What question did this study set out to answer?

April 29, 2026Medicine0 citationsOpen Access

A machine learning-based prediction model for in-hospital mortality in ICU patients with acute rhinosinusitis: A multi-database study using SHAP interpretability

MSManxiang SunFirst Affiliated Hospital of Chengdu Medical College LWL WangFirst Affiliated Hospital of Chengdu Medical College JYJing YangChengdu Medical College

Key Points

The aim is to develop a specific prediction model for in-hospital mortality in ICU patients with acute rhinosinusitis.
Conducted a retrospective analysis using data from three ICU databases with a total of 559 patients.
Employed the Boruta algorithm for feature selection and trained nine machine learning models.
Evaluated model performance using area under the receiver operating characteristic curve and other metrics.
The random forest model achieved the best performance with a maximum area under the curve of 0.799 in validation sets.
Key predictors identified included sequential organ failure assessment scores and platelet counts.
SHapley Additive exPlanations analysis highlighted Oxford acute severity of illness score as a core feature impacting model outputs.

Abstract

Acute rhinosinusitis (ARS) is a heterogeneous catarrhal or purulent inflammatory disease of the sinus mucosa with critical risks in the intensive care unit (ICU). Current mortality models lack specificity for their unique pathophysiology, necessitating tailored prediction tools. We conducted a retrospective analysis of 3 large ICU databases: Medical Information Mart for Intensive Care (MIMIC)-IV-3.1 (n = 392), MIMIC-III-CareVue (n = 89), and electronic intensive care unit collaborative research database −2.0 (n = 78). We employed the Boruta algorithm for feature selection and developed 9 machine learning models using MIMIC-IV as the training dataset, with MIMIC-III-CareVue and the electronic intensive care unit collaborative research database serving as external validation datasets. Model performance was evaluated by multiple metrics, including area under the receiver operating characteristic curve, sensitivity, specificity, and accuracy. We implemented L2 regularization and grid-based hyperparameter tuning to prevent overfitting. SHapley Additive exPlanations analysis was used to interpret the best-performing model. This study included 559 patients. The Boruta algorithm identified sequential organ failure assessment scores and platelet counts as potential predictors. Receiver operating characteristic curve analysis showed that the random forest model performed best in 2 external validation sets, with a maximum area under the receiver operating characteristic curve of 0.799, and high sensitivity and accuracy. SHapley Additive exPlanations analysis revealed Oxford acute severity of illness score scores as the core feature affecting model output, with simplified acute physiology score II scores, weight, blood urea nitrogen levels in the first 24 hours, and sequential organ failure assessment scores also showing high contribution. This study constructed an ICU prognostic model for ARS using multi-database clinical data and the top-performing random forest algorithm. It aids in identifying and managing high-risk ARS patients clinically and offers a reliable tool for rational medical resource allocation.

Demander à l'IA

Bookmark

View Full Paper