Simple Admission Data Struggle to Predict Stroke Deterioration After Clot Removal

When a large stroke-causing clot is mechanically pulled from a patient’s brain, the procedure—endovascular thrombectomy, or EVT—can be the difference between lifelong disability and functional independence. Yet even after the vessel is reopened, a troubling minority of patients deteriorate neurologically within the first two days. A new study from Shanxi Provincial People’s Hospital in Taiyuan, China, published in Frontiers in Neurology, asks a deceptively simple question: can the routine measurements taken when a patient first arrives—the stroke severity score, age, blood glucose, white blood cell count, and triglyceride level—reliably flag who will worsen by 48 hours after the procedure? The answer, according to a rigorously honest statistical reanalysis, is: not yet, and perhaps not with these tools alone.

The research team, led by Guoyong Ren, Chenyang Zhao, and Yaxuan Sun of the Department of Neurology, focused on a specific and clinically meaningful endpoint they call END48: early neurological deterioration defined as an increase of at least four points on the National Institutes of Health Stroke Scale (NIHSS) between admission and a fixed postoperative assessment at 48 hours. A four-point jump on this scale is not subtle; it typically corresponds to a visible change in function, such as new weakness, speech difficulty, or impaired consciousness. Identifying such patients early could, in principle, allow clinicians to intensify monitoring, escalate blood pressure management, or investigate causes such as vessel re-occlusion, hemorrhagic transformation, or brain edema before irreversible damage occurs.

To build their prediction tool, the researchers turned to a single-center registry covering patients treated between 2020 and 2022. After accounting for overlapping records—253 registry entries representing 251 unique patients—and applying a deliberately conservative rule that excluded a segment of longitudinal data whose status could not be resolved, 214 adults formed the primary analysis cohort. Among these, 51 patients, or 23.8 percent, experienced END48. That prevalence matters: with roughly one in four patients deteriorating, there is a genuine signal to predict, but the absolute number of events also limits how many predictor variables a model can credibly support—a constraint the authors took seriously by locking a five-predictor logistic regression model at the revision stage rather than letting data-driven selection inflate performance.

The five predictors chosen were the admission NIHSS score, patient age, admission blood glucose, white blood cell count, and the natural logarithm of serum triglycerides. Each has a plausible biological rationale. Higher admission NIHSS reflects greater initial brain injury. Elevated glucose is associated with worse outcomes after reperfusion, partly through oxidative stress and blood-brain barrier disruption. Leukocytosis can signal systemic inflammation that amplifies ischemic injury. Triglycerides, less commonly studied in this context, may reflect metabolic risk profiles that influence microvascular integrity. The model’s simplicity—five routinely available measurements, no imaging-derived features, no specialized biomarkers—was the point: a parsimonious tool could be deployed anywhere EVT is performed without additional cost.

But simplicity means nothing without validated accuracy, and this is where the study’s methodological rigor becomes its defining feature. Rather than reporting only the apparent performance of the model on the data used to fit it—a practice that almost always overstates real-world accuracy—the team used optimism correction with 1,000 bootstrap resamples. Bootstrapping essentially asks how much the model’s apparent skill is inflated by chance patterns in the specific sample, then subtracts that inflation. They went further, employing paired nested validation intervals of 1,000 outer and 100 inner resamples to compare their five-predictor model against a stripped-down comparator containing only the admission NIHSS score. This design directly tests whether the four additional variables add anything beyond what the stroke scale already captures.

The results were sobering. The apparent area under the receiver operating characteristic curve (AUC)—a measure of how well the model separates patients who deteriorate from those who do not—was 0.651, itself modest. After optimism correction, the AUC fell to 0.601, with a 95 percent confidence interval spanning 0.511 to 0.723. An AUC of 0.5 represents coin-flip discrimination; 0.601 sits barely above that floor. The corrected difference in AUC between the full model and the NIHSS-only comparator was 0.038, with a confidence interval from −0.025 to 0.171—crossing zero, meaning the improvement could plausibly be nothing at all.

Discrimination is only half the story, however. A prediction model must also produce probabilities that match observed reality, a property statisticians call calibration. The corrected Brier score—a measure of overall probability accuracy where lower is better—was 0.180, compared with a prevalence-only reference of 0.182. In other words, knowing the base rate of deterioration alone (23.8 percent) predicts almost as well as the full model. The Brier Skill Score, which quantifies improvement over that naive baseline, was a mere 0.9 percent. The corrected calibration slope of 0.731 (confidence interval 0.519 to 0.892) fell below the ideal value of 1, indicating that the model’s predicted probabilities were somewhat too extreme—overconfident in both directions.

The team also evaluated clinical usefulness through decision curve analysis, which translates statistical performance into net benefit at specific risk thresholds. At decision thresholds of 10, 20, and 30 percent, the incremental net benefit of the full model over the NIHSS-only comparator was −0.015, 0.036, and 0.013, respectively—and every one of the corresponding 95 percent confidence intervals crossed zero. From a decision-making standpoint, this means there is no reliable evidence that acting on the model’s predictions would help more patients than it would harm through unnecessary interventions.

Perhaps the most intellectually honest portion of the paper is the post-hoc assessment using Riley’s framework, which estimates whether a given cohort is large enough to support model development across plausible assumptions about how well the model can fit the data. Under all reasonable scenarios, the answer was no: the 214-patient cohort fell below the sample size desirable for developing a stable prediction model. A sensitivity analysis using a reconstructed cohort that included the excluded data segment did not change this overall interpretation, reinforcing that the findings were not an artifact of the conservative inclusion rule.

The authors conclude that the model should be regarded as exploratory rather than clinically implementable, and they call for independent validation in external cohorts and improved outcome ascertainment. For a field increasingly crowded with machine-learning prediction papers boasting inflated apparent AUCs, this study offers a counterexample of methodological transparency: it demonstrates how optimism correction, nested validation, calibration assessment, decision curve analysis, and sample-size auditing should be standard practice before any prediction tool is promoted to the bedside. For patients and clinicians, the practical message is that routine admission blood tests and the NIHSS score, however convenient, do not yet reliably identify who will deteriorate after thrombectomy—and that better data, larger cohorts, and richer predictors, including imaging and dynamic biomarkers, will be needed before early-warning tools for post-EVT neurological decline can earn clinical trust.

Subject of Research: Prediction of early neurological deterioration after endovascular thrombectomy using routine admission measurements

Article Title: Development and internal validation of a parsimonious model for neurological deterioration at 48 h after endovascular thrombectomy

Article References: Ren, G., Zhao, C., & Sun, Y. (2026). Development and internal validation of a parsimonious model for neurological deterioration at 48 h after endovascular thrombectomy. Frontiers in Neurology, 17, Article 1888204. https://doi.org/10.3389/fneur.2026.1888204

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1888204

Keywords: endovascular thrombectomy, neurological deterioration, NIHSS, prediction model, internal validation, bootstrap, calibration, Brier score, decision curve analysis, acute ischemic stroke, logistic regression, Frontiers in Neurology