机器学习预测神经重症患者长期结局:SYNAPSE-ICU子分析
该研究为SYNAPSE-ICU研究的回顾性子分析,纳入2202例神经重症患者,应用多种机器学习模型预测6个月结局。模型在ICU入院时和出院时预测性能最高,出院时GOSE和年龄是最重要参数。
要点
- 机器学习模型可有效预测神经重症患者6个月结局。
- ICU入院时和出院时预测性能最高,ICU期间短暂下降。
- 出院时GOSE和年龄是影响6个月结局的关键参数。
核心结论
- 机器学习模型能有效预测神经重症患者长期结局,准确性从ICU入院至出院逐步提高。
研究设计与样本
- SYNAPSE-ICU研究的回顾性子分析,纳入2202例具备完整数据的神经重症患者。
- 采用H2O框架训练GLM、DRF、GBM、XGBoost和Stacked Ensembles模型。
- 在ICU入院、ICU第3天、ICU第7天和出院时分别训练模型,以AUPRC和AUROC评估性能。
主要发现
- 6个月时836例患者结局良好(GOSE≥5),其出院时GOSE显著更高(P < 0.001)。
- 预测性能在ICU入院时AUROC为0.841(95% CI 0.804-0.878),出院时为0.885(0.845-0.925)。
- ICU第3天AUROC为0.830(0.789-0.870),第7天为0.821(0.777-0.865)。
- 年龄是各时间点最重要参数,出院时出院GOSE对6个月结局影响最大。
临床怎么用
- 机器学习模型可辅助神经重症患者早期风险分层和预后评估。
- 出院时GOSE和年龄应作为长期结局预测的关键参考指标。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: 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 3 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: In all, 2202 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 5 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. CONCLUSIONS: 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.
引用
- Scholze AS, Borgstedt L, Schaller SJ, et al. Machine Learning for Predicting Long-Term Outcomes in Neurocritical Care: A Sub-Analysis from the SYNAPSE-ICU Study. World Neurosurg. 2026. PMID: 42833556. doi:10.1016/j.wneu.2026.125391
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。