Toader Corneliu, Munteanu Octavian, Radoi Mugurel Petrinel, Crivoi Carla, Covache-Busuioc Razvan-Adrian, Serban Matei, Ciurea Alexandru Vlad, Dobrin Nicolaie
Department of Neurosurgery, "Carol Davila" University of Medicine and Pharmacy, 020021 Bucharest, Romania.
Department of Vascular Neurosurgery, National Institute of Neurology and Neurovascular Diseases, 077160 Bucharest, Romania.
J Clin Med. 2025 Apr 14;14(8):2672. doi: 10.3390/jcm14082672.
: The Glasgow Coma Scale (GCS) is a cornerstone in neurological assessment, providing critical insights into consciousness levels in patients with traumatic brain injuries and other neurological conditions. Despite its clinical importance, traditional methods for predicting GCS scores often fail to capture the complex, multi-dimensional nature of patient data. This study aims to address this gap by leveraging machine learning (ML) techniques to develop accurate, interpretable models for GCS prediction, enhancing decision making in critical care. : A comprehensive dataset of 759 patients, encompassing 25 features spanning pre-, intra-, and post-operative stages, was used to develop predictive models. The dataset included key variables such as cognitive impairments, Hunt and Hess scores, and aneurysm dimensions. Six ML algorithms, including random forest (RF), XGBoost, and artificial neural networks (ANN), were trained and rigorously evaluated. Data preprocessing involved numerical encoding, standardization, and stratified splitting into training and validation subsets. Model performance was assessed using accuracy and receiver operating characteristic area under the curve (ROC AUC) metrics. : The RF model achieved the highest accuracy (86.4%) and mean ROC AUC (0.9592 ± 0.0386, standard deviation), highlighting its robustness and reliability in handling heterogeneous clinical datasets. XGBoost and SVM models also demonstrated strong performance (ROC AUC = 0.9502 and 0.9462, respectively). Key predictors identified included the Hunt and Hess score, aneurysm dimensions, and post-operative factors such as prolonged intubation. Ensemble methods outperformed simpler models, such as K-nearest neighbors (KNN), which struggled with high-dimensional data. : This study demonstrates the transformative potential of ML in GCS prediction, offering accurate and interpretable tools that go beyond traditional methods. By integrating advanced algorithms with clinically relevant features, this work provides a dynamic, data-driven framework for critical care decision making. The findings lay the groundwork for future advancements, including multi-modal data integration and broader validation, positioning ML as a vital tool in personalized neurological care.
格拉斯哥昏迷量表（GCS）是神经学评估的基石，能为创伤性脑损伤及其他神经疾病患者的意识水平提供关键见解。尽管其具有临床重要性，但传统的格拉斯哥昏迷量表评分预测方法往往无法捕捉患者数据复杂的多维度特性。本研究旨在通过利用机器学习（ML）技术开发准确、可解释的格拉斯哥昏迷量表预测模型来填补这一空白，以加强重症监护中的决策制定。
使用一个包含759名患者的综合数据集来开发预测模型，该数据集涵盖了术前、术中和术后阶段的25个特征。数据集中包括认知障碍、Hunt和Hess评分以及动脉瘤大小等关键变量。对六种机器学习算法进行了训练和严格评估，包括随机森林（RF）、XGBoost和人工神经网络（ANN）。数据预处理包括数值编码、标准化以及分层划分为训练子集和验证子集。使用准确率和曲线下接受者操作特征面积（ROC AUC）指标评估模型性能。
随机森林模型实现了最高准确率（86.4%）和平均ROC AUC（0.9592±0.0386，标准差），突出了其在处理异质临床数据集方面的稳健性和可靠性。XGBoost和支持向量机模型也表现出强大性能（ROC AUC分别为0.9502和0.9462）。确定的关键预测因素包括Hunt和Hess评分、动脉瘤大小以及术后诸如长时间插管等因素。集成方法优于简单模型，如在处理高维数据时遇到困难的K近邻（KNN）模型。
本研究证明了机器学习在格拉斯哥昏迷量表预测中的变革潜力，提供了超越传统方法的准确且可解释的工具。通过将先进算法与临床相关特征相结合，这项工作为重症监护决策制定提供了一个动态的、数据驱动的框架。研究结果为未来的进展奠定了基础，包括多模态数据整合和更广泛的验证，将机器学习定位为个性化神经护理中的重要工具。
