目的 基于机器学习算法构建神经外科重症患者早期肠内营养误吸风险的预测模型,并验证其预测性能。 方法 回 顾性收集 2023 年 1 月至 2024 年 12 月张家港澳洋医院收治神经外科重症患者 322 例作为研究对象,采用留出法,以 4 ∶ 1 的比 例将患者随机分为建模组(n = 258)和验证组(n = 64)。 收集患者临床资料,通过 LASSO 回归分析筛选关键变量,进行多因素 分析,基于相关危险因素,采用机器学习算法构建早期肠内营养误吸风险预测模型,验证模型性能。 结果 本研究纳入 322 例 患者,87 例发生误吸,发生率为 27. 02%。 LASSO 回归分析共筛选出 9 个关键变量,多因素分析结果显示,年龄、轻度以上意识 障碍、机械通气、吸烟史、营养风险筛查 2002 评分(≥3 分)、≥1 个合并症为神经外科重症患者早期肠内营养发生误吸的独立 危险因素(P<0. 05)。 据此构建 6 种机器学习模型,受试者操作特征曲线分析结果显示,6 种机器学习模型建模组和验证组的 曲线下面积均>0. 7,其中以 XGBoost 模型的预测性能最高。 决策曲线分析结果显示,6 种模型在高风险阈值 0~ 1. 0 范围内,相 比均干预和未干预,均可获得更高的标准化净收益;十折交叉验证结果显示,6 种机器学习模型的曲线下面积均未见明显波 动,提示模型拟合度良好。 结论 基于机器学习算法构建神经外科重症患者早期肠内营养误吸风险的预测模型相对精准地预 测误吸风险,其中以 XGBoost 的预测性能最佳,但考虑临床实用性与便利性,推荐基于性能仅次于 XGBoost 的 Logistic 算法构 建预测模型,以便临床制订相应预防措施,降低误吸风险。

Objective Constructing a predictive model for early enteral nutrition aspiration risk in critically ill neurosurgical patients based on machine learning algorithms and verifying its predictive performance. Method A retrospective study was conducted on 322 critically ill neurosurgical patients admitted to Zhangjiagang Aoyang Hospital from January 2023 to December 2024. The patients were randomly divided into a modeling group n = 258 and a validation group n = 64 in a 4 ∶ 1 ratio using the retention method. Collect clinical data from patients screen key variables through LASSO regression analysis conduct multiple factor analysis based on relevant risk factors use machine learning algorithms to construct an early enteral nutrition aspiration risk prediction model and verify the performance of the model. Result This study included 322 patients of whom 87 experienced aspiration with an incidence rate of 27. 02%. LASSO regression analysis identified a total of 9 key variables. Multivariate analysis results showed that age consciousness status mechanical ventilation smoking history NRS 2002 score and number of comorbidities were independent risk factors for early enteral nutrition aspiration in critically ill neurosurgical patients P < 0. 05 . Based on this 6 machine learning models were constructed. Receiver working characteristic curve analysis results showed that the area under the curve of the modeling and validation groups of the 6 machine learning models was>0. 7 with the XGBoost model having the highest predictive performance The decision curve analysis results show that within the high -risk threshold range of 0-1. 0 all six models can achieve higher standardized net returns compared to those with and without intervention The ten fold cross validation results showed that there was no significant fluctuation in the area under the curve of the six machine learning models indicating good model fitting. Conclusion A prediction model for early enteral nutrition aspiration risk in critically ill neurosurgical patients based on machine learning algorithms is relatively accurate in predicting aspiration risk. Among them XGBoost has the best predictive performance but considering clinical practicality and convenience it is recommended to build a prediction model based on Logistic algorithm which is second only to XGBoost in performance in order to formulate corresponding preventive measures and reduce aspiration risk in clinical practice.