[An interpretable machine learning-based prediction model for risk of death for patients with ischemic stroke in intensive care unit]

Abstract

Objective: To construct an inherent interpretability machine learning model as an explainable boosting machine model (EBM) for predicting one-year risk of death in patients with severe ischemic stroke.

Methods: We randomly divided the data of 2369 eligible patients with severe ischemic stroke in the MIMIC-Ⅳ(2.0) database, who were admitted in ICU in 2008 to 2019, into a training dataset (80%) and a test dataset (20%), and assessed the prognosis of the patients using the EBM model. The prediction performance of the model was evaluated by calculating the area under the receiver operating characteristic (AUC) curve. The calibration curve and Brier score were used to evaluate the degree of calibration of the model, and a decision curve was generated to assess the net clinical benefit.

Results: The EBM model constructed in this study had good discrimination power, calibration and net benefit, with an AUC of 0.857 (95% CI: 0.831-0.887) for predicting prognosis of severe ischemic stroke. Calibration curve analysis showed that the standard curve of the EBM model was the closest to the ideal curve. Decision curve analysis showed that the model had the greatest net benefit rate at the prediction probability threshold of 0.10 to 0.80. The top 5 independent predictive variables based on the EBM model were age, SOFA score, mean heart rate, mechanical ventilation, and mean respiratory rate, whose significance scores ranged from 0.179 to 0.370.

Conclusion: This EBM model has a good performance for predicting the risk of death within one year in patients with severe ischemic stroke and allows clinicians to better understand the contributing factors of the patients' outcomes through the model interpretability.

Keywords: explainable boosting machine; inherent interpretability machine learning; mortality prediction; severe ischemic stroke.

图 1

研究对象筛选流程图 Flowchart of subject selection.

图 2

四种预测模型的受试者工作特征曲线 Receiver operating characteristic (ROC) curves of the 4 prediction models.

图 3

四种预测模型的校准曲线 Calibration curves of the 4 prediction models.

图 4

四种预测模型的决策曲线 Decision curve analysis for the 4 prediction models.

图 5

EBM模型所选前15个变量的相对重要性得分 Relative importance scores of the top 15 variables selected by the EBM model.

图 6

风险效应趋势图 Risk effect trend graph. A: Trend graph of risk effect of age. B: Trend graph of risk effect of SOFA score.