Comparison Study of Multiple Machine Learning Models for Predicting Anxiety Among Neurosurgical ICU Family Members Using Critical Care Family Needs Inventory.

aimThis investigation intended to develop machine learning predictive models to predict anxiety status of neurosurgical ICU patients' relatives by taking all Critical Care Family Needs Inventory (CCFNI) items as predictive variables, identify core unmet family demands strongly correlated with psychological distress, and select the optimal model for clinical rapid anxiety screening to support targeted precise nursing interventions.

methodsA total of 1000 first-degree relatives of neurosurgical ICU patients with a Glasgow Coma Scale (GCS) score ≤ 8 at the First Affiliated Hospital of Wannan Medical University from January 2024 to December 2025 were enrolled as research subjects. On the third day after admission to the department, general clinical data, HADS scores, and CCFNI scale data of the subjects were collected. SPSS 26.0 was applied for traditional statistical analysis and R 4.3.3 was utilized to develop machine learning prediction models. The full dataset was randomly split into training and test sets at a 7:3 ratio. Multiple metrics including accuracy, sensitivity, specificity, F1-score, and AUC were adopted to thoroughly assess model efficacy, followed by screening core predictors linked to relatives' anxiety levels. The research obtained ethical clearance from our institutional review board (ID: 2018043) and fully abided by the ethical standards outlined in the Declaration of Helsinki.

resultsAmong 1000 neurosurgical ICU patients, the primary diagnoses were craniocerebral injury (40.2%) and hypertensive intracerebral hemorrhage (39.6%) constituted the primary diagnoses, and the majority of patients served as the main economic source of their families (60.4%). The average age of family members was 51.72 ± 15.22 years, with junior and senior high school education being the predominant educational attainment, and nearly half of them had no accompanying experience. Screening of family members' anxiety status showed that only 1.2% had no anxiety symptoms, and the incidence of moderate to severe anxiety was as high as 76.8%. Model validation was performed using a 7:3 ratio and leveraged 10-fold cross-validation to optimize hyperparameters. The comparison results of three machine learning models indicated that the support vector machine (SVM) algorithm delivered superior overall predictive efficacy, achieving favorable accuracy (0.8863), sensitivity (0.7264), specificity (0.9348), and F1 score (0.7463), and its AUC value reached 0.9364; the Random Forest model had the highest AUC value (0.9451) but relatively low sensitivity (0.5362). The Artificial Neural Network exhibited weaker overall predictive capacity relative to the other two algorithms. Variable importance analysis of the SVM model revealed that CCFNI7, CCFNI12, and CCFNI9 were the core indicators for predicting family members' anxiety status.

conclusionThe incidence of moderate to severe anxiety among neurosurgical ICU patients' relatives far exceeds that of patients' families in general ICU and anxiety symptoms are closely correlated with neurosurgical ICU specific stressors and unmet family demands. Compared with traditional statistical methods, machine learning algorithms exhibit prominent strengths in predicting family anxiety and screening core CCFNI items related to psychological distress. The SVM model can serve as a reliable clinical screening tool, which may support medical-family collaborative precise nursing and help relieve anxiety symptoms among relatives. IMPLICATIONS FOR PRACTICE: The optimal machine learning model screened in this study can assist clinicians in precisely recognizing the core demands of neurosurgical ICU patients' relatives, which offers evidence to develop personalized intervention protocols, elevating the standard of humanistic intensive care nursing and optimizing the medical experience of both patients and their relatives. REPORTING

methodThis study strictly followed the TRIPOD statement for transparent reporting of diagnostic and prognostic prediction models, which provides standardized reporting checklists for all model development and validation procedures [1]. Meanwhile, we referenced the integrated risk factor analysis framework proposed by Wang et al. to standardize the methodological interpretation of multidimensional predictive indicators in this research [2]. PATIENT OR PUBLIC CONTRIBUTION: Family members of Neurosurgical ICU patients cooperated in completing relevant questionnaires, providing core data support for the study. There was no patient or public contribution. IMPLICATION FOR NURSING MANAGEMENT: The proposed model empowers nursing administrators to streamline family need assessment, optimize manpower allocation, and standardize emotional intervention procedures. It facilitates refined Neurosurgical ICU nursing management and improves holistic nursing quality.