文章摘要
麻醉恢复室内多种并发症的可解释多标签分类预测模型
Interpretable multi-label classification prediction model for multiple complications in the post-anesthesia care unit
  
DOI：10.12089/jca.2025.08.001
中文关键词: 麻醉恢复室  多标签分类  可解释性  风险因素  联合预测
英文关键词: Post-anesthesia care unit  Multi-label classification  Interpretability  Risk factors  Joint prediction
基金项目:甘肃省自然科学基金面上项目（23JRRA1296）；甘肃省临床医学研究中心建设项目（21JR7RA675）；甘肃省人民医院人才库项目（2024KYQDJ-C-27）
作者单位E-mail
马国婷 730000,兰州市,甘肃省人民医院麻醉科  
贾晓琴 730000,兰州市,甘肃省人民医院麻醉科  
张东 730000,兰州市,甘肃省人民医院麻醉科  
王玲凯 730000,兰州市,甘肃省人民医院麻醉科 152127433@qq.com 
阎文军 730000,兰州市,甘肃省人民医院麻醉科  
摘要点击次数: 739
全文下载次数: 183
中文摘要:
      
目的：探讨麻醉恢复室内多种并发症的发生情况及其相关风险因素，构建并验证可解释的多标签分类预测模型。
方法：选择2023年11月至2024年10月择期全身麻醉下手术后转入麻醉恢复室的患者10 313例，男5 416例，女4 897例，年龄≥18岁，按7∶3比例随机划分为训练集(n=7 220)和验证集(n=3 093)。持续监测和评估患者是否发生呼吸系统不良事件、低体温、血流动力学不稳定、恶心/呕吐、躁动/谵妄、疼痛及非疼痛症状7类麻醉恢复室常见并发症。采用多标签分类学习技术在训练集中构建二进制相关性、分类器链、集成分类器链和多标签k近邻4种预测模型，并在验证集中通过汉明损失、准确度、精确度、召回率和F1分数评估模型性能。通过Shapley加法解释对表现最佳的模型进行全局和局部的可解释性分析。
结果：有4 854例（47.1%）患者至少发生一种并发症，2 723例（26.4%）患者同时出现两种或以上并发症。年龄≥65岁、女性、BMI≥28 kg/m2、麻醉时间≥4 h及术后镇痛是预测麻醉恢复室并发症的5个关键特征。分类器链模型表现最佳，汉明损失为0.119、准确度为0.881、精确度为0.888、召回率为0.987以及F1分数为0.934。
结论：结合可解释性的多标签分类模型能够有效预测麻醉恢复室内并发症，并揭示各并发症的特定风险因素，为术后早期并发症的综合管理提供可预测、可操作的方法。
英文摘要:
      
Objective: To investigate the incidence of multiple complications in the post-anesthesia care unit (PACU) and their associated risk factors, and to develop and validate an interpretable multi-label classification prediction model.
Methods: The convenience sampling method was used to enroll 10 313 patients who underwent elective surgery under general anesthesia and were transferred to the PACU between November 2023 and October 2024, 5 416 males and 4 897 females, aged ≥ 18 years. Patients were randomly assigned to training group (n = 7 220) and validation group (n = 3 093) in a 7∶3 ratio. Continuously monitored and assessed seven common complications: respiratory adverse events, hypothermia, hemodynamic instability, nausea / vomiting, agitation / delirium, pain, and non-pain symptoms. Four typical predictive models including binary relevance, classifier chains, ensemble of classifier chains, and multi-label k-nearest neighbors were constructed in the training set using multi-label classification learning techniques, and their performance were evaluated in the validation set using Hamming loss, accuracy, precision, recall, and F1 score. The best-performing model was analyzed for global and local interpretability using Shapley additive explanations.
Results: A total of 4 854 patients (47.1%) experienced at least one complication, and 2 723 patients (26.4%) had two or more complications simultaneously. Key features influencing PACU complication prediction included age ≥ 65 years, female gender, BMI ≥ 28 kg/m2, anesthesia duration ≥ 4 hours, and postoperative pain management. The classifier chain model performed best, with a Hamming loss of 0.119, accuracy of 0.881, precision of 0.888, recall of 0.987, and F1 score of 0.934.
Conclusion: The multi-label classification model, combined with interpretability methods, can effectively predict complications in the PACU and identify specific risk factors for each complication. This approach provides a predictive and actionable tool for the comprehensive management of early postoperative complications.
查看全文   查看/发表评论  下载PDF阅读器
关闭