Streamlined 2D multimodal fusion framework predicts 90 day functional outcomes in acute ischemic stroke

Summary

Functional outcome prediction after endovascular therapy can support early risk stratification in acute ischemic stroke. We developed a 2D multimodal fusion framework using multicenter data from 405 patients with post-interventional cerebral hyperdensity on immediate non-contrast CT. Radiomic and deep transfer learning features were extracted from the single CT slice showing the largest hyperdensity area and combined with clinical variables using an Extra Trees classifier. In the external test cohort, the fusion model predicted unfavorable 90-day outcome, defined as a modified Rankin Scale score greater than 2, with an AUC of 0.803, specificity of 88.0%, positive predictive value of 93.3%, and sensitivity of 60.9%. SHAP analysis identified malignant cerebral edema, age, baseline NIHSS, reperfusion status, and imaging intensity descriptors as influential factors. By reducing dependence on 3D segmentation, this framework provides an early phase (within 5 days) prognostic assessment tool for individualized in-hospital risk stratification after thrombectomy.

Graphical abstract

Highlights

• •2D fusion model predicts 90-day outcomes in acute ischemic stroke post-thrombectomy
• •Integrates clinical, radiomic, and deep transfer learning features from a single CT
• •Bypasses 3D segmentation, enabling rapid automated multi-modal prognostic assessment
• •SHAP analysis links AI features with malignant cerebral edema and reperfusion status

Health sciences

Introduction

Endovascular therapy (EVT) is the recommended standard of care for acute ischemic stroke (AIS) caused by large vessel occlusion (LVO).^1,2 However, achieving technical success does not universally translate to functional independence (modified Rankin Scale [mRS] ≤2).^3,4 Given the significant burden of long-term disability, it is essential to accurately identify patients who are at high risk for adverse outcomes. This identification is crucial for tailoring early-phase (within 5 days) post-EVT management and providing informed prognostic counseling.

Immediate post-operative non-contrast computed tomography (NCCT) is essential for the detection of post-interventional cerebral hyperdensities (PCHD), which serve as a direct surrogate marker for blood-brain barrier (BBB) disruption.⁵ PCHD signifies a shift from ischemic injury to reperfusion injury, during which the integrity of the microvasculature is compromised. This disruption leads to vasogenic edema and creates the underlying conditions necessary for hemorrhagic transformation (HT).^6,7 Furthermore, the microcirculatory disturbances associated with extravasation can undermine the advantages of large-vessel recanalization.⁸ As a crucial connection between primary injuries and subsequent complications, PCHD has been established as an independent risk factor contributing to unfavorable functional outcomes.⁹ However, its prognostic value is often underutilized because of the absence of quantitative methods.

Although radiomics (RAD) and deep transfer learning (DTL) are proficient at extracting high-dimensional prognostic signatures that extend beyond mere visual assessment,¹⁰ applying these tools to acute stroke is hindered by the labor-intensive manual segmentation required for traditional volumetric analysis.¹¹ Consequently, the majority of existing 90-day prognostic models heavily rely on baseline clinical data, resulting in the underutilization of the detailed pathological data captured in immediate post-procedural scans.¹²

Recognizing the need for clinically deployable prognostic tools during the early-phase (within 5 days) following EVT, we developed a 2D multimodal fusion framework to predict unfavorable 90-day functional outcomes (mRS > 2). Radiomic and DTL signatures were extracted exclusively from the single immediate post-operative NCCT slice demonstrating the largest PCHD area, thereby avoiding the complexity of conventional 3D volumetric segmentation. These baseline imaging signatures were subsequently integrated with early follow-up clinical events, including follow-up computed tomography (CT) and clinical evaluation within 5 days, using an Extra Trees classifier, while SHapley Additive exPlanations (SHAP) were employed to enhance model interpretability. Collectively, this framework provides a clinically applicable tool for individualized prognostic assessment during the early-phase (within 5 days) period.

Results

Patient characteristics across cohorts

To explore the combined prognostic significance of imaging features and clinical metrics, we conducted a retrospective study involving 405 patients from three different centers (Figure 1). The training and validation cohorts, derived from center A, included 248 and 63 patients, respectively, while an external test cohort from centers B and C comprised 94 patients. Overall, unfavorable 90-day functional outcomes (mRS > 2) were observed in 225 patients (55.6%). Clinical and imaging characteristics for all cohorts are summarized in Table 1. Major clinical indicators were comparable between the training and validation cohorts. In contrast, the external test cohort exhibited greater clinical heterogeneity, reflecting the inherent variance encountered in real-world multicenter practice, with patients being older, presenting with lower baseline Alberta Stroke Program Early CT Score (ASPECTS), and experiencing longer times from last known normal. The proportion of unfavorable functional outcomes was notably greater in the test cohort compared to the training and validation cohorts (73.4% vs. ∼50%, p < 0.001).

Figure 1.

Flowchart of patient selection

NCCT, non-contrast CT; PCHD, post-interventional cerebral hyperdensities

Table 1.

Clinical characteristics of patients in the training, validation, and test cohorts

Characteristics	Training cohort (n = 248)	Validation cohort (n = 63)	Test cohort (n = 94)	p value
Age (years)	69 (58–79)	69 (58–79)	75 (67–82)	0.008
Baseline NIHSS score	17 (12–21)	15 (11–18)	16 (12–21)	0.198
ASPECTS	9 (8–10)	9 (8–10)	8 (8–9)	0.014
Baseline blood glucose (mmol/L)	7 (6–8)	7 (6–9)	8 (6–10)	0.355
Onset to puncture time (min)	301 (240–425)	245 (180–332)	341 (242–478)	<0.001
Number of passes	1 (1–3)	2 (1–3)	1 (1–2)	0.818
Poor functional outcome (90-day mRS > 2), n (%)	124 (50.0%)	32 (50.8%)	69 (73.4%)	<0.001
Male sex, n (%)	159 (64.1%)	42 (66.7%)	50 (53.2%)	0.126
Successful recanalization (mTICI ≥ 2b), n (%)	227 (91.5%)	52 (82.5%)	84 (89.4%)	0.112
MCE, n (%)	53 (21.4%)	13 (20.6%)	23 (24.5%)	0.795
HT, n (%)	172 (69.4%)	40 (63.5%)	72 (76.6%)	0.195
Anterior circulation occlusion, n (%)	230 (92.7%)	57 (90.5%)	78 (83.0%)	0.026
Hypertension, n (%)	148 (59.7%)	36 (57.1%)	51 (54.3%)	0.655
Diabetes mellitus, n (%)	42 (16.9%)	11 (17.5%)	16 (17.0%)	0.995
Atrial fibrillation, n (%)	99 (39.9%)	24 (38.1%)	37 (39.4%)	0.965
Hyperlipidemia, n (%)	25 (10.1%)	7 (11.1%)	4 (4.3%)	0.191
Coronary artery disease, n (%)	33 (13.3%)	5 (7.9%)	4 (4.3%)	0.039
Alcohol consumption, n (%)	49 (19.8%)	17 (27.0%)	11 (11.7%)	0.051
Current smoking, n (%)	54 (21.8%)	16 (25.4%)	12 (12.8%)	0.098
Intravenous thrombolysis, n (%)	86 (34.7%)	19 (30.2%)	31 (33.0%)	0.787
Use of aspiration catheter, n (%)	103 (41.5%)	32 (50.8%)	18 (19.1%)	<0.001

Data are presented as median (interquartile range, IQR) for continuous variables and as number (percentage, %) for categorical variables. p values indicate the statistical significance of differences across the cohorts.

ASPECTS, Alberta Stroke Program Early CT Score; HT, hemorrhagic transformation; MCE, malignant cerebral edema (defined as a midline shift ≥ 5 mm); mRS, modified Rankin Scale; mTICI, modified thrombolysis in cerebral infarction; NIHSS, National Institutes of Health Stroke Scale.

Comprehensive performance evaluation of the predictive models

In the training cohort, the multimodal fusion model demonstrated a robust performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.900 (95% confidence interval [CI]: 0.863–0.938). In contrast, the AUCs for the individual clinical, DTL, and RAD models were 0.922, 0.806, and 0.774, respectively. When evaluated in the internal validation cohort, the fusion model maintained its strong performance, yielding an AUC of 0.909 (95% CI: 0.836–0.982). Additionally, it exhibited an overall accuracy of 82.5%, with a sensitivity of 75.0%, specificity of 90.3%, and a precision (positive predictive value, PPV) of 88.9%. In comparison, the standalone clinical, RAD, and DTL models recorded AUCs of 0.904, 0.546, and 0.592 in the validation cohort, respectively (Table 2).

Table 2.

Performance metrics of the models across different cohorts

Model	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV	Cohort
Fusion	0.835	0.900	0.863–0.938	0.871	0.798	0.812	0.861	training
RAD	0.726	0.774	0.717–0.832	0.621	0.831	0.786	0.687	training
DTL	0.714	0.806	0.754–0.859	0.508	0.919	0.863	0.651	training
Clinical	0.859	0.922	0.891–0.954	0.887	0.831	0.840	0.880	training
Fusion	0.825	0.909	0.836–0.982	0.750	0.903	0.889	0.778	validation
RAD	0.508	0.546	0.401–0.691	0.156	0.871	0.556	0.500	validation
DTL	0.603	0.592	0.449–0.734	0.375	0.839	0.706	0.565	validation
Clinical	0.762	0.904	0.830–0.978	0.562	0.968	0.947	0.682	validation
Fusion	0.681	0.803	0.715–0.892	0.609	0.880	0.933	0.449	testing
RAD	0.606	0.663	0.535–0.791	0.594	0.640	0.820	0.364	testing
DTL	0.638	0.632	0.504–0.759	0.638	0.640	0.830	0.390	testing
Clinical	0.660	0.738	0.615–0.861	0.681	0.600	0.825	0.405	testing

AUC, area under the receiver operating characteristic curve; CI, confidence interval; DTL, deep transfer learning; NPV, negative predictive value; PPV, positive predictive value; RAD, radiomics

In the external test cohort, the fusion model demonstrated an AUC of 0.803 (95% CI: 0.715–0.892). In comparison, the AUCs for the standalone clinical, RAD, and DTL models were 0.738, 0.663, and 0.632, respectively. Although DeLong’s test (p = 0.200) and continuous integrated discrimination improvement (IDI) (p = 0.070) were non-significant, a significant categorical net reclassification improvement (NRI).

NRI (20.0%, p <0.001) demonstrated that integrating imaging signatures successfully reclassified 20% of patients misclassified by clinical variables alone. The fusion model achieved a sensitivity of 60.9%, while yielding a specificity of 88.0% and a high PPV of 93.3% (Table 2).

The fusion model demonstrated strong reliability, with calibration curves (Figure 2B) showing close agreement between predicted and observed risks. Furthermore, decision curve analysis (Figure 2C) confirmed its superior clinical utility, providing a consistently higher net benefit than single-modal approaches across clinically relevant threshold probabilities. To further assess specific predictive capabilities, we compared the fusion model against the three single-modal models (RAD, DTL, and clinical) within the test cohort (n = 94). The confusion matrices (Figure 2D) revealed that the clinical, RAD, and DTL models produced 10, 9, and 9 false positives (FPs) when predicting poor functional outcomes. In contrast, the fusion model effectively identified 22 out of 25 true negative cases (patients with a favorable prognosis, mRS 0–2), resulting in only 3 FPs. As visually highlighted in the prediction reliability profile (Figure 2E), the single-modal models plateaued at specificities ranging from 60.0% to 64.0% and PPV values between 82.0% and 83.0%. In comparison, the fusion model achieved a superior specificity of 88.0% and a PPV of 93.3%.

Figure 2.

Comprehensive performance evaluation of the predictive models in the external test cohort

(A) Receiver operating characteristic (ROC) curves comparing the discriminative ability of the fusion, clinical, DTL, and RAD models for predicting 90-day unfavorable outcomes.

(B) Calibration curves assessing the agreement between the predicted probabilities and the actual observed frequencies.

(C) Decision curve analysis (DCA) illustrating the net clinical benefit of the models across a range of threshold probabilities.

(D) Confusion matrices detailing the prediction results of the single-modal and fusion models based on the optimized decision threshold.

(E) Bar plots comparing model reliability metrics, demonstrating that the fusion model achieves the highest specificity (88.0%) and precision (positive predictive value, 93.3%). DTL, deep transfer learning; RAD, radiomics.

Feature importance and interpretability of the predictive model

To elucidate the decision-making process of the predictive model, we examined both the global feature weights and the detailed SHAP values (Figure 3). Among the high-dimensional imaging markers extracted, DTL_AbsSum and Rad_Mean_of_Squares were identified as the most significant DTL and RAD features, respectively. The global feature importance analysis indicated that malignant cerebral edema (MCE), age, baseline National Institutes of Health Stroke Scale (NIHSS), DTL_AbsSum, and HUmax value over 90 (HUmax90) were the five most influential variables driving the model’s predictions. The SHAP summary plot further quantified the directional impact of these features on the 90-day prognosis. Specifically, the occurrence of MCE, along with higher values of age, baseline NIHSS, DTL_AbsSum, HUmax90, and Rad_Mean_of_Squares, produced positive SHAP values, substantially contributing to the prediction of unfavorable outcomes. Conversely, successful reperfusion (modified thrombolysis in cerebral infarction, mTICI ≥ 2b), anterior circulation stroke, and higher ASPECTS scores were associated with negative SHAP values, indicating a protective effect and steering the model toward predicting favorable outcomes.

Figure 3.

Feature importance and interpretability of the predictive model

(A) Global feature weights ranking the most influential variables driving the model’s predictions.

(B) SHAP (SHapley Additive exPlanations) summary plot demonstrating the directional impact of individual features on the 90-day prognosis. Each dot represents a single patient. Red and blue indicate high and low feature values, respectively. Positive SHAP values correspond to an increased risk of an unfavorable functional outcome (mRS > 2). SHAP, SHapley Additive exPlanations; MCE, malignant cerebral edema; NIHSS, National Institutes of Health Stroke Scale; HT, hemorrhagic transformation; mTICI, modified thrombolysis in cerebral infarction; ASPECTS, Alberta Stroke Program Early CT Score; DTL, deep transfer learning; RAD, radiomics.

Discussion

In this multicenter study, we present a streamlined 2D multimodal model for early-phase (within 5 days) prognostic assessment of unfavorable 90-day outcomes following EVT. By integrating clinical, radiomic, and DTL signatures, the model achieved high discriminative performance, yielding an AUC of 0.803, characterized by a specificity of 88.0% and a high PPV of 93.3% in external validation. The framework was intended for prognostic assessment during the early-phase (within 5 days after EVT), integrating baseline (immediate post-op NCCT) imaging signatures with early follow-up (follow-up CT and clinical evaluation) clinical events. Importantly, we acknowledge that the model’s primary value lies in in-hospital risk stratification, rather than serving as a point-of-care decision tool at the time of discharge or immediately after the procedure. The higher proportion of poor outcomes in the external cohort compared to the training and validation cohorts (73.4% vs. ∼50%, p < 0.001) likely reflects increased clinical severity, including older age, lower ASPECTS, and delayed treatment. Furthermore, it is important to note that positive and negative predictive values (PPV and NPV) are inherently prevalence-dependent metrics. Consequently, the high PPV (93.3%) and relatively low NPV (44.9%) observed during external testing are partially attributable to this prevalence shift. When interpreting the model’s clinical utility across different clinical settings, it should be recognized that these predictive values will naturally fluctuate according to the local baseline prevalence of unfavorable outcomes. Specifically, in high-severity populations where poor outcomes are highly prevalent, the model excels at confidently “ruling in” high-risk patients, whereas its ability to “rule out” such outcomes (NPV) may be inherently constrained. The multimodal fusion framework demonstrated generally preserved performance across cohorts, with some attenuation from internal validation to external testing while maintaining good discriminative capability.

Conventional imaging-based pipelines typically require manual region of interest (ROI) delineation, a process that is labor-intensive, time-consuming, and prone to significant inter-observer variability.¹¹ In contrast, our approach centers on the single NCCT slice that exhibits the largest PCHD area and completely eliminates the need for manual lesion contouring through an automated whole-slice masking framework. This method captures the most representative pathological signal, facilitating the extraction of radiomic and deep learning features while also minimizing computational complexity.¹³ By bypassing both the burden of 3D volumetric segmentation and the subjectivity of 2D regional delineation, this workflow ensures that prognostic features are comprehensively captured at the maximum level of pathological signal alteration. In alignment with emerging literature advocating for automated, image-level deep learning assessments in acute stroke,¹⁴ our approach transforms complex radiomic analysis into a highly efficient and clinically deployable tool.

Our findings underscore the importance of PCHD as a vital imaging biomarker for long-term prognosis,¹⁵ reflecting the acute collapse of the neurovascular unit and subsequent BBB failure.^16,17 Following reperfusion, the extravasated contrast becomes a neurotoxic agent that induces vasogenic edema and leads to progressive tissue destruction.^2,18,19 By predisposing patients to microcirculatory “no-reflow” and HT,^15,20 these complications can severely undermine brain structural integrity and neutralize the clinical benefits of successful thrombectomy.²¹ To non-invasively quantify the microscopic complexity of secondary injury, we incorporated radiomic analysis into our framework.^21,22 Notably, our SHAP analysis revealed that Rad_Mean_of_Squares stands out as the primary radiomic predictor. This finding suggests a strong correlation between the total energy and density of contrast extravasation and irreversible tissue damage. Our results highlight the distinctive ability of RAD to detect subtle pixel-level variations in BBB damage, which often go unnoticed in standard visual assessment.^19,22

To complement the stability of RAD, our model integrates DTL using a ResNet-152 architecture to capture high-dimensional prognostic patterns within PCHD. While Liu et al. and Wang et al. have demonstrated the capacity of deep learning to extract imperceptible semantic representations from baseline scans, our framework utilizes DTL features specifically extracted from post-operative NCCT slices that demonstrate the maximal extent of PCHD.^23,24 By leveraging the hierarchical architecture of ResNet-152, the model captures global contextual signatures that reflect the overall signal intensity and heterogeneity of contrast extravasation within the post-interventional tissue environment. SHAP analysis supported this synergy by identifying DTL_AbsSum as a key contributor to fine-grained prognostic adjustments in complex cases. Ultimately, the integration of explicit radiomic intensity and abstract DTL patterns enhances the model’s sensitivity to complex pathological landscapes, facilitating a more accurate and comprehensive prediction of 90-day functional recovery in AIS-LVO patients.

To improve clinical interpretability, we utilized SHAP analysis and feature coefficient mapping to evaluate the contribution of each variable. In accordance with the findings of Liu et al. and Huang et al., baseline NIHSS and age emerged as the most significant clinical predictors of 90-day disability.^20,23 HT and MCE are essential indicators of significant parenchymal injury. Specifically, HT denotes severe BBB breakdown and neurotoxic inflammatory cascades, which are the primary biochemical drivers of progressive tissue necrosis and an unfavorable prognosis.²⁵ Concurrently, the severe mass effect of MCE serves as a primary mechanical driver for severe neurological deterioration and poor recovery.²⁶ Conversely, achieving successful reperfusion along with higher ASPECTS scores emerged as significant protective factors. Feature coefficient mapping indicated that mTICI had the most substantial negative weight; underscoring that effective macrovascular recanalization is essential for a positive recovery outcome. Additionally, higher ASPECTS scores provided a protective advantage, demonstrating the benefits associated with a smaller initial infarct core.^27,28 Together, these findings suggest that the model effectively combines systemic clinical status with advanced imaging-derived markers (DTL and RAD) to assess both baseline vulnerability and acute tissue-level injury.

Limitations of the study

Despite the encouraging results, this study has several limitations. First, as a retrospective analysis, it is subject to inherent selection bias. Second, while selecting the single 2D slice with the maximal PCHD area offers a highly practical and computationally efficient workflow, this simplified “representative slice” approach may sacrifice the capture of overall lesion heterogeneity. Because PCHD lesions can present with irregular 3D geometries, the slice demonstrating the largest area does not necessarily correlate with total lesion volume or the region of maximum pathological severity. Although high inter-reader reproducibility was confirmed via intraclass correlation coefficient (ICC) analysis, single-slice evaluation may underrepresent spatially distributed micro-lesions, which likely contributes to the model’s moderate sensitivity (60.9%). Consequently, the model should be interpreted cautiously as an adjunctive risk-stratification tool rather than a standalone decision-making tool. Future prospective studies should directly compare feature stability and predictive performance between maximal 2D slice selection and whole-lesion 3D volumetric analyses. Third, although interpretable radiomic features are included, the deep learning aspect of the fusion model operates as a “black box.” The abstract hierarchical representations generated by the neural network lack direct clinico-pathological correlation, potentially hindering clinicians’ intuitive understanding of the model’s decision-making process. Fourth, although the decision threshold was lowered, the calibration curve shows that the fusion model curve lies overall above the perfect prediction line. This indicates that the model may systematically underestimate the probability of 90-day unfavorable outcomes, a factor that clinicians must carefully consider when applying these risk estimates in practice.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Yougen Cheng (zjtdcyg@126.com).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •De-identified clinical and imaging data reported in this paper are not publicly deposited because they contain patient-level medical imaging and clinical information and are subject to patient privacy protection, institutional ethics approvals, and data-use restrictions. Requests for access to the de-identified data may be directed to the lead contact with a reasonable scientific rationale.
• •All original code required reproducing the implementation of the 2D multimodal fusion framework, model construction, performance evaluation, and SHAP analysis is publicly available at GitHub: https://github.com/ChamnLee/Comp2_MCC.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request.

Acknowledgments

This work was supported by the Key Discipline of Jiaxing (grant no. 2023-ZC-015).

Author contributions

S.H., Y. Chen, J.D., D.T., and Y. Cheng conceived and designed the study; S.H., Y. Chen, and F.L. led the overall study implementation, including data organization, image processing, RAD analysis, model development, statistical analysis, figure preparation, and manuscript drafting; J.D., Y. Chen, F.L., J.H., X.Z., Y.L., T.L., H.L., S.G., and S.C. contributed to clinical and imaging data collection and curation; S.H. and J.F. contributed to methodological guidance, model interpretation, and critical review of the analyses; D.T. and Y. Cheng supervised the study, provided clinical guidance, and acquired funding. All authors reviewed, edited, and approved the final manuscript.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this manuscript, the authors used ChatGPT (OpenAI) to assist with language refinement and editing. The authors reviewed and edited the manuscript after using this tool and take full responsibility for the content of the published article. All ideas, methodologies, analyses, results, and conclusions were developed, verified, and approved by the authors.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
De-identified patient clinical and imaging data	This paper	Restricted access; requests may be directed to the lead contact, Yougen Cheng (zjtdcyg@126.com), with a reasonable scientific rationale and are subject to institutional ethics approvals and data-use restrictions
Original analysis code	This paper	GitHub: https://github.com/ChamnLee/Comp2_MCC
Software and algorithms
PyRadiomics	PyRadiomics developers	https://pyradiomics.readthedocs.io
torchvision	PyTorch developers	https://pytorch.org/vision
scikit-learn	scikit-learn developers	https://scikit-learn.org
SHAP	SHAP developers	https://shap.readthedocs.io

Experimental model and study participant details

Study design and patient cohorts

This retrospective multicenter human-participant study was conducted across three institutions and was not a clinical trial. Ethical approval was obtained from the respective ethics committees, which waived the requirement for informed consent: the Fourth Affiliated Hospital of Zhejiang University (Center A, Approval No. K2024139), The First Hospital of Jiaxing and The Affiliated Hospital of Jiaxing University (Center B, Approval No. 2026-KY-268), and Tongde Hospital of Zhejiang Province (Center C, Approval No. KTSC2026291).

The study included 405 patients with acute ischemic stroke treated with EVT and post-interventional cerebral hyperdensity on postoperative NCCT. We enrolled 311 patients from Center A between July 2016 and January 2024. Of these, 248 patients (80%) were randomly assigned to the training cohort, while 63 patients (20%) comprised the internal validation cohort. To create an external test cohort, we included an additional 94 patients from Center B (n=81) and Center C (n=13), who were treated from November 2018 to January 2024. Thus, the final study population included 405 patients. The median age was 69 years (IQR 58–79) in the training cohort, 69 years (IQR 58–79) in the internal validation cohort, and 75 years (IQR 67–82) in the external test cohort. Male patients accounted for 159/248 (64.1%) in the training cohort, 42/63 (66.7%) in the internal validation cohort, and 50/94 (53.2%) in the external test cohort. Sex was included as a candidate clinical variable in model development, and sex distribution did not differ significantly across cohorts (p=0.126; Table 1). Sex-stratified and sex-interaction analyses were not performed. Gender, ancestry, race, and ethnicity were not routinely recorded in the retrospective datasets. No animals, plants, microbial strains, cell lines, or primary cell cultures were used in this study; therefore, cell-line authentication and mycoplasma testing were not applicable.

Inclusion criteria mandated that patients undergo a non-contrast CT (NCCT) within 1 hour following EVT. We excluded patients based on several criteria: (1) missing 90-day modified Rankin Scale (mRS) scores; (2) unsuccessful EVT, such as failed access to the occlusion site or spontaneous recanalization confirmed by angiography; (3) lack of post-interventional cerebral hyperdensity (PCHD) on follow-up imaging; (4) image artifacts that hindered reliable PCHD assessment; and (5) initial NCCT performed more than 1 hour after EVT. All participants received clinical care for acute ischemic stroke with EVT and postoperative/follow-up imaging as part of routine clinical management.

Method details

Clinical and imaging data collection

Trained clinicians meticulously extracted clinical characteristics and surgical details from medical records, addressing any discrepancies through collaborative consultation. The collected variables encompassed age, sex, hyperlipidemia, hypertension, diabetes mellitus, coronary artery disease, atrial fibrillation, history of alcohol consumption or smoking, baseline National Institutes of Health Stroke Scale (NIHSS) scores, baseline blood glucose levels, thrombolysis status, stent implantation details, modified Thrombolysis in Cerebral Infarction (mTICI) scores, the use of aspiration catheters, MCE (defined as a midline shift ≥ 5mm), the TOAST classification, onset to puncture time (OPT), and the number of passes. Baseline blood glucose was the sole variable with missing data (missing in 81 out of 405 patients, 20.0%), which we imputed using Multiple Imputation by Chained Equations (MICE) with 20 imputations. We assessed functional outcomes at 90 days using the mRS (range: 0–6), where we defined unfavorable outcomes as mRS scores ranging from 3 to 6.

For imaging metrics, we obtained initial postoperative head CT scans within 1-h post-thrombectomy utilizing spiral CT scanners (Table S3). Two senior neuroradiologists, each possessing over a decade of experience and blinded to the clinical data, independently evaluated the scans in a randomized order. They assessed various metrics, including the Alberta Stroke Program Early CT Score (ASPECTS) for both anterior and posterior circulation, subarachnoid PCHD, and HUmax values exceeding 90. Any discrepancies in interpretation were addressed through a consensus discussion. HT and MCE were assessed during the early-phase (within 5 days) post-EVT period based on follow-up imaging and clinical evaluation within 5 days after EVT.

2D image selection and radiomics feature extraction

To enhance the analysis pipeline, neuroradiologists visually identified and selected a single axial CT slice exhibiting the maximum PCHD area for each patient. Instead of manual lesion contouring, an automatically generated whole-slice mask that matched the exact pixel dimensions of the selected 2D slice was applied. This streamlined slice-selection process required only about 1 minute per case, and the entire 2D slice was subsequently used as the primary input for both radiomics and DTL feature extraction.

To maintain spatial consistency across CT scanners, all images were resampled to a 1×1 mm² resolution and gray-scale discretization was performed with a fixed bin width of 5. Using PyRadiomics, we extracted a comprehensive set of 1,834 radiomic features from the entire selected slice, encompassing 2D shape, first-order statistics, texture, and wavelet transformations. After excluding 14 geometrically constant shape features, an interobserver agreement analysis on a 50-case subset evaluating the reproducibility of the visual slice-selection step confirmed the high stability of this workflow, allowing us to retain 1,347 robust features with an intraclass correlation coefficient greater than 0.85 for subsequent analysis. To further refine the extracted high-dimensional feature profiles, we computed seven statistical descriptors for the features, specifically the sum, mean, median, absolute sum, L2 norm, mean of squares, and median of squares.

Deep transfer learning (DTL) extraction

We identified ResNet152 as the most suitable backbone among the five convolutional neural network architectures evaluated. The model was initialized using pre-trained weights from ImageNet and was fine-tuned specifically on the selected 2D NCCT slice that exhibited the largest PCHD area. To improve model generalization, we implemented data augmentation techniques, including RandomResizedCrop and RandomHorizontalFlip. The network was trained with the Adam optimizer, utilizing a learning rate of 1×10^-5 over the course of 50 epochs. High-dimensional DTL features were extracted from the 'avgpool’ layer. Subsequently, we applied Principal Component Analysis (PCA) to condense the feature space to 303 dimensions, effectively removing redundancy. In alignment with the radiomics pipeline, we computed seven statistical descriptors to characterize the distribution of DTL features: sum, absolute sum, mean, median, L2 norm, mean of squares, and median of squares.

Quantification and statistical analysis

Continuous variables were summarized as median (IQR; center = median, dispersion = IQR) and compared using Kruskal-Wallis tests. Categorical variables were summarized as n (%) and compared using chi-square or Fisher’s exact tests as appropriate. All tests were two-sided, and p < 0.05 was considered statistically significant. Exact n values represent patients unless otherwise specified; cohort sizes, the number of events, statistical tests, and summary measures are reported in Table 1, and model-performance sample sizes and metrics are reported in Figure 2; Tables 2 and 3 the results, and the relevant figure legends.

Table 3.

Performance comparison of the five machine learning algorithms in the internal validation cohort

Model	Accuracy	AUC	95% CI	Sensitivity	Specificity	PPV	NPV
LR	0.889	0.863	0.819–0.905	0.875	0.903	0.903	0.875
SVM	0.905	0.853	0.819–0.889	0.875	0.935	0.933	0.879
KNN	0.746	0.886	0.840–0.924	0.531	0.968	0.944	0.667
ExtraTrees	0.825	0.909	0.836–0.982	0.750	0.903	0.889	0.778
MLP	0.889	0.844	0.807–0.890	0.844	0.935	0.931	0.853

AUC, area under the receiver operating characteristic curve; CI, confidence interval; KNN, K-nearest neighbors; LR, logistic regression; MLP, multilayer perceptron; NPV, negative predictive value; PPV, positive predictive value; SVM, support vector machine.

The analysis workflow used PyRadiomics for radiomic feature extraction; Python-based machine-learning software, including PyTorch/torchvision for ResNet-152 transfer learning and image augmentation and scikit-learn for preprocessing, PCA, LASSO feature selection, and model construction/evaluation; MICE with 20 imputations for missing baseline blood glucose; and SHAP for model interpretability. Software versions and executable code are available in the public GitHub repository listed in Data and code availability.

Multimodal model construction

As shown in Figure 4, we established a multimodal prognostic framework that integrates 19 clinical variables, 7 radiomics features, and 7 deep transfer learning (DTL) statistical descriptors (detailed in Table S2). To ensure consistency, we standardized all features using Z score normalization and applied Pearson correlation analysis to remove redundant variables (correlation coefficients > 0.9). Subsequently, we applied LASSO regression with 10-fold cross-validation for feature selection, determining the optimal penalty parameter (λ) via minimum binomial deviance. Using these selected features, we evaluated five machine learning algorithms and ultimately chose the Extra Trees classifier for the final fusion model, given its superior capability in handling high-dimensional data during internal validation (Table 3). Crucially, the external test cohort was strictly locked throughout the feature selection, hyperparameter tuning, and algorithm selection phases, and was utilized exclusively for the final independent evaluation. We assessed the model’s overall efficacy and clinical utility through Receiver Operating Characteristic (ROC) curves, Decision Curve Analysis (DCA), calibration curves, and confusion matrices. Furthermore, to establish a clinically optimal decision boundary, we evaluated the model’s diagnostic metrics across various probability thresholds within the training and internal validation cohorts. The final operational threshold was determined by maximizing the Youden index in the internal validation cohort. This predefined threshold was then applied unchanged to the independent external test cohort. To evaluate the incremental predictive value of the imaging features over the clinical variables, we compared the multimodal fusion model against the standalone clinical model using DeLong’s test, continuous Integrated Discrimination Improvement (IDI), and categorical Net Reclassification Improvement (NRI).

Figure 4.

Overall pipeline of the proposed multimodal fusion framework

The workflow consists of two main stages: feature extraction and multi-model construction. First, patient data across three centers were collected, and 2D non-contrast CT slices with the maximal post-interventional cerebral hyperdensity (PCHD) were selected. High-dimensional signatures were extracted via deep transfer learning (ResNet-152) and radiomics pipelines, then reduced to representative statistical descriptors. Finally, these imaging signatures were integrated with clinical profiles to construct the Extra Trees classifier, followed by performance evaluation (e.g., ROC and DCA) and interpretability analysis (SHAP).

Model interpretability

To elucidate the decision-making process of the integrated Extra Trees classifier, we employed SHapley Additive exPlanations (SHAP) to quantify the contributions of individual features based on cooperative game theory. This analysis combined clinical indicators with the seven statistical descriptors derived from both radiomics and DTL features. We created SHAP beeswarm plots to rank the global importance of features and visualize the directional influence of each variable on the predicted probability of unfavorable functional outcomes.