Analysis of risk prediction model for severe disturbance of consciousness in patients with acute ischemic stroke: a machine learning method based on intrinsic interpretability
急性缺血性脑卒中患者严重意识障碍风险预测模型分析：基于内在可解释性机器学习方法

摘要

目的 本研究基于MIMIC-IV数据库，旨在开发一种可解释的机器学习模型，用于预测卒中患者发生严重意识障碍的风险。方法 本研究从MIMIC-IV数据库中，根据ICD-9和ICD-10编码提取急性缺血性脑卒中患者。利用LASSO回归算法进行特征筛选。通过七种机器学习算法，依据AUC、准确率以及F1分数等指标进行评估和比较，选出最优算法。按照7:3的比例划分为训练集和测试集，在训练集中进行五折交叉验证。超参数优化采用网格搜索方法，以提升算法性能。在测试集上评估最优算法的预测能力及其泛化性能。采用SHAP方法解释关键特征对严重意识障碍的影响。结果 本研究共从MIMIC-IV数据库中提取急性缺血性脑卒中患者1998例，其中471例（占23.6%）在入住ICU后30天内发生严重意识障碍。通过对多种机器学习算法在验证集上的评估与比较，最终选定XGBoost算法作为最优算法。研究中将数据按7:3的比例划分为训练集和测试集，并结合五折交叉验证与网格搜索优化超参数，结果表明XGBoost算法在测试集上展现了良好的严重意识障碍风险预测性能和泛化能力（AUC=0.788，95%CI：0.747~0.829；准确率=76.7%）。SHAP解释分析显示，ICU住院时间、血糖变异性及早期有创氧疗是卒中患者严重意识障碍发生风险增加的主要危险因素。结论 XGBoost算法在预测急性缺血性脑卒中患者严重意识障碍风险方面展现出较强的潜力。此外，SHAP解释分析突显了ICU住院时间、血糖变异性及氧气治疗对严重意识障碍风险的重要性。

Abstract

Objective This study, based on the MIMIC-IV database, aims to develop an interpretable machine learning model for predicting the risk of severe consciousness disturbances in stroke patients.
Methods Acute ischemic stroke patients were extracted from the MIMIC-IV database using ICD-9 and ICD-10 codes. LASSO regression was used for feature selection. Seven machine learning algorithms were evaluated and compared based on AUC, accuracy, and F1 score to select the optimal algorithm. The data were split into training and test sets in a 7:3 ratio, with five-fold cross-validation performed in the training set. Hyperparameter optimization was conducted using grid search to improve algorithm performance. The predictive ability and generalizability of the optimal algorithm were evaluated on the test set. SHAP (Shapley Additive Explanations) was applied to explain the impact of key features on the risk of severe consciousness disturbances.
Results A total of 1,998 acute ischemic stroke patients were extracted from the MIMIC-IV database, of whom 471 (23.6%) developed severe consciousness disturbances within 30 days of ICU admission. After evaluating and comparing various machine learning algorithms on the validation set, the XGBoost algorithm was selected as the optimal model. The data were split into training and test sets in a 7:3 ratio, with five-fold cross-validation and grid search optimization for hyperparameters. The results showed that the XGBoost algorithm demonstrated good predictive performance and generalization ability for the risk of severe consciousness disturbances on the test set (AUC=0.788, 95% CI: 0.747–0.829; accuracy=76.7%). SHAP analysis revealed that ICU length of stay, blood glucose variability, and early invasive oxygen therapy were the main risk factors for the development of severe consciousness disturbances in stroke patients.
Conclusion   The XGBoost algorithm shows strong potential in predicting the risk of severe consciousness disturbances in acute ischemic stroke patients. Additionally, SHAP analysis highlights the significance of ICU length of stay, blood glucose variability, and oxygen therapy in increasing the risk of severe consciousness disturbances.