|
| 基于深度学习算法构建胸腔镜手术患者全身麻醉诱导后低血压风险的预测模型 |
| Development of a deep learning-based prediction model for post-induction hypotension risk in patients undergoing general anesthesia for video-assisted thoracic surgery |
| |
| DOI：10.12089/jca.2026.07.001 |
| 中文关键词: 诱导后低血压 胸腔镜手术 深度学习 机器学习 长短期记忆网络 沙普利加性解释 |
| 英文关键词: Post-induction hypotension Video-assisted thoracic surgery Deep learning Machine learning Long short-term memory Shapley additive explanations |
| 基金项目:山东省中医药科技项目（M-2023061） |
|
| 摘要点击次数: 112 |
| 全文下载次数: 34 |
| 中文摘要: |
目的：基于深度学习算法构建胸腔镜手术（VATS）全身麻醉诱导后低血压（PIH）风险的预测模型，并与多种经典机器学习模型进行比较。
方法：回顾性收集2023年3月至2025年3月接受VATS治疗的510例患者作为训练集，另选取2025年4月至2026年3月符合标准的122例VATS患者作为测试集，年龄≥18岁，ASA Ⅰ—Ⅲ级。采用Boruta算法筛选PIH风险变量。基于长短期记忆网络（LSTM）构建深度学习预测模型，并与朴素贝叶斯（NB）、随机森林（RF）、类别提升（CatBoost）、柔性判别分析（FDA）及逻辑回归（LR）5种经典机器学习模型进行比较。通过五折交叉验证评估模型性能，评估指标包括Brier分数、ROC曲线下面积（AUC）、敏感性及特异性，并采用沙普利加性解释（SHAP）值对LSTM模型进行可解释性分析。
结果：训练集中有284例患者（55.6%）发生PIH。Boruta算法筛选出MAP、术前禁食时间、年龄、BMI及使用血管紧张素转换酶抑制剂/血管紧张素Ⅱ受体拮抗剂（ACEIs/ARBs）5个关键PIH风险变量。在6种模型中，LSTM模型在多项指标上表现最佳（Brier分数0.098，AUC 0.997，敏感性0.982，特异性0.974），预测性能优于传统机器学习模型。SHAP全局解释显示，基础MAP、年龄、禁食时间、BMI及使用ACEIs/ARBs是影响PIH风险的主要特征变量。测试集中有66例（54.1%）发生PIH，LSTM模型预测的AUC为0.921（95%CI 0.883～0.976），校准曲线显示总体预测准确率为87.4%。
结论：基于LSTM算法的深度学习模型在预测VATS患者PIH风险方面具有较高准确性与稳定性。结合SHAP值可实现关键风险特征识别，为术前风险评估及个体化麻醉管理提供参考。 |
| 英文摘要: |
Objective: To develop a deep learning-based prediction model for post-induction hypotension (PIH) risk during general anesthesia for video-assisted thoracic surgery (VATS), and to compare its performance with multiple classical machine learning models.
Methods: Clinical data from 510 patients who underwent VATS at our hospital between March 2023 and March 2025 were retrospectively collected as the training set. Additionally, 122 eligible VATS patients from April 2025 to March 2026 were selected as an external validation set to assess the predictive accuracy of the LSTM model. The Boruta algorithm was used to screen PIH risk variables. A deep learning prediction model was constructed using long short-term memory (LSTM) and compared with five classic machine learning models: Naive Bayes (NB), random forest (RF), categorical boosting (CatBoost), flexible discriminant analysis (FDA), and logistic regression (LR). Model performance was evaluated using five-fold cross-validation, with metrics including Brier score, AUC, sensitivity, and specificity. The Shapley additive explanations (SHAP) method was used to interpret and visualize the deep learning model.
Results: A total of 284 patients (55.6%) developed PIH in the training set. The Boruta algorithm identified five key risk variables for PIH: baseline MAP, preoperative fasting time, age, BMI, and use of angiotensin-converting enzyme inhibitors/angiotensin Ⅱ receptor blockers (ACEIs/ARBs). Among the six models, the LSTM model performed best across multiple metrics (Brier score = 0.098, AUC = 0.997, sensitivity = 0.982, specificity = 0.974), demonstrating superior predictive performance compared to conventional machine learning models. SHAP global interpretation revealed that baseline MAP, age, fasting time, BMI, and ACEIs/ARBs use were the main characteristic variables influencing PIH risk. In the test set (n = 122), the incidence of PIH was 54.1% (66/122). The LSTM model achieved an AUC of 0.921 (95% CI 0.883-0.976) for prediction, and the calibration curve showed an overall prediction accuracy of 87.4%.
Conclusion: The deep learning model based on the LSTM algorithm provides high accuracy and stability in predicting PIH risk among VATS patients. Combined with SHAP interpretation, it enables identification of key risk factors and offers a novel reference for preoperative risk assessment and individualized anesthesia management. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
