Related Experiment Videos
Machine Learning for Predicting Long-Term Outcomes in Neurocritical Care: A sub-analysis from the SYNAPSE-ICU Study
Anna S Scholze1, Laura Borgstedt1, Stefan J Schaller2
1TUM School of Medicine and Health, Department Clinical Medicine, Department of Anaesthesiology and Intensive Care Medicine, Munich, Germany.
Purpose:
Neurocritical conditions present significant challenges in the intensive care unit (ICU), especially during the initial hours. Accurate early risk assessment is essential for guiding acute care and influencing long-term outcomes. This study aims to evaluate the potential of machine learning (ML) models for risk prediction in neurocritical patients.
Methods:
This retrospective sub-analysis of the SYNAPSE-ICU Study applied machine learning to predict the 6-month outcome for patients with three different neurocritical conditions: subarachnoid hemorrhage (SAH), traumatic brain injury (TBI), or intracerebral hemorrhage (ICH). Models were trained at ICU admission, ICU day 3, ICU day 7, and Hospital discharge using the H2O framework, including GLM, DRF, GBM, XGBoost, and Stacked Ensembles. Predictive performance was evaluated using AUPRC and AUROC metrics.
Results:
2,202 patients were included with a complete data set to predict the 6-month outcome. At 6 months, 836 patients had a favorable outcome (Glasgow Outcome Scale Extended (GOSE) ≥5) and had a significantly higher GOSE at Hospital discharge (p<0.001). Predictive performance (pooled across five multiply-imputed datasets) was highest at ICU admission (AUROC 0.841 [95% CI 0.804-0.878]) and Hospital discharge (AUROC 0.885 [0.845-0.925]), with a transient reduction during the ICU stay (Day 3: AUROC 0.830 [0.789-0.870]; Day 7: AUROC 0.821 [0.777-0.865]). Age was the most important parameter at the respective time points, except for Hospital discharge, where the GOSE at discharge had the highest impact on the 6-month outcome.
Conclusion:
This sub-analysis demonstrates that machine learning models can effectively predict long-term outcomes for neurocritical patients, with accuracy improving from ICU admission to Hospital discharge. GOSE at discharge and age are critical parameters. The findings underscore the potential of ML models to improve prognostic assessments in neurocritical care.