Predicting one-year mortality risk in ICU patients with ischemic stroke using multi-algorithm machine learning and a nomogram

Abstract

Introduction: Ischemic stroke is a leading cause of mortality, and patients requiring intensive care unit (ICU) admission carry a guarded prognosis. We aimed to develop and validate a predictive model for estimating one-year mortality risk in ICU-admitted ischemic stroke patients.

Methods

In this retrospective cohort study, data from 1974 ischemic stroke patients were extracted from the MIMIC-IV database. Patients were randomly allocated into a training set (n = 1582) and a test set (n = 392) in an 8:2 ratio. Five machine learning algorithms (CART, RF, SVM, GBM, and NB) were employed for initial feature screening. Key predictors were subsequently integrated into a multivariate logistic regression model, which was visualized as a nomogram. The model's performance was evaluated using discrimination, calibration, and clinical utility metrics and was compared against established severity scores (SOFA, SAPS II, LODS, OASIS, GCS).

Results

The final nomogram incorporated nine predictors: age, heart rate, weight, glucose, anion gap, calcium, alkaline phosphatase (ALP), red cell distribution width (RDW), and mean corpuscular hemoglobin concentration (MCHC). The model demonstrated an AUC of 0.739 (95% CI: 0.715–0.764) in the training set and 0.737 (95% CI: 0.688–0.786) in the test set. Calibration curves indicated good agreement between predictions and observations. Decision curve and clinical impact curve analyses confirmed the nomogram's favorable clinical net benefit, and it outperformed all comparator scoring systems in discriminatory ability.

Conclusions

We developed and validated a practical nomogram that effectively integrates key clinical variables to predict one-year mortality risk in ICU patients with ischemic stroke. This tool demonstrates robust performance and potential clinical utility for risk stratification.