A comparison of machine learning models and established grading scales for hospital mortality prediction in non‐traumatic subarachnoid haemorrhage
Hong Kong Journal of Emergency Medicine · 2026-12 · 33 (6)
DOI: 10.1002/hkj2.70151 ↗Kurz gefasst
Fachartikel in der Zeitschrift Hong Kong Journal of Emergency Medicine, online erschienen am 15.11.2026. Verfasst von Jennifer Wing Suen Lai und 3 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Introduction
Existing severity grading scales show limited ability to guide decisions in subarachnoid haemorrhage (SAH). Machine learning (ML) offers a promising alternative. We compared ML models against established scales—World Federation of Neurosurgeons (WFNS), Fisher, Glasgow Coma Scale (GCS) and Acute Physiology and Chronic Health Evaluation (APACHE) IV—for predicting hospital mortality in Intensive Care Unit (ICU) patients with non‐traumatic SAH.
Methods
We collected relevant data from adults with non‐traumatic SAH admitted between January 2011 and March 2025. Two multivariable prognostic models were developed and compared across logistic regression (LR) and three ML algorithms (random forest [RF], artificial neural network [ANN] and eXtreme Gradient Boosting). Model 1 included demographics, GCS, radiographic findings, vasospasm, 5‐day interventions and APACHE IV scores, whereas Model 2 omitted APACHE IV. Model performance was assessed with area under the receiver operating characteristic curve (AUROC) and area under the precision‐recall curve.
Results
Among conventional scales, APACHE IV discriminated best (AUROC 0.871, 95% CI 0.839–0.902), followed by GCS severity (0.804), WFNS (0.780), and Fisher (0.540). Model 1 achieved AUROCs ranging from 0.916 to 0.935 (ANN AUROC 0.935, LR AUROC 0.928), significantly outperforming individual admission scales. For Model 2 (without APACHE IV), predictive performance remained high with AUROCs ranging from 0.902 to 0.917 (RF AUROC 0.917, LR AUROC 0.908), demonstrating discrimination comparable to APACHE IV alone (AUROC 0.871, p > 0.05) despite requiring substantially fewer variables.
Conclusion
In non‐traumatic SAH, multivariable models incorporating 5‐day complication and intervention data outperformed WFNS and the Fisher scale. However, complex ML algorithms showed no significant performance advantage over a LR model. Notably, a streamlined model omitting APACHE IV achieved discrimination comparable to the full 129‐variable score while requiring a fraction of the data collection effort. Together, these findings support adopting complication‐ and intervention‐aware models to refine risk prognostication and guide ICU decision‐making.
Zitieren
APA
Lai, J. W. S., Man, M., Yu, J. S. Y., & Shum, H. (2026). A comparison of machine learning models and established grading scales for hospital mortality prediction in non‐traumatic subarachnoid haemorrhage. Hong Kong Journal of Emergency Medicine, 33(6). https://doi.org/10.1002/hkj2.70151
Vancouver
Lai JWS, Man M, Yu JSY, Shum H. A comparison of machine learning models and established grading scales for hospital mortality prediction in non‐traumatic subarachnoid haemorrhage. Hong Kong Journal of Emergency Medicine. 2026;33(6). doi:10.1002/hkj2.70151
BibTeX
@article{Lai2026,
author = {Lai, Jennifer Wing Suen and Man, Man‐Yee and Yu, Judianna Suet Ying and Shum, Hoi‐Ping},
title = {A comparison of machine learning models and established grading scales for hospital mortality prediction in non‐traumatic subarachnoid haemorrhage},
journal = {Hong Kong Journal of Emergency Medicine},
year = {2026},
volume = {33},
number = {6},
doi = {10.1002/hkj2.70151}
}
EUVIMED