Objective To develop and validate a deep learning system (DLS) model predicting hematoma expansion (HE) based on non-contrast (NC) CT and a score combining with clinical variables. Materials and methods The multicenter retrospective dataset (R), the multicenter prospective dataset (P1), and the single-center prospective dataset (P2) enrolled 2350, 460, and 96 intracerebral hemorrhage (ICH) patients for analysis, respectively. The DLS model was developed, validated, and tested in R-development (R-dev), R-validation (R-val), and P1, respectively. After exploring clinical predictors of HE using multivariable logistic regression on P1-development (P1-dev), a five-point score “ARCHES” (Ai-Reinforced intraCerebral Hemorrhage hematoma Expansion Score) combining clinical predictors with the DLS model was created. We compared the discrimination of the ARCHES, DLS, with other models using the receiver operating characteristic (ROC) and DeLong test. Results The areas under the curve (AUC) of the DLS model were 0.781 (95% CI: 0.761–0.800) in R-dev, and showed similar results in P1. The ARCHES score, which includes DLS, baseline National Institutes of Health Stroke Scale (NIHSS), onset-to-NCCT time and regular antihypertension history, showed significantly better discrimination (AUC, 0.820; 95% CI: 0.775–0.859) than the blend sign and any hypodensity and time from onset-to-NCCT (BAT) score and the meta-analysis prediction model in P1-dev, P1-validation (P1-val) and P2. Conclusions The DLS model provides an automated, objective, rapid, and readily deployable tool for HE prediction only based on NCCT. The DLS model and the ARCHES score significantly improve risk stratification with better performance than previous HE prediction models, enabling timely clinical decisions for intensive monitoring and anti-HE therapy. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> Lacking reliable NCCT-based deep learning models (DLS) and a score combining with clinical variables for HE prediction. <Emphasis Type="BoldItalic">Findings</Emphasis> A NCCT-based DLS model and a score combining with clinical variables were developed and tested, both showing good discrimination and calibration in predicting HE. <Emphasis Type="BoldItalic">Clinical relevant</Emphasis> The DLS model enables automated and rapid HE prediction, while the easy-to-use score combines imaging and clinical variables, making it ideal for use in emergency settings. Both models improve risk stratification and support clinical decision-making in acute intracranial hemorrhage. Graphical Abstract

Prediction of intracerebral hemorrhage hematoma expansion: value of a novel deep learning system score

期刊论文
发表日期：

2026年1月29日

查看全文

Na Li,
Lei Wu,
Wanxing Ye,
Qiang Zhang,
Ravikiran Mane,
Xia Meng,
Anxin Wang,
Zeqiang Ji,
Kaijiang Kang,
Yaou Liu,
Yunyun Duan,
Zhenzhou Wu,
Zixiao Li,
Hao Li,
Xingquan Zhao

摘要

Objective

To develop and validate a deep learning system (DLS) model predicting hematoma expansion (HE) based on non-contrast (NC) CT and a score combining with clinical variables.

Materials and methods

The multicenter retrospective dataset (R), the multicenter prospective dataset (P1), and the single-center prospective dataset (P2) enrolled 2350, 460, and 96 intracerebral hemorrhage (ICH) patients for analysis, respectively. The DLS model was developed, validated, and tested in R-development (R-dev), R-validation (R-val), and P1, respectively. After exploring clinical predictors of HE using multivariable logistic regression on P1-development (P1-dev), a five-point score “ARCHES” (Ai-Reinforced intraCerebral Hemorrhage hematoma Expansion Score) combining clinical predictors with the DLS model was created. We compared the discrimination of the ARCHES, DLS, with other models using the receiver operating characteristic (ROC) and DeLong test.

Results

The areas under the curve (AUC) of the DLS model were 0.781 (95% CI: 0.761–0.800) in R-dev, and showed similar results in P1. The ARCHES score, which includes DLS, baseline National Institutes of Health Stroke Scale (NIHSS), onset-to-NCCT time and regular antihypertension history, showed significantly better discrimination (AUC, 0.820; 95% CI: 0.775–0.859) than the blend sign and any hypodensity and time from onset-to-NCCT (BAT) score and the meta-analysis prediction model in P1-dev, P1-validation (P1-val) and P2.

Conclusions

The DLS model provides an automated, objective, rapid, and readily deployable tool for HE prediction only based on NCCT. The DLS model and the ARCHES score significantly improve risk stratification with better performance than previous HE prediction models, enabling timely clinical decisions for intensive monitoring and anti-HE therapy.

Key Points

Question Lacking reliable NCCT-based deep learning models (DLS) and a score combining with clinical variables for HE prediction.

Findings A NCCT-based DLS model and a score combining with clinical variables were developed and tested, both showing good discrimination and calibration in predicting HE.

Clinical relevant The DLS model enables automated and rapid HE prediction, while the easy-to-use score combines imaging and clinical variables, making it ideal for use in emergency settings. Both models improve risk stratification and support clinical decision-making in acute intracranial hemorrhage.

Graphical Abstract