Predicting unplanned readmissions in neurological patients: A large-scale AI-driven analysis using CHAID decision trees
- 1Critical Care Nursing, Clinical Nursing Department, School of Nursing, University of Jordan, Amman, Jordan
- 2Department of Training and Continuing Education, King Hussein Medical Hospital, Jordanian Royal Medical Services/Army Forces, Amman, Jordan
Abstract
Background: Unplanned readmissions in neurology signify care gaps, increasing costs, and morbidity. This study aimed to develop and internally validate an AI model to predict unplanned readmissions among neurological patients in Jordan. Method and materials: We conducted a retrospective analysis of 2,795 adult neurology admissions (2022–2024). Patients’ diagnosis was grouped and analyzed based on the International Classification of Diseases and Related Health Problems, Tenth Revision. Nine prediction models were compared using Electronic Health Record (EHR) data. The Chi-square Automatic Interaction Detection (CHAID) decision tree was selected for optimal performance and interpretability. Results: The unplanned readmission rate was 8.9%, highest in the neoplasm group (24%). The CHAID model demonstrated modest discriminative ability (AUC = 0.671). For risk stratification, it achieved a balanced accuracy of 0.63, sensitivity of 0.52, and specificity of 0.74. Key predictors were Charlson Comorbidity Index (CCI ≥ 5), age, geographic region, length of stay, and sex. Risk-stratification metrics were calculated descriptively from the complete cohort, whereas independent model validation was based on the held-out test-set AUC. Subgroup analysis revealed distinct risk profiles: vascular patients with high CCI (p < .001, χ2 = 66.368), young females of whom 37.3% had neoplasms (p = .037, χ2 = 4.367), and functional disorder patients from northern regions (p = .009, χ2 = 8.897). Conclusion: This is one of the first studies from the Middle East to demonstrate the potential of the CHAID model for risk stratification in neurology readmissions. The model offers an interpretable framework to identify high-risk patient subgroups but requires external validation before clinical application. Its poor-to-fair discriminative performance (AUC = 0.671) supports population-level risk stratification rather than individual-level prediction.
Related
- Neutrophil-to-lymphocyte ratio, a predictor for mortality and unplanned hospital readmissions in patients aged 75 years and older with cancer.
- Systematic evaluation of long-read and short-read sequencing in neurological disorders diagnosis: a direct comparison study of 310 patients
- Eye movement analysis detects early neurological alterations in cirrhotic patients classified as not having minimal hepatic encephalopathy by the Psychometric Hepatic Encephalopathy Score
- Emicizumab and shared medical decision in patients living with hemophilia A
- Predicting Severe Toxicities and Mortality in Older Patients with Colorectal Cancer: An Analysis of 3 Clinical Trials and 3 Prospective Cohorts
BibTeX
@article{alkharabsheh2026predicting,
title = {{Predicting unplanned readmissions in neurological patients: A large-scale AI-driven analysis using CHAID decision trees}},
author = {Randa Al-Kharabsheh and Muayyad Ahmad},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0360077},
url = {https://doi.org/10.1371/journal.pone.0360077},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Predicting unplanned readmissions in neurological patients: A large-scale AI-driven analysis using CHAID decision trees
AU - Randa Al-Kharabsheh
AU - Muayyad Ahmad
PY - 2026
DA - 2026/10/09
JO - PLOS One
DO - 10.1371/journal.pone.0360077
UR - https://doi.org/10.1371/journal.pone.0360077
AB - Background: Unplanned readmissions in neurology signify care gaps, increasing costs, and morbidity. This study aimed to develop and internally validate an AI model to predict unplanned readmissions among neurological patients in Jordan. Method and materials: We conducted a retrospective analysis of 2,795 adult neurology admissions (2022–2024). Patients’ diagnosis was grouped and analyzed based on the International Classification of Diseases and Related Health Problems, Tenth Revision. Nine prediction models were compared using Electronic Health Record (EHR) data. The Chi-square Automatic Interaction Detection (CHAID) decision tree was selected for optimal performance and interpretability. Results: The unplanned readmission rate was 8.9%, highest in the neoplasm group (24%). The CHAID model demonstrated modest discriminative ability (AUC = 0.671). For risk stratification, it achieved a balanced accuracy of 0.63, sensitivity of 0.52, and specificity of 0.74. Key predictors were Charlson Comorbidity Index (CCI ≥ 5), age, geographic region, length of stay, and sex. Risk-stratification metrics were calculated descriptively from the complete cohort, whereas independent model validation was based on the held-out test-set AUC. Subgroup analysis revealed distinct risk profiles: vascular patients with high CCI (p < .001, χ2 = 66.368), young females of whom 37.3% had neoplasms (p = .037, χ2 = 4.367), and functional disorder patients from northern regions (p = .009, χ2 = 8.897). Conclusion: This is one of the first studies from the Middle East to demonstrate the potential of the CHAID model for risk stratification in neurology readmissions. The model offers an interpretable framework to identify high-risk patient subgroups but requires external validation before clinical application. Its poor-to-fair discriminative performance (AUC = 0.671) supports population-level risk stratification rather than individual-level prediction.
SN - 1932-6203
ER -