This study was a retrospective analysis of 180 patients diagnosed with acute ischemic stroke at Liangxiang Hospital (Beijing, China) from October 2020 to May 2022, of which 148 were included in analysis (Figure 1). The inclusion criteria were: (1) acute ischemic stroke due to anterior circulation artery occlusion; (2) MRI completed within 48 h of admission; (3) complete set of MRI sequences; (4) complete data on demographics and clinical characteristics; and (5) signed informed consent. The exclusion criteria were: (1) cerebral hemorrhage; (2) traumatic brain injury; (3) previous neurological or psychiatric disease; and (4) significant artifacts in MRI data. This study was approved by the ethics committee of Liangxiang Hospital (approval number 2016126).

Demographic characteristics as well as the clinical and imaging data were collected under the permission of patients. The National Institutes of Health Stroke Scale (NIHSS) score was collected to evaluate the degree of neurological deficit in stroke patients, which represented the level of consciousness, eye movements, integrity of visual fields, facial movements, arm and leg muscle strength, sensation, coordination, language, speech and neglect (10). Arranging from 0 to 42, the higher the NIHSS score, the more severe the neurological impairment: score 0 was normal neurological function, 1 to 4 was mild stroke, 5 to 15 was moderate stroke, 16 to 20 was moderate-severe stroke and 21 to 42 was severe stroke.

The modified Rankin Scale (mRS) score was used as a prognostic judgment index, with good prognosis defined as mRS scores of 0, 1, and 2, and poor prognosis defined as mRS scores of 3, 4, and 5. Of the 148 patients included in the analysis, 83 were in the good prognosis group and 65 were in the poor prognosis group.

MRI scans were performed within 3 days of stroke onset, using a Magnetom Skyra 3.0T MRI scanner (Siemens, Germany) with a 20-channel phased-array head coil.

All participants underwent the following scans:

(1) T1-weighted image scan. Scan parameters: T1w sequence, repetition time (TR) = 2,000 ms, inversion time (TI) = 900 ms, echo time (TE) = 8.8 ms, matrix = 209 × 256, field of view (FOV) = 220 mm² × 196 mm², thickness = 5 mm. number of layers = 24 layers, and parallel imaging factor = 2.

(2) Cerebrospinal fluid suppression image. Scan parameters: T2-FLAIR sequence, TR = 6,000 ms, TI = 2,028 ms, TE = 72 ms, matrix = 320 × 261, FOV = 220 mm² × 196 mm², thickness = 5 mm, number of layers = 24 layers, and parallel imaging factor = 2.

(3) SWI imaging scan sequence. Scan parameters: 3D-GRE sequence, TR = 27 ms, TE = 20 ms, flip angle (FA) = 15°, matrix = 256 × 256, FOV = 220 mm² × 196 mm², layer thickness = 2.5 mm, number of layers = 44, repetition number = 1, fat suppression on, and parallel imaging factor = 2.

(4) DWI imaging scan sequence. Scan parameters: EPI-Resolve sequence, b-value b = 1,000 and 0 scan, TR = 500 ms, TE1 = 63 ms, TE2 = 103 ms, FA = 180°, matrix = 160 × 160, FOV = 220 mm² × 220 mm², layer thickness = 5 mm, number of layers = 24, fat suppression on, and parallel imaging factor = 2.

Image analysis was performed independently by two MR diagnosticians blinded to the groups. The region of interest (ROI) of acute ischemic lesions was manually outlined layer by layer on the DWI images using MRIcron software.¹

For each patient, data including all modalities (ADC, FLAIR, SWI, T1w) were aligned to the DWI images using SPM12 software² so that the outlined lesions could be directly used for texture feature extraction in the different modal images (Figure 2).

For each patient, the MRI data of the five modalities was analyzed by pyradiomics software³ for radiomiscs feature extraction using the recommended settings and steps for MRI data: (1) Images were resampled to 3 mm² × 3 mm² × 3 mm. (2) DWI, FLAIR, SWI, and T1w images are weighted images and thus needed to be numerically standardization. A scale of 100 was used and the binwith was set to 5. Since ADC images are quantitative, no numerical standardization was done and binwith was set to 20. The above settings ensured that the total number of bins is between 16 and 128. (3) Texture feature extraction was performed on the original images and filtered images, where the filter consisted of Laplacian of Gaussian filter based on sigma = 3 and 5 mm, edge enhancement filter, and 8 wavelet transforms (combination of high pass and low pass in three dimensions). (4) Finally, the texture features of radiomics were extracted, including 18 first order, 22 glcm, 16 glrlm, 16 glszm, and 14 gldm features. Thus, a total of 946 features were extracted per image. The shape features of 14 lesion regions were also extracted.

The scikit-learn package⁴ was used for feature selection of the MRI data of the five modalities. The recursive feature elimination (RFE) feature extraction method was used and only the three best features were retained for each modality. For lesion shape features, we used the same method to retain the one best feature. Finally, a total of 16 image features were retained and used to train the ML model with the scikit-learn tool. A total of five methods, including Support Vector Machine (SVM) Classifier, Random Forest (RF) Classifier, Light Gradient Boosting Machine (LightGBM) Classifier, Category Boosting (CatBoost) Classifier, and eXtreme Gradient Boosting (XGBoost) Classifier, were used to build the models. Model performance was evaluated by a five-fold stratified cross-validation process. The evaluation metrics included accuracy, precision, recall, F1 score, receiver operating characteristic (ROC) curve, area under the curve (AUC), precision recall curve.

Statistical analysis was performed using SPSS 21.0 software. The independent samples t-test was used to compare the measurement data. Results with p < 0.05 were considered to be statistically significant differences.

Of the 148 patients, 83 (56.1%) had a good prognosis and 65 (43.9%) had a poor prognosis. The training set comprised 104 patients and the remaining 44 were used to test the ML model. For the overall sample, the mean age was 64.29 years and the mean NIHSS score was 5.59. These two variables differed significantly between groups (p < 0.05). No significant differences were found between patient groups for other baseline clinical characteristics (all p > 0.05, Table 1).

Three best radiomic features for each MRI modality (DWI, ADC, FLAIR, SWI and T1w) and one best feature for lesion shape were selected as features in the ML models. The detailed information about the features is presented in Table 2.

The average results obtained for the test set using the different classification models after five-fold cross-validation are as follows: SVM model with 79% accuracy, RF model with 82% accuracy, LightGBM model with 83% accuracy, CatBoost model with 81% accuracy, and XGBoost model with 80% accuracy. The full model evaluation results are shown in Table 3 and Figures 3, 4.