
Conclusion: This study developed five hypoxemia prediction models in outpatients undergoing painless gastroscopy diagnosis and treatment based on LR, SVM, RF, XGB, and LightGBM algorithms. Among these, the LightGBM model demonstrated superior performance. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
