Xiao-Hong Su,1,* Wei-Peng Li,2,3,* Xiao-Feng Xu,1,* Xiao-Ling Su,4 Jia Liu,1 Shi-Yuan Feng,1 Jun-Yu Liu,1 Rui-Qi Dong,1 Iok Keng Ngai,1 Lu Yang,1 Li Xu,1 Zhe-Qi Li,1 Dong-Cheng Li,1 Ying Jiang,1 Fu-Hua Peng1 

1Department of Neurology, The Third Affiliated Hospital of Sun Yat-Sen University, Guangzhou, Guangdong, People’s Republic of China; 2School of Traditional Chinese Medicine, Southern Medical University, Guangzhou, People’s Republic of China; 3Department of Neurology, Southern Medical University Hospital of Integrated Traditional Chinese and Western Medicine, Southern Medical University, Guangzhou, People’s Republic of China; 4Department of Cardiovascular and Oncology, Zhongshan People’s Hospital of Torch Development Zone, Zhongshan, People’s Republic of China

*These authors contributed equally to this work

Correspondence: Fu-Hua Peng; Ying Jiang, Department of Neurology, The Third Affiliated Hospital, Sun Yat-Sen University, 600 Tianhe Road, Guangzhou, Guangdong, 510630, People’s Republic of China, Tel +86-20-85253275, Email pengfh@mail.sysu.edu.cn; jiangy9@mail.sysu.edu.cn

Introduction: Autoimmune glial fibrillary acidic protein astrocytopathy (A-GFAP-A) is an increasingly recognized neurological disorder with significant clinical management challenges, particularly in predicting the need for intensive care unit (ICU) admission. This study aimed to develop and validate predictive models to identify A-GFAP-A patients at increased risk for ICU admission.
Methods: We retrospectively analyzed 107 patients (January 2021 - August 2024), randomly assigned to training and validation cohorts (7:3). Variable selection for model development was performed using random forest, least absolute shrinkage and selection operator (LASSO), and extreme gradient boosting (XGBoost). Logistic regression was used to construct a nomogram, and a decision tree was developed to facilitate rapid clinical decision-making. Model performance was assessed by area under the curve (AUC), calibration plots, and decision curve analysis (DCA).
Results: Four key predictors of ICU admission were identified: Glasgow Coma Scale (GCS) score at admission, seizures, maximum body temperature, and C-reactive protein (CRP) levels. The nomogram demonstrated excellent predictive accuracy with AUCs of 0.923 (95% CI, 0.858– 0.987) in training cohort, 0.922 (95% CI, 0.836– 1.000) in validation cohort, and 0.93 (95% CI, 0.883– 0.972) in bootstrap validation. The model showed excellent calibration, and DCA confirmed its clinical utility. The decision tree identified GCS < 15, seizures, and temperature > 39°C as the most relevant indicators for high-risk stratification.
Discussion: This study presents the first validated nomogram and decision tree for ICU admission risk in A-GFAP-A, based on the largest reported cohort to date, providing a valuable tool for clinical decision-making and resource optimization.

Keywords: prediction model, glial fibrillary acidic protein astrocytopathy, intensive care unit