Non-Invasive Assessment of Intracranial Hypertension in Patients with Traumatic Brain Injury Using Computed Tomography Radiomic Features: A Pilot Study

Abstract

This study aimed to assess intracranial hypertension in patients with traumatic brain injury non-invasively using computed tomography (CT) radiomic features. Fifty patients from the primary cohort were enrolled in this study. The clinical data, pre-operative cranial CT images, and initial intracranial pressure readings were collected and used to develop a prediction model. Data of 20 patients from another hospital were used to validate the model. Clinical features including age, sex, midline shift, basilar cistern status, and ventriculocranial ratio were measured. Radiomic features-i.e., 18 first-order and 40 second-order features- were extracted from the CT images. LASSO method was used for features filtration. Multi-variate logistic regression was used to develop three prediction models with clinical (CF model), first-order (FO model), and second-order features (SO model). The SO model achieved the most robust ability to predict intracranial hypertension. Internal validation showed that the C-statistic of the model was 0.811 (95% confidence interval [CI]: 0.691-0.931) with the bootstrapping method. The Hosmer Lemeshow test and calibration curve also showed that the SO model had excellent performance. The external validation results showed a good discrimination with an area under the curve of 0.725 (95% CI: 0.500-0.951). Although the FO model was inferior to the SO model, it had better prediction ability than the CF model. The study shows that the radiomic features analysis, especially second-order features, can be used to evaluate intracranial hypertension non-invasively compared with conventional clinical features, given its potential for clinical practice and further research.

Keywords: computed tomography; intracranial hypertension; radiomics; traumatic brain injury.

FIG. 1.

Selection of a rectangular region of interest (ROI) with size of 20 pixels. *20 pixels.

FIG. 2.

Model establishment and analysis process. CT, computed tomography; CF, clinical features; FO, first-order features; SO, second-order features.

FIG. 3.

First-order feature selection using the least absolute shrinkage and selection operator (LASSO) binary logistic regression model. (A) LASSO coefficient profiles of the 18 first-order features. A coefficient profile plot was produced against the log (λ) sequence. Vertical line was drawn at the value selected using 10-fold cross-validation, where optimal λ resulted in three non-zero coefficients. (B) Tuning parameter (λ) selection in the LASSO model used 10-fold cross-validation via minimum criteria. Mean squared error was plotted versus log(λ). Dotted vertical lines were drawn at the optimal values by using the minimum criteria and the one standard error (SE) of the minimum criteria (the 1-SE criteria). A λ value of 0.05 was chosen (minimum criteria) according to 10-fold cross-validation.

FIG. 4.

Second-order feature selection using the least absolute shrinkage and selection operator (LASSO) binary logistic regression model. (A) LASSO coefficient profiles of the 40 second-order features. A coefficient profile plot was produced against the log (λ) sequence. Vertical line was drawn at the value selected using 10-fold cross-validation, where optimal λ resulted in three non-zero coefficients. (B) Tuning parameter (λ) selection in the LASSO model used 10-fold cross-validation via minimum criteria. Mean squared error was plotted versus log(λ). Dotted vertical lines were drawn at the optimal values by using the minimum criteria and the one standard error (SE) of the minimum criteria (the 1-SE criteria). A λ value of 0.06 was chosen (minimum criteria) according to 10-fold cross-validation.

FIG. 5.

(A) Performance of models based on second-order features, first-order features, and clinical features; the areas under the curves (AUCs) for these models were 0.81, 0.76, and 0.65, respectively. (B) Calibration curve of the second-order features in patients with traumatic brain injury. The dashed line stands for perfect prediction. The dotted line represents apparent estimates of predicted versus observed values. The solid line (on behalf of bias) shows the corrected estimates via employing 1000 bootstrap samples. The mean absolute error was 0.076.

FIG. 6.

The results of ROC curve analysis and calibration for predicting intracranial hypertension in the external validation cohort (using second-order model). (A) The area under the curve (AUC) of external cohort for predicting intracranial hypertension was 0.725 (95% confidence interval: 0.500–0951). (B) The calibration curve analysis of the nomogram in the external validation cohort. The dashed line stands for perfect prediction. The dotted line represents apparent estimates of predicted versus observed values. The solid line (on behalf of bias) shows the corrected estimates via employing 1000 bootstrap samples. The mean absolute error was 0.039.