Mortality risk prediction model for patients with acute ischemic stroke: a machine learning method based on intrinsic interpretability
急性缺血性脑卒中患者死亡风险预测模型：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，用于预测卒中患者的ICU死亡风险。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ICU死亡风险的影响。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436例（占21.8%）在入住ICU后30天内死亡。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ICU死亡风险预测性能和泛化能力（AUC=0.821，95%CI：0.778~0.864；准确率=80.7%）。SHAP解释分析显示，早期有创氧疗和高龄是卒中患者ICU死亡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ICU死亡风险方面展现出较强的潜力。此外，SHAP解释分析突显了早期有创氧疗和高龄对ICU死亡风险的重要性。

Abstract

Objective This study, based on the MIMIC-IV database, aims to develop an interpretable machine learning model to predict the ICU mortality risk in stroke patients.
Methods Acute ischemic stroke patients were extracted from the MIMIC-IV database based on ICD-9 and ICD-10 codes. Feature selection was performed using the LASSO regression algorithm. Seven machine learning algorithms were evaluated and compared using metrics such as AUC, accuracy, and F1 score to identify the optimal algorithm. The dataset was split into training and testing sets in a 7:3 ratio, and five-fold cross-validation was conducted on the training set. Hyperparameter optimization was performed using grid search to enhance algorithm performance. The predictive ability and generalization performance of the optimal algorithm were evaluated on the test set. SHAP analysis was used to interpret the impact of key features on ICU mortality risk.
Results: A total of 1,998 acute ischemic stroke patients were extracted from the MIMIC-IV database, of which 436 (21.8%) died within 30 days of ICU admission. After evaluating and comparing the performance of multiple machine learning algorithms on the validation set, the XGBoost algorithm was selected as the optimal model. The data were divided into training and testing sets in a 7:3 ratio, and five-fold cross-validation with grid search was employed for hyperparameter optimization. The results showed that the XGBoost algorithm demonstrated excellent ICU mortality risk prediction performance and generalization ability on the test set (AUC = 0.821, 95% CI: 0.778–0.864; accuracy = 80.7%). SHAP analysis revealed that early invasive oxygen therapy and advanced age were the primary risk factors for increased ICU mortality in stroke patients.
Conclusion   The XGBoost algorithm shows strong potential for predicting ICU mortality risk in acute ischemic stroke patients. Moreover, SHAP analysis highlights the significant roles of early invasive oxygen therapy and advanced age in determining ICU mortality risk.