Scholarlink

Background <p>Ischemic stroke is the most prevalent type of stroke worldwide, accounting for approximately 85% of all stroke cases and resulting in the deaths of about 6.4&#xa0;million people annually. Among stroke survivors, more than 60% experience varying degrees of functional impairment within three months following the stroke. Therefore, early identification and prediction of the risk of adverse outcomes after stroke are crucial. This study seeks to create an intuitive and clinically relevant machine learning model to predict the risk of adverse prognostic outcomes within one year in patients suffering from acute cerebral infarction.</p> Methods <p>A retrospective cohort study was performed using medical data from 888 patients at the Air Force Specialty Medical Center of the People’s Liberation Army of China to construct the model and evaluate its performance. Following screening, 771 patients were included and allocated into training and validation datasets in a 7:3 ratio. Patients were classified into poor and favorable prognostic groups based on the modified Rankin Scale (mRS) scores. The SHAP analysis method was employed to scrutinize the marginal effects of each feature on model outputs, delivering a nuanced analysis of the model’s structure from a global and local perspective. Furthermore, the LIME interpretation technique was utilized for an exhaustive dissection of the model’s local attributes.</p> Results <p>In this study, the Recursive Feature Elimination (RFE) algorithm was employed to identify 14 key variables, including D-dimer, HDL, and PNI, for constructing 11 machine learning models. Comparative analysis revealed that the Extreme Gradient Boosting model demonstrated optimal predictive performance, with its AUC value reaching 0.766 and accuracy as high as 0.835. Utilizing SHAP analysis, the study ranked the global importance of the 14 key feature variables, including D-Dimer. Through LIME analysis, the respective weights of different feature ranges within the predictive model were uncovered.</p> Conclusion <p>The interpretable machine learning predictive model developed in this study exhibited commendable accuracy in predicting adverse outcomes within one year for patients with acute cerebral infarction. In comparison to previous models, the integration of SHAP and LIME analyses significantly enhanced the model’s interpretability, thereby further bolstering the clinical decision-making process.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable machine learning for predicting one-year prognosis in acute ischemic stroke

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月27日

查看全文
  • Xing-Yu Zhu,
  • Tian-Tian Jiang,
  • Yan-Wei Yin

摘要

Background

Ischemic stroke is the most prevalent type of stroke worldwide, accounting for approximately 85% of all stroke cases and resulting in the deaths of about 6.4 million people annually. Among stroke survivors, more than 60% experience varying degrees of functional impairment within three months following the stroke. Therefore, early identification and prediction of the risk of adverse outcomes after stroke are crucial. This study seeks to create an intuitive and clinically relevant machine learning model to predict the risk of adverse prognostic outcomes within one year in patients suffering from acute cerebral infarction.

Methods

A retrospective cohort study was performed using medical data from 888 patients at the Air Force Specialty Medical Center of the People’s Liberation Army of China to construct the model and evaluate its performance. Following screening, 771 patients were included and allocated into training and validation datasets in a 7:3 ratio. Patients were classified into poor and favorable prognostic groups based on the modified Rankin Scale (mRS) scores. The SHAP analysis method was employed to scrutinize the marginal effects of each feature on model outputs, delivering a nuanced analysis of the model’s structure from a global and local perspective. Furthermore, the LIME interpretation technique was utilized for an exhaustive dissection of the model’s local attributes.

Results

In this study, the Recursive Feature Elimination (RFE) algorithm was employed to identify 14 key variables, including D-dimer, HDL, and PNI, for constructing 11 machine learning models. Comparative analysis revealed that the Extreme Gradient Boosting model demonstrated optimal predictive performance, with its AUC value reaching 0.766 and accuracy as high as 0.835. Utilizing SHAP analysis, the study ranked the global importance of the 14 key feature variables, including D-Dimer. Through LIME analysis, the respective weights of different feature ranges within the predictive model were uncovered.

Conclusion

The interpretable machine learning predictive model developed in this study exhibited commendable accuracy in predicting adverse outcomes within one year for patients with acute cerebral infarction. In comparison to previous models, the integration of SHAP and LIME analyses significantly enhanced the model’s interpretability, thereby further bolstering the clinical decision-making process.