基于入院平扫CT的深度学习模型预测动脉瘤性蛛网膜下腔出血90天死亡:开发与外部验证
本研究开发并外部验证仅基于入院平扫CT的全自动深度学习模型,预测动脉瘤性蛛网膜下腔出血90天死亡。多中心回顾性纳入863例患者(586例训练、147例内部测试、130例外部队列验证)。深度学习模型在内部测试和外部验证中区分度与校准均与需多变量的临床模型相当,可作为互补决策支持工具。
要点
- 仅需入院平扫CT即可全自动预测aSAH后90天死亡,无需人工输入。
- 深度学习模型区分度与校准和需多变量的临床模型相当。
- 该工具定位为互补决策支持,而非替代传统预后评估。
核心结论
- 仅基于入院平扫CT的全自动深度学习模型可预测aSAH后90天死亡,区分度与校准和需多变量的临床模型相当。
研究设计与样本
- 多中心回顾性研究,9家三级医院用于模型开发,2个独立中心用于外部验证。
- 采用3维DenseNet-121架构迁移学习,以入院平扫CT为唯一输入训练深度学习模型。
- 构建三个逻辑回归对照模型:Core(年龄、WFNS分级)、Imaging(加改良Fisher分级、动脉瘤大小和位置)、Full Clinical(再加治疗方式)。
- 共纳入863例患者:586例训练、147例内部测试、130例外部验证。
- 评估指标包括区分度、分类、校准和决策曲线分析。
主要发现
- 内部测试AUC(95%CI):Core 0.856(0.790-0.913)、Imaging 0.853(0.780-0.916)、Full 0.844(0.766-0.909)、DL 0.855(0.786-0.917)。
- 外部验证AUC(95%CI):Core 0.823(0.738-0.895)、Imaging 0.793(0.705-0.871)、Full 0.798(0.707-0.873)、DL 0.806(0.724-0.876)。
- 所有模型均显示良好校准。
- 决策曲线分析显示在临床相关阈值范围内净获益相当,DL模型与传统方法间无显著性能差异(DeLong P > .05)。
临床怎么用
- 可在床旁仅凭常规入院平扫CT获得客观、可重复的aSAH死亡风险分层,无需额外数据采集。
- 该模型可作为传统预后评估的互补决策支持工具,而非替代。
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
BACKGROUND AND OBJECTIVES: Traditional prognostication after aneurysmal subarachnoid hemorrhage depends on subjective clinical grading and radiological scoring systems that exhibit inter-rater variability and require multiple variables, hampering clinical adoption. We aimed to develop and externally validate a fully automated deep learning (DL) model predicting 90-day mortality exclusively from admission noncontrast computed tomography (NCCT), requiring no manual input and providing objective, reproducible, image-only risk stratification. METHODS: This multicenter retrospective study included 9 tertiary hospitals for model development and 2 independent centers for external validation. A DL model was trained using 3-dimensional DenseNet-121 architecture with transfer learning, using admission NCCT scans as the sole input. Three comparator logistic regression models were constructed: Core (age, World Federation of Neurosurgical Societies grade), Imaging (adding modified Fisher grade, aneurysm size, and location), and Full Clinical (further including treatment modality). Performance metrics included discrimination, classification, calibration, and decision-curve analysis. RESULTS: The study included 863 patients: 586 for training, 147 for internal testing, and 130 for external validation. In internal testing, area under the curves (95% CI) were: Core 0.856 (0.790-0.913), Imaging 0.853 (0.780-0.916), Full 0.844 (0.766-0.909), and DL 0.855 (0.786-0.917). In external validation, area under the curves were: Core 0.823 (0.738-0.895), Imaging 0.793 (0.705-0.871), Full 0.798 (0.707-0.873), and DL 0.806 (0.724-0.876). All models demonstrated good calibration. Decision-curve analysis showed comparable net benefit across clinically relevant thresholds, with no significant performance differences between the DL model and conventional approaches (DeLong P > .05). CONCLUSION: A fully automated DL model based solely on admission NCCT predicts 90-day mortality after aneurysmal subarachnoid hemorrhage with discrimination and calibration comparable with clinical models requiring multiple variables. Rather than replacing conventional prognostication, this approach offers a complementary decision-support tool that requires no data collection beyond routine imaging, enabling objective, reproducible risk stratification at the point of care.
引用
- Cepeda S, Rizoudis A, Müller D, et al. Development and External Validation of a Deep Learning Model to Predict Mortality in Aneurysmal Subarachnoid Hemorrhage Using Admission Computed Tomography. Neurosurgery. 2026. PMID: 42023861. doi:10.1227/neu.0000000000004038
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。