The study, published in Frontiers in Neurology, followed 316 patients treated at a single Chinese stroke center between 2020 and 2023. All had suffered an acute ischemic stroke caused by a large-vessel occlusion and had undergone successful endovascular thrombectomy within 24 hours of symptom onset, defined as a modified Thrombolysis in Cerebral Infarction grade of at least 2b. At the 90-day mark, 166 of the 316 patients — 52.5 percent — had an unfavorable outcome, scoring 3 to 6 on the modified Rankin Scale, meaning moderate-to-severe disability or death.

This stubborn failure mode, in which the artery is open but the brain does not recover, is known in the field as futile recanalization, and it is the central problem the new model was built to attack. The researchers' central insight was that prognosis after thrombectomy is written in two places at once: upstream, in the architecture of the atherosclerotic plaque that spawned the clot, and downstream, in the systemic immune storm that reperfusion ignites. To read the upstream signal, they turned to radiomics — the automated extraction of hundreds of quantitative texture features from ordinary two-dimensional B-mode carotid ultrasound images.

Features such as gray-level co-occurrence and run-length matrices capture sub-visual patterns of echogenicity and heterogeneity that the human eye cannot grade consistently, and prior work has shown such signatures can identify vulnerable, lipid-rich plaques with an area under the curve of about 0.82. To read the downstream signal, the team measured the Systemic Immune-Inflammation Index — a composite of platelet, neutrophil, and lymphocyte counts — at four time points: admission, day 1, day 3, and day 7. Rather than collapsing this rich temporal data into a single baseline number, they used latent class mixed models to cluster patients by the shape of their inflammatory trajectory.

The statistics pointed decisively to three phenotypes: a rapid-resolution group making up 38 percent of the cohort, a delayed-clearance group peaking at day 3 comprising 45 percent, and a persistent high-inflammation group — 17 percent of patients — whose immune activation never subsided through the first week. The contrast between static and dynamic biomarker assessment was stark. A single baseline SII measurement, the best-performing static marker, achieved an AUC of only 0.71 for predicting poor outcome.

But the trajectories told a far sharper story. Patients in the persistent high-inflammation class had nearly five-fold higher adjusted odds of an unfavorable 90-day outcome compared with the rapid-resolution group — an adjusted odds ratio of 4.85, with a 95 percent confidence interval of 2.20 to 10.85. That class also suffered dramatically more stroke-associated pneumonia, symptomatic intracranial hemorrhage, malignant cerebral edema, and systemic infection, although sensitivity analyses adjusting for these complications preserved a significant, if attenuated, association.

On the imaging side, the pipeline was rigorously standardized. Images were resampled to isotropic 1.0-millimeter spacing, intensity discretization used a fixed bin width of 25 to preserve the absolute relationship between pixel intensity and tissue echogenicity, and features were harmonized across the two ultrasound vendors using the ComBat empirical-Bayes method. After reliability filtering — only features with intra- and inter-observer intraclass correlation coefficients above 0.80 survived — least absolute shrinkage and selection operator regression distilled more than 1,100 candidate features down to nine core textural descriptors.

The resulting radiomics score was independently predictive of poor outcome, with an adjusted odds ratio of 1.52 per 0.1-point increase. Notably, the score correlated moderately with membership in the persistent-inflammation class, with a Spearman coefficient of 0.48, hinting at a link between vulnerable plaque texture and a runaway systemic immune response. The four progressive models told a clear story of incremental value.

A clinical baseline using age, admission stroke severity, and time to puncture reached an AUC of 0.74 in the internal validation cohort. Adding a static baseline SII nudged it to 0.79 — a difference that was not statistically significant. Adding the radiomics score lifted it to 0.83.

Only when the nine radiomic features and the latent trajectory classes were combined in an Extreme Gradient Boosting architecture did performance jump to 0.92, with 95 percent confidence intervals of 0.86 to 0.97, along with 86.3 percent accuracy, 88.0 percent sensitivity, and 84.4 percent specificity. The improvement over the clinical baseline was highly significant by the DeLong test, and the model also outperformed an enhanced clinical-procedural comparator that included collateral status, device passes, and hemorrhagic complications, which itself reached an AUC of 0.81. Reclassification metrics underscored the clinical stakes.

Against the standard clinical baseline, the full multimodal model produced a continuous net reclassification improvement of 0.985 and an integrated discrimination improvement of 0.194, both statistically significant, meaning it moved a substantial fraction of patients into the correct risk category. SHAP explainability analysis ranked the persistent-inflammation trajectory, admission NIHSS score, and specific wavelet-transformed plaque textures as the dominant drivers of individual predictions. The authors also translated the locked model into an interactive dynamic nomogram, allowing a clinician to enter a patient's variables and receive an individualized absolute probability of unfavorable outcome at 90 days.

Yet the researchers are careful — unusually so for a machine-learning paper — about what these numbers do and do not prove. The discovery cohort of 221 patients fell below the calculated minimum sample size of roughly 374, making the model explicitly exploratory. Bootstrap optimism correction trimmed the apparent AUC of 0.92 to a corrected estimate of 0.90, and repeated nested cross-validation yielded a mean AUC of 0.89.

Perhaps most tellingly, a plain penalized logistic regression using the same predictors achieved an AUC of 0.88, statistically indistinguishable from the gradient-boosted model. The real advance, the authors conclude, is the multimodal data itself, not algorithmic sophistication — and a transparent regression may be the better vehicle for future external validation. The framework also faces a structural timing constraint: because the immune trajectory is only complete at day 7, the model cannot guide acute treatment decisions in the catheterization lab.

It is designed for subacute risk stratification — deciding who needs intensive neurocritical monitoring, escalated rehabilitation, or enrollment in immunomodulatory trials during the first post-procedural week. That window matters: several phase III trials of post-thrombectomy anti-inflammatory agents have returned neutral results, and a leading explanation is the failure to target the right patients. A data-driven phenotype that separates rapid resolvers from patients with unremitting inflammation could, in principle, enrich future trials.

But with a retrospective, single-center design, manual image segmentation, and no histological confirmation of the plaque features, the authors state plainly that external, prospective, multicenter validation is required before any clinical use.

Subject of Research: A multimodal prognostic model combining carotid ultrasound radiomics and longitudinal systemic immune-inflammation trajectories for predicting outcomes after stroke thrombectomy

Article Title: 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

Article References: Liu, H., Lu, Z., Xie, L., Miao, F., Li, J., & Chen, H. (2026). 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. Frontiers in Neurology, 17, 1896641. https://doi.org/10.3389/fneur.2026.1896641

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1896641

Keywords: acute ischemic stroke, endovascular thrombectomy, radiomics, carotid plaque, systemic immune-inflammation index, machine learning, XGBoost, latent class mixed models, prognosis, futile recanalization, internal validation, stroke outcomes