This study was approved by the Ethics Committee of the Zhejiang Provincial People's Hospital. Due to the retrospective nature of the study, the patient's informed consent was waived. Patients' clinical data and NCCT data were obtained from routine clinical and radiological records. All patients with a clinically confirmed diagnosis of AIS who underwent interventional treatment from 1 January 2016 to 31 December 2020 were included. The inclusion criteria included (1) all patients who were diagnosed with AIS with large vessel occlusion (ICA isolated or in tandem with MCA) by preoperative one-stop head CT at admission, (2) intravascular intervention was carried out within the time window, (3) NCCT scan was performed immediately after patient intervention, and (4) complete clinical and imaging data could be obtained. The exclusion criteria included (1) patients with vascular malformation, intracranial hemorrhage, infection, or neoplastic lesions; (2) patients with a history of severe heart, lung, or kidney disease; and (3) postoperative NCCT images with obvious motion artifacts.

The mRS and the National Institutes of Health Stroke Scale (NIHSS) were obtained from clinical records. The primary outcome measure was defined as whether the postoperative functional status was classified as a “good prognosis”, defined as a 90-day mRS 0-2, or “poor prognosis”, defined as a 90-day mRS 3-6, including severe disability and death. Finally, patients were divided into good prognosis (n = 84) and poor prognosis (n = 57) groups. At a ratio of 7:3, all patients were also divided into training (n = 97) and validation (n = 44) sets according to the subject serial number (Figure 1).

NCCT scans after interventional treatment were performed using the Siemens Definition AS 128 CT. The routine head scan protocol: the tube voltage = 120 kV, the reference current = 400 mA, and the actual current can be adjusted by using the combined applications reduce exposure dose 4 dimensions (CARE dose 4D) technology: acquisition matrix = 512 × 512, rebuild FOV = 300 × 300 mm, layer thickness = 1 mm, and interslice gap = 0. The emergency head scan protocol: the tube voltage = 120 kV, the reference current = 400 mA, the actual current can be adjusted using the CARE dose 4D technology, acquisition matrix = 512 × 512, rebuild FOV = 300 × 300 mm, and pitch = 0:9 mm.

Based on the NCCT images after interventional treatment, 3D slicer software was used to segment 3D ROIs on the infarcted cerebral hemispheres. The detailed process is shown in Supplementary Figure S1.

The images were preprocessed using AK software (Artificial Intelligence Kit V3.0.0.R, GE Healthcare), which included image interpolation, intensity normalization, and gray-level discretization as described previously. First, the image grayscale intensity level was discretized and normalized for noise reduction by downsampling each image into 25 bins. Given these fixed bin values and numbers, the grayscale range of the image was divided into equally spaced intervals. Next, we calculated 396 texture features, including histogram, formfactor, Haralick, run-length matrix (RLM), gray-level cooccurrence matrix (GLCM), and gray-level size zone matrix (GLSZM) with AK software. Prior to feature selection, all the extracted texture features were standardized. Dimension reduction was performed using analysis of variance and Mann–Whitney U-test and then we performed a correlation test to reduce data redundancy. Finally, the least absolute shrinkage and selection operator (LASSO) was used to further select significant features (Figure 2).

A variety of machine learning algorithms, including, multivariate logistic regression (LR), Bayes, random forest, and decision tree classifiers, were undertaken to construct models based on the remaining features. The stability of each machine learning algorithm was quantified using the relative SD (RSD) and a bootstrap approach. For each classification method, we trained the model on a subsampled training cohort from the training set or the testing set and evaluated its performance on the remaining data using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. Subsampling of the training or testing set was performed 100 times using bootstrapping. RSD is the absolute value of the coefficient of variation and is often expressed as a percentage according to Equation: RSD = σAUC/μAUC → 100%, where σAUC and μAUC are the standard deviation and mean of the 100 AUC values, respectively. It should be noted that higher stability corresponds to lower RSD values. The calibration curve was used to describe the goodness-of-fit of radiomic models. Thereafter, we calculated the radiomic score (rad-score) for every patient in both the training and validation sets using the formula constructed in the training set.

In the training set, multivariable LR analysis was performed to select independent predictors of the clinical prognosis of AIS for each potential predictive variable, including age, gender, smoking, hypertension, diabetes, atrial fibrillation, use of anticoagulants, hyperlipidemia, the NIHSS at admission, the Alberta Stroke Program Early CT (ASPECT) score, bridging treatment, time of surgery, good revascularization, and rad-scores. Finally, the independent predictors from the training set were used to construct a joint prediction model using stepwise LR, and the data in the testing set were used to verify the performance of the joint model. Then, the ROC curves were used to visualize the experimental results, and the AUC was calculated to quantify the prediction performance.

All statistical analyses were performed using SPSS (version 25.0) and R 3.5.1. The Kolmogorov–Smirnov test was used to test the normality of the data. Normally distributed data were evaluated using an independent sample t-test, whereas non-normally distributed data were evaluated using a Mann–Whitney U-test. The difference between categorical variables was tested with a chi-squared test. The correlations between mRS and optimal texture features were used in Spearman's analysis. Besides, the subjects were stratified into different subgroups using the median values of clinical factors and imaging biomarkers for correlation analysis. The Sankey diagram was used to show the relationship between these subgroups. A two-tailed P-value of < 0.05 indicated statistical significance.

There were significant differences in the history of hypertension, the NIHSS score, the ASPECT score at admission, and the incidence of good revascularization between good and poor prognosis groups (all p < 0.05), as seen in Table 1. In the training set, there were significant differences in the history of hypertension, the NIHSS, and the ASPECT score at admission (all p < 0.05, Table 2).

In the training set, the RSD values of the Bayes, LR, Tree, and Forest algorithms were 13.33, 11.54, 12.21, and 11.93, respectively. In the testing set, the RSD values for these models were 20.95, 18.50, 18.99, and 21.79, respectively. The LR algorithm showed better diagnostic performance and stability than the other machine learning algorithms in the training and test sets (Figure 3).

After standardization and dimension reduction, the four most valuable texture features were selected for the construction of the radiomics signature, including Range, Correlation_angle45_offset1, SurfaceVolumeRati, and VolumeMM. The main process of dimension reduction and the formula for the radiomics signature are provided in the Supplementary material. Based on the formula, the rad-scores were calculated, and it had favorable predictive efficacy in the training and testing cohorts (the AUC values were 0.863 and 0.840, respectively). The Hosmer–Lemeshow test revealed good goodness-of-fit of the radiomics model (all P > 0.05), and the calibration curves demonstrated good consistency in both the training and testing sets. The relevant results are shown in Supplementary Figure S2.

After stepwise LR, the NIHSS score at admission, hypertension, and rad-scores were used to build the joint model, as shown in Table 3. The AUCs of the joint model, rad-scores, NIHSS score, and hypertension were 0.900, 0.863, 0.727, and 0.591 in the training set, respectively. In the testing set, the AUCs of the joint model, rad-scores, NIHSS score, and hypertension were 0.885, 0.840, 0.721, and 0.590, respectively (Table 4; Figure 4).

There was a negative correlation between mRS and surface volume ratio (r = −0.531, p < 0.001). However, the mRS was positively correlated with rad-scores (r = 0.570, p < 0.001), range (r = 192, p = 0.022), and volume MM (r = 0.510, p < 0.001), as shown in Table 5 and Figure 5.

The relationships among clinical factors (NIHSS, hypertension), imaging biomarkers, and clinical outcomes were shown in the Sankey diagram (Figure 6). Each of the three imaging biomarkers that were correlated with mRS was divided into high (H) and low (L) based on their median values. The Sankey diagram shows that most subjects with imaging LHL characteristics, which indicate a low range, high surface volume ratio, and low volume MM, have a good prognosis, whereas subjects with imaging-HLH phenotype, which denotes high range, low surface volume ratio, and high volume MM, have a poorer prognosis. In addition, subjects with higher NIHSS scores and hypertension have a poor prognosis.