A model comprising the blend sign and black hole sign shows good performance for predicting early intracerebral haemorrhage expansion: a comprehensive evaluation of CT features

Abstract

Objective: To predict early intracerebral haemorrhage expansion (HE) by comprehensive evaluation of commonly used noncontrast computed tomography (NCCT) features.

Methods: Two hundred eighty-eight patients who had a spontaneous intracerebral haemorrhage (ICH) were included. All of the patients had undergone baseline NCCT within 6 h after ICH symptom onset. Ten NCCT features were extracted. Univariate analysis and multivariable logistic regression analysis were used to select the features. Using the finally selected features, a logistic regression model was built with a training cohort (n = 202) and subsequently validated in an independent test cohort (n = 86). Additionally, stratification analysis was performed in cases with and without anticoagulant therapy.

Results: HE was found in 78 patients (27.1%). The blend sign and black hole sign were finally selected. The logistic regression model built with the two features exhibited accuracies of 76.7% and 75.6%, specificities of 98.6% and 98.4%, and positive predictive values (PPVs) of 83.3% and 75.0% for the training and test cohorts, respectively. The model also showed specificities of 100% and 98.5% and PPVs of 100% and 76.9% for the anticoagulant and non-anticoagulant drug use groups, respectively. These performances were better than those of each of the separate features.

Conclusions: By comprehensive evaluation, the model comprising the blend sign and black hole sign showed good performance for predicting early intracerebral haemorrhage expansion, particularly for high specificity and PPV, regardless of the anticoagulant status.

Key points: • Early identification of patients who are more likely to have haematoma expansion is important for therapeutic intervention. • Many radiological features have been reported to correlate with intracerebral haemorrhage expansion. • By integrating only the blend sign and black hole sign, the logistic regression model showed good performance for predicting early intracerebral haemorrhage expansion.

Keywords: Cerebral haemorrhage; Machine learning; Multidetector computed tomography.