Multimodal integration of carotid plaque ultrasound radiomics and dynamic systemic immune-inflammation trajectories for subacute risk stratification after stroke thrombectomy: development and internal validation of an exploratory prognostic model

Background We hypothesized that integrating macroscopic carotid plaque architecture, quantified via ultrasound radiomics, with downstream micro-cellular temporal dynamics, defined by longitudinal immune trajectories, would yield a highly accurate prognostic framework for 90-day functional outcomes in patients undergoing endovascular thrombectomy (EVT). Methods This single-center, retrospective cohort study consecutively included 316 AIS-LVO patients who underwent successful EVT (modified Thrombolysis in Cerebral Infarction score ≥2b) within 24 h of symptom onset. Patients were partitioned into a discovery ( N = 221) and an internal validation cohort ( N = 95) using stratified random sampling. To capture temporal inflammatory heterogeneity, the Systemic Immune-Inflammation Index (SII) was measured at days 0, 1, 3, and 7, and clustered using Latent Class Mixed Models (LCMM). Concurrently, high-dimensional radiomic features strictly adhering to the Image Biomarker Standardisation Initiative (IBSI) guidelines were extracted from baseline two-dimensional carotid ultrasound images. A radiomics score (Rad-score) was constructed utilizing Least Absolute Shrinkage and Selection Operator (LASSO) regression. Four progressive predictive models were developed, culminating in an Extreme Gradient Boosting (XGBoost) multimodal architecture. The primary endpoint was an unfavorable 90-day functional outcome (modified Rankin Scale score 3 to 6). Results An unfavorable outcome occurred in 166 of 316 patients (52.5%). The LCMM objectively identified three longitudinal SII phenotypes. The “persistent high-inflammation” trajectory (Class 3, 17.0% of patients) was independently associated with an increased risk of an unfavorable outcome compared to the “rapid resolution” trajectory (Class 1) [adjusted odds ratio (aOR), 4.85; 95% confidence interval [CI], 2.20 to 10.85; p < 0.001]. The ultrasound Rad-score, comprising 9 core textural features, also demonstrated robust independent predictive value (aOR, 1.52; 95% CI, 1.21 to 1.92 per 0.1-point increase; p < 0.001). In the strictly isolated internal validation cohort, the full multimodal XGBoost model achieved an area under the receiver operating characteristic curve (AUC) of 0.92 (95% CI, 0.86 to 0.97), with an accuracy of 86.3%, sensitivity of 88.0%, and specificity of 84.4%. This performance was statistically superior to the standard clinical baseline model (AUC, 0.74; 95% CI, 0.63 to 0.84; p < 0.001 via DeLong test). Furthermore, the multimodal paradigm demonstrated clinical reclassification utility, yielding a continuous net reclassification improvement of 0.985 (95% CI, 0.684 to 1.286) and an integrated discrimination improvement of 0.194 (95% CI, 0.128 to 0.260). Conclusion An exploratory model integrating baseline carotid plaque ultrasound radiomics with Day-0-to-Day-7 systemic immune-inflammation trajectories was associated with 90-day functional outcome after EVT and, on internal validation, discriminated outcomes more strongly than a conventional clinical model. Because the immune-trajectory component is only complete at Day 7, the model is intended for subacute post-EVT risk stratification rather than acute treatment selection, and requires external, prospective validation before any clinical application.