Department of Radiology, Mayo Clinic, Rochester, MN, USA; Department of Computer and Information Science, University of Pennsylvania, Philadelphia, PA, USA.
Dr Kiran C Patel College of Allopathic Medicine, Nova Southeastern University, Davie, FL, USA.
J Stroke Cerebrovasc Dis. 2024 Jan;33(1):107489. doi: 10.1016/j.jstrokecerebrovasdis.2023.107489. Epub 2023 Nov 17.
Predicting patient recovery and discharge disposition following mechanical thrombectomy remains a challenge in patients with ischemic stroke. Machine learning offers a promising prognostication approach assisting in personalized post-thrombectomy care plans and resource allocation. As a large national database, National Inpatient Sample (NIS), contain valuable insights amenable to data-mining. The study aimed to develop and evaluate ML models predicting hospital discharge disposition with a focus on demographic, socioeconomic and hospital characteristics.
The NIS dataset (2006-2019) was used, including 4956 patients diagnosed with ischemic stroke who underwent thrombectomy. Demographics, hospital characteristics, and Elixhauser comorbidity indices were recorded. Feature extraction, processing, and selection were performed using Python, with Maximum Relevance - Minimum Redundancy (MRMR) applied for dimensionality reduction. ML models were developed and benchmarked prior to interpretation of the best model using Shapley Additive exPlanations (SHAP).
The multilayer perceptron model outperformed others and achieved an AUROC of 0.81, accuracy of 77 %, F1-score of 0.48, precision of 0.64, and recall of 0.54. SHAP analysis identified the most important features for predicting discharge disposition as dysphagia and dysarthria, NIHSS, age, primary payer (Medicare), cerebral edema, fluid and electrolyte disorders, complicated hypertension, primary payer (private insurance), intracranial hemorrhage, and thrombectomy alone.
Machine learning modeling of NIS database shows potential in predicting hospital discharge disposition for inpatients with acute ischemic stroke following mechanical thrombectomy in the NIS database. Insights gained from SHAP interpretation can inform targeted interventions and care plans, ultimately enhancing patient outcomes and resource allocation.
在缺血性脑卒中患者中，预测机械取栓后患者的恢复和出院去向仍然具有挑战性。机器学习提供了一种很有前途的预后方法，可以辅助制定个体化的取栓后护理计划和资源分配。国家住院患者样本（National Inpatient Sample，NIS）是一个大型国家数据库，包含了可用于数据挖掘的宝贵见解。本研究旨在开发和评估用于预测医院出院去向的机器学习模型，重点关注人口统计学、社会经济学和医院特征。
使用 NIS 数据集（2006-2019 年），包括 4956 例接受取栓治疗的缺血性脑卒中患者。记录了人口统计学、医院特征和 Elixhauser 合并症指数。使用 Python 进行特征提取、处理和选择，应用最大相关性-最小冗余度（Maximum Relevance - Minimum Redundancy，MRMR）进行降维。在解释最佳模型之前，开发并基准测试了机器学习模型，然后使用 Shapley Additive exPlanations（SHAP）进行解释。
多层感知机模型表现优于其他模型，其 AUROC 为 0.81，准确率为 77%，F1 得分为 0.48，精确率为 0.64，召回率为 0.54。SHAP 分析确定了预测出院去向的最重要特征，包括吞咽困难和构音障碍、NIHSS、年龄、主要支付者（医疗保险）、脑水肿、液体和电解质紊乱、复杂高血压、主要支付者（私人保险）、颅内出血和单纯取栓。
NIS 数据库的机器学习建模显示出在预测 NIS 数据库中接受机械取栓的急性缺血性脑卒中患者的医院出院去向方面具有潜力。通过 SHAP 解释获得的见解可以为有针对性的干预和护理计划提供信息，最终改善患者的预后和资源分配。
