This study was approved by the Ethics Committee (Approval Number: K2024139). During the prospective phase, the model was deployed without any clinical interventions or changes to standard care. After the prospective period, data were retrieved from the research database, along with our model’s predictions, for downstream analysis. All clinical investigations were conducted under the principles outlined in the Declaration of Helsinki.

We retrospectively reviewed patients diagnosed with acute ischemic stroke due to intracranial large vessel occlusion (LVO) who underwent endovascular MT between June 2016 and December 2023. The indications and contraindications for MT and thrombolysis were based on the most current guidelines available at the time of treatment. General clinical characteristics, laboratory findings, clinical presentations, and imaging data were collected. Inclusion criteria were as follows: (1) patients underwent head NCCT post-MT; (2) initial postoperative NCCT was performed within 1 h after MT; (3) HIM, defined as hyperattenuation in the brain parenchyma or subarachnoid space, were detected on the initial post-MT NCCT. Exclusion criteria included: (1) incomplete mRS score at discharge; (2) unsuccessful MT, such as guidewire failure to reach the occlusion site or angiography confirming vessel recanalization; (3) absence of HIM; (4) imaging artifacts compromising HIM assessment; (5) administration of iodinated contrast before preoperative CT. A flowchart of the study population is presented in Figure 1.

Initial postoperative non-enhanced head CT scans were acquired within 1 h post-thrombectomy using either a 64-row spiral CT scanner (Somatom® Definition AS, Siemens Healthineers, Forchheim, Germany) or a 62-row spiral scanner (Optima® CT620, GE Medical Systems, Milwaukee, WI, USA). Scanning parameters were as follows: axial mode, tube voltage of 120 kV, tube current of 250–300 mAs, coverage from the skull base to the cranial vertex, a section thickness of 5 mm, and reconstruction with a standard algorithm. Two neuroradiologists with over 10 years of experience and blinded to the clinical data independently evaluated imaging data in randomized order. Any discrepancies were resolved through consensus discussion.

The workflow and global analysis pipeline for the classification model are illustrated in Figure 2. HIM was manually delineated on each CT image, and the region of interest (ROI) encompassing the HIM was extracted for input into the predictive model. To ensure the reliability and reproducibility of radiomics features, 30 lesions were randomly selected, and intra-class correlation coefficients (ICC) were used to assess inter- and intra-observer agreement in feature extraction. Reader A performed ROI segmentation twice, with a 1-month interval, to evaluate intra-observer agreement. Reader B segmented the 30 lesions once, and the extracted radiomics features were used to assess inter-observer agreement. Features with ICC > 0.75 were deemed reliable and retained for model construction, resulting in 1,725 features with excellent agreement (ICC > 0.75). Additionally, segmentation consistency was quantified using the Dice similarity coefficient (DSC), yielding a mean DSC of 0.954, indicating high agreement between the two neuroradiologists. DTL and radiomics features were extracted separately: DTL features were derived from the largest cross-sectional ROI, while radiomics features were computed from the entire ROI volume. Feature selection was conducted sequentially using analysis of variance (ANOVA), Pearson’s correlation coefficient, and the least absolute shrinkage and selection operator (LASSO).

In this study, 303 patients were enrolled from June 2016 to December 2023. Of these, 246 patients treated between June 2016 and January 2023 were allocated to model training and validation: 197 (80%) were randomly assigned to the training cohort, and 49 (20%) to the validation cohort. An additional 57 patients from February 2023 to December 2023 were included in the internal test cohort. Functional outcomes at discharge were evaluated using the mRS (range: 0–6), with unfavorable outcomes defined as mRS scores of 4–6, extracted from medical records by trained clinicians. Discrepancies were resolved through consultation.

By reviewing patient charts and procedure notes, we gathered the following clinical indicators and surgical details: age, sex, hyperlipidemia, hypertension, diabetes mellitus, coronary artery disease, atrial fibrillation, history of prior stroke, history of alcohol consumption or smoking, baseline National Institutes of Health Stroke Scale (NIHSS) scores, length of hospital stay, pre-stroke mRS scores, admission mRS scores, thrombolysis status, stent implantation, modified Thrombolysis in Cerebral Infarction (mTICI) scores, use of aspiration catheters, and number of thrombectomy attempts. Imaging metrics, including the Alberta Stroke Program Early CT Score (ASPECTS) for anterior and posterior circulation, subarachnoid HIM, HUmax values, and other relevant parameters, were determined through consensus between two neuroradiologists, each with over 5 years of experience and blinded to the patients’ clinical information.

Radiomic features were extracted from HIM on NCCT using an in-house feature analysis program based on Pyradiomics¹. To eliminate any potential variations in CT images obtained using different CT scanners, NCCT images were reconstructed using a voxel size of 1 × 1 × 1 mm³ and gray-scale discretization. A total of 1,834 radiomic features were computed, covering shape, first-order statistics, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size zone matrix, neighboring gray-tone difference matrix, gray-level dependence matrix, and wavelet features.

DTL features were extracted using five models: VGG19_bn, Inception v3, ResNet50, VGG16, and ResNet152. Among these, we selected the model with the highest predictive performance for this study, based on its superior area under the curve (AUC) value. Pre-trained on large-scale datasets, the selected model was fine-tuned using the maximum cross-sectional HIM ROI. Data augmentation techniques, including RandomResizedCrop and RandomHorizontalFlip, were applied to improve generalization. Features from the ‘avgpool’ layer generated 25,088 DTL features, which were then reduced to 303 using principal component analysis (PCA). Radiomics features were extracted and evaluated using five machine learning models: SVM, KNN, Random Forest, XGBoost, and LightGBM. Among these, SVM performed best, achieving an AUC of 0.609. As a result, a clinical-radiomics-DTL nomogram was constructed using SVM. Clinical, radiomics, and DTL features were standardized to have a mean of 0 and a variance of 1. Feature selection was performed using ANOVA, Pearson’s correlation coefficient, PCA, and LASSO to enhance the models’ generalization ability and reduce overfitting. For the clinical, radiomics, and DTL models, ANOVA, Pearson’s correlation coefficient, and LASSO were applied sequentially to complete feature selection. PCA, ANOVA, Pearson’s correlation coefficient, and LASSO were used sequentially for the combined model.

To develop the clinical model, we first performed univariate logistic regression analysis to identify clinical variables significantly associated with the outcome. Variables with a p-value less than 0.05 were considered candidate predictors. These selected features were then used to construct the clinical model using a support vector machine (SVM) algorithm, implemented in Python (version 3.11.13). Model performance was evaluated in the training and test cohorts using receiver operating characteristic (ROC) curves and corresponding AUC values.

A nomogram integrating clinical, radiomics, and DTL features was developed, alongside separate clinical, radiomics, and DTL models for comparison. Model performance was assessed using sensitivity, specificity, receiver operating characteristic (ROC) curves, and AUC. Five-fold cross-validation, combined with grid search, was employed to optimize the hyperparameters of the SVM model. The training cohort data were divided into five folds, yielding five prediction models. For each model, sensitivity, specificity, and AUC were computed. The fold with the median AUC was selected as the representative split. Feature engineering was strictly limited to the training data in each fold to prevent data leakage and ensure unbiased evaluation. Models were trained and evaluated on the training, validation, and test cohorts. The Shapley Additive Explanations (SHAP) method was used to visualize and quantify each feature’s contribution to model predictions. Calibration curves were generated to assess the agreement between predicted probabilities and observed outcomes in the test cohort. The DeLong test was applied to compare ROC curve performance across models, and decision curve analysis (DCA) was performed to evaluate the clinical utility of the predictive models.

The network architecture was implemented using Python and the PyTorch library. The DTL model was trained with the stochastic gradient descent algorithm as the optimizer, using a learning rate of 0.001, a mini-batch size of 8, and a binary cross-entropy loss function. Batch normalization was applied after each convolution layer to accelerate convergence and reduce overfitting. The network was trained for 50 epochs.

To assess the consistency and robustness of the nomogram in predicting longer-term outcomes, a sensitivity analysis was conducted in the test cohort using 90-day mRS scores as the outcome measure. Patients with 90-day mRS scores of 3–6 were classified as having unfavorable outcomes.

Statistical analyses were performed using SPSS software (version 27.0, IBM). Continuous variables are reported as means with standard deviations, and categorical variables as frequencies and percentages. The Kruskal–Wallis test was used to compare age, hospital stay duration, baseline NIHSS score, discharge NIHSS score, admission mRS score, and ASPECTS. Other categorical variables were analyzed using the chi-square test or Fisher’s exact test, as appropriate. Feature extraction, selection, and model construction were conducted using Python (version 3.11.13). Model performance was evaluated by calculating the AUC. Calibration curves were generated to assess agreement between predicted probabilities and observed outcomes in the test cohort. The DeLong test compared performance differences between ROC curves of different models, and DCA evaluated the clinical utility of the prediction models. A p-value < 0.05 was considered statistically significant for all tests.

The study included a training cohort of 197 patients, a validation cohort of 49 patients, and a test cohort of 57 patients. Clinical feature selection was performed using ANOVA, Pearson correlation, and LASSO. Table 1 details the patient characteristics. Unfavorable discharge outcomes, defined as mRS 4–6, were observed in 185 of 303 patients. Significant differences among cohorts were identified in hyperlipidemia (p = 0.010), coronary artery disease (p = 0.040), thrombolysis (p = 0.005), stent type, (p = 0.001), and aspiration catheter use (p < 0.001). A nomogram incorporating clinical, radiomics, and DTL features was developed (Figure 3). To derive the DTL features, we evaluated five convolutional neural network models: VGG19_bn, Inception v3, ResNet50, VGG16, and ResNet152. Among these, VGG19_bn demonstrated the highest performance (AUC = 0.674) and was selected for this study.

In the training cohort, the nomogram demonstrated superior performance with an AUC of 0.995 (95% CI: 0.989–1.000), a sensitivity of 0.969, and a specificity of 0.986. The clinical model achieved an AUC of 0.924 (95% CI: 0.889–0.960), while the radiomics model and DTL model yielded AUCs of 0.792 (95% CI: 0.723–0.860) and 0.675 (95% CI: 0.600–0.751), respectively (Figure 4A).

In the validation cohort, the nomogram maintained high performance with an AUC of 0.959 (95% CI: 0.910–1.000), a sensitivity of 0.903, and a specificity of 0.889. The clinical model achieved an AUC of 0.737 (95% CI: 0.597–0.876), while the radiomics model and DTL model yielded AUCs of 0.609 (95% CI: 0.431–0.788) and 0.674 (95% CI: 0.517–0.831), respectively (Figure 4B).

Five independent validation results were obtained, as shown in Table 2. The fold with the median AUC (0.959) was selected to represent a typical data split for model training and evaluation.

In the test cohort, our nomogram achieved an AUC of 0.894 (95% CI: 0.807–0.981), outperforming the clinical model (AUC 0.792, 95% CI: 0.674–0.910), the radiomics model (AUC 0.542, 95% CI: 0.386–0.698), and the DTL model (AUC 0.533, 95% CI: 0.378–0.689) (Figure 4C). DCA further revealed that all models improved outcome predictions compared to no-model scenarios, with the nomogram providing the most significant clinical benefit (Figure 5A). To evaluate the models’ effectiveness, the DeLong test was performed (Figure 5B). The findings indicated that the AUC of the nomogram was markedly superior to that of the radiomics model (p < 0.001) and the DTL model (p < 0.001) within the testing cohort. Moreover, the nomogram displayed a trend toward enhanced performance relative to the clinical model (p = 0.079), though this difference fell short of statistical significance. To assess the agreement between predicted and observed outcomes for the nomogram, calibration curves were generated (Figure 5C).

Figure 6 presents a SHAP bar plot, ranking features by their contribution to the model’s output in descending order, and a SHAP beeswarm plot, illustrating the impact and distribution of each feature on model predictions. The magnitude of SHAP values indicates the strength of each feature’s influence on individual predictions.

We performed a sensitivity analysis in the test cohort using 90-day mRS scores as the outcome measure, despite one patient being lost to follow-up at 90 days. The nomogram achieved an AUC of 0.854, suggesting good discriminatory ability for predicting long-term functional disability.

To enhance the nomogram’s clinical utility, a probability threshold of 0.760 was established by maximizing the Youden Index (0.645) in the test cohort. At this cutoff, the model achieved a sensitivity of 0.778, specificity of 0.867, positive predictive value (PPV) of 0.840, and negative predictive value (NPV) of 0.812, striking an optimal balance between false positives and false negatives. The positive likelihood ratio (PLR) of 5.85 and negative likelihood ratio (NLR) of 0.26 demonstrate robust risk stratification.

Figure 7 illustrates the ROC curves for the nomogram and physicians. The nomogram achieves an AUC of 0.894, outperforming both the senior physician (0.693) and the junior physician (0.600). In comparison to the senior physician, the nomogram demonstrates superior accuracy (83% vs. 53%), specificity (83% vs. 80%), and sensitivity (82% vs. 62%). Similarly, when compared to the junior physician, the nomogram exhibits greater accuracy (83% vs. 53%), specificity (83% vs. 64%), and sensitivity (82% vs. 56%).