Machine learning-based 28-day mortality prediction model for elderly neurocritically Ill patients

Background

The growing population of elderly neurocritically ill patients highlights the need for effective prognosis prediction tools. This study aims to develop and validate machine learning (ML) models for predicting 28-day mortality in intensive care units (ICUs).

Methods

Data were extracted from the Medical Information Mart for Intensive Care IV(MIMIC-IV) database, focusing on elderly neurocritical ill patients with ICU stays ≥ 24 h. The cohort was split into 70 % for training and 30 % for internal validation. We analyzed 58 variables, including demographics, vital signs, medications, lab results, comorbidities, and medical scores, using Lasso regression to identify predictors of 28-day mortality. Seven ML algorithms were evaluated, and the best model was validated with data from Guizhou Medical University Affiliated Hospital. A log-rank test was used to assess survival differences in Kaplan-Meier curves. Shapley Additive Explanations (SHAP) were used to interpret the best model, while subgroup analysis identified variations in model performance across different populations.

Results

The study included 1,773 elderly neurocritically ill patients, with a 28-day mortality rate of 28.6 %. The Light Gradient Boosting Machine (LightGBM) outperformed other models, achieving an area under the curve (AUC) of 0.896 in internal validation and 0.812 in external validation. Kaplan-Meier analysis showed that higher LightGBM prediction scores correlated with lower survival probabilities. Key predictors identified through SHAP analysis included partial pressure of arterial carbon dioxide (PaCO₂), Acute physiology and chronic health evaluation II (APACHE II), white blood cell count, age, and lactate. The LightGBM model demonstrated consistent performance across various subgroups.

Conclusions

The LightGBM model effectively predicts 28-day mortality risk in elderly neurocritically ill patients, aiding clinicians in management and resource allocation. Its reliable performance across diverse subgroups underscores its clinical utility.