Clinical severity outperforms routine biomarkers for predicting stroke bleeding

In a retrospective study of 847 adults treated with intravenous thrombolysis for acute ischemic stroke, established clinical features predicted in-hospital hemorrhagic transformation substantially better than routine inflammatory and metabolic biomarkers. A six-variable model combining age, baseline neurological severity, atrial fibrillation, large-vessel occlusion, rapid blood glucose and onset-to-treatment time achieved an area under the receiver operating characteristic curve, or AUC, of 0.801 in a later temporal-validation cohort. Adding three laboratory markers improved the AUC by only 0.018, a change that was not statistically significant and was not reproduced consistently in sensitivity analyses.

The researchers divided admissions from Changsha Central Hospital chronologically rather than randomly. The earliest 70 percent, from January 2020 through November 2024, formed a derivation cohort of 593 patients, including 67 with hemorrhagic transformation. The remaining 254 admissions, recorded from late November 2024 through May 2026, formed the internal temporal-validation cohort and included 27 hemorrhagic transformation events. Overall, 94 of 847 admissions, or 11.1 percent, had imaging-confirmed in-hospital hemorrhagic transformation. The design was intended to test whether laboratory markers added useful information beyond readily available clinical variables under changing real-world conditions.

The baseline model captured age, baseline National Institutes of Health Stroke Scale score, atrial fibrillation, large-vessel occlusion, rapid-glucose measurement and onset-to-treatment time. The candidate additions were the natural logarithm of the neutrophil-to-lymphocyte ratio, or ln(NLR), the atherogenic index of plasma, or AIP, and residual cholesterol. In temporal validation, the clinical model had an AUC of 0.801 with a 95 percent confidence interval of 0.707 to 0.881. Adding ln(NLR) raised the AUC to 0.818, a difference of 0.017 with a DeLong test p value of 0.220. Adding AIP and residual cholesterol produced an AUC of 0.803, while the full nine-variable model reached 0.818, a gain of 0.018 with p = 0.346.

Discrimination was therefore only numerically better after the biomarkers were introduced. Calibration and overall prediction error were virtually unchanged: both the clinical and full models had Brier scores of 0.081 in temporal validation, and their calibration slopes were 1.029 and 1.060, respectively. Decision-curve analysis likewise showed closely overlapping net-benefit trajectories, without a broad or consistently separated advantage for the full model. In other words, the laboratory additions did not demonstrate that they would improve clinical decisions more reliably than the simpler model across the range of decision thresholds examined.

The adjusted analysis identified baseline stroke severity as the clearest independent correlate of the broader hemorrhagic transformation outcome. Each one-point increase in the NIHSS score was associated with 8.4 percent higher odds of hemorrhagic transformation, with an odds ratio of 1.084, a 95 percent confidence interval of 1.044 to 1.126 and p below 0.001. Atrial fibrillation was associated with approximately 2.7-fold higher odds, with an odds ratio of 2.66, a confidence interval of 1.34 to 5.27 and p = 0.005. Large-vessel occlusion had a positive but statistically imprecise association. By contrast, ln(NLR), AIP and residual cholesterol were not independently associated with the primary outcome after the clinical variables were included.

That contrast extended to the unadjusted patient characteristics. Admissions with hemorrhagic transformation had a median baseline NIHSS score of 15, compared with 5 among those without transformation, and large-vessel occlusion was present in 75.9 percent versus 41.4 percent. Atrial fibrillation was also more frequent, at 46.2 percent versus 15.7 percent. Patients with transformation also had higher NLR and other inflammatory indices, lower total and low-density lipoprotein cholesterol, and a lower AIP. These differences did not translate into independent biomarker effects or robust incremental prediction once the clinical model was included, illustrating why an association in a group comparison is not equivalent to useful patient-level discrimination.

A separate exploratory analysis produced a narrower signal for strict symptomatic intracranial hemorrhage, or sICH, defined using the ECASS II framework and restricted to events within seven days. There were 37 sICH events overall, including only 12 in the temporal-validation cohort. In the derivation data, ln(NLR) was associated with sICH after adjustment for NIHSS, atrial fibrillation and large-vessel occlusion, with an odds ratio of 1.89, a 95 percent confidence interval of 1.04 to 3.42 and p = 0.037. Adding NLR to the exploratory clinical model increased its validation AUC from 0.816 to 0.888, but the increment of 0.071 did not reach statistical significance, with p = 0.097. The small number of validation events makes this finding hypothesis-generating rather than evidence of clinical readiness.

The robustness checks did not establish a reproducible advantage from adding biomarkers. Excluding patients who received bridging thrombectomy produced validation AUCs of 0.770 for the clinical model and 0.761 for the full model. Restricting the analysis to patients with known large-vessel occlusion status produced AUCs of 0.800 and 0.815, respectively, while limiting hemorrhagic transformation to events documented within seven days yielded 0.802 and 0.809. Excluding three events whose archived documentation could not be re-verified and adopting a conservative approach to uncertain laboratory timing likewise failed to demonstrate consistent biomarker gains. A separate gradient-boosting analysis within the same cohort also ranked NIHSS, atrial fibrillation and large-vessel occlusion above the three biomarkers, but it was not an independent validation.

Several design features strengthen the central conclusion. Preprocessing and multiple-imputation models were fitted in the derivation cohort and then applied to the later cohort, limiting information leakage. The analysis assessed discrimination, calibration, Brier score, decision curves and multiple sensitivity analyses rather than relying on a single performance metric. A 500-resample bootstrap check estimated a mean optimism of 0.031 for the full model’s apparent AUC, producing an optimism-corrected derivation AUC of approximately 0.786. A post-hoc adequacy calculation also indicated that more development observations would have been desirable. The authors therefore characterized the work as an exploratory incremental-value analysis rather than the development of a deployable clinical risk score.

Important limitations temper the conclusions. This was a single-center retrospective study, so institutional practices and patient selection may limit generalizability, and internal temporal validation does not establish transportability to other hospitals. Hemorrhagic transformation was identified from official radiology reports rather than independent, blinded review of archived images, and clinically triggered repeat imaging may have affected ascertainment. Although the analyzed blood and lipid measurements came from initial emergency specimens collected before thrombolysis, exact draw-to-bolus intervals were not retained for every patient, and not all results were necessarily available before treatment. Predictor missingness and biomarker distributions also changed over time. The results therefore support evaluating future markers against strong clinical baselines, but they do not justify withholding intravenous thrombolysis from otherwise eligible patients.

Subject of Research: Incremental prediction of hemorrhagic transformation after intravenous thrombolysis for acute ischemic stroke

Article Title: Clinical severity predominates over routine inflammatory and metabolic biomarkers in predicting hemorrhagic transformation after intravenous thrombolysis: internal temporal validation of incremental biomarker value

Article References: Liang, H., Zeng, B., Wang, J., Yi, L., Huang, L., Lin, Z., Zhou, C., & Yue, Z. (2026). Clinical severity predominates over routine inflammatory and metabolic biomarkers in predicting hemorrhagic transformation after intravenous thrombolysis: internal temporal validation of incremental biomarker value. Frontiers in Neurology, 17, Article 1940110. https://doi.org/10.3389/fneur.2026.1940110

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1940110

Keywords: acute ischemic stroke, intravenous thrombolysis, hemorrhagic transformation, NIHSS, neutrophil-to-lymphocyte ratio, atherogenic index of plasma, residual cholesterol, prediction model, temporal validation, Clinical, severity, predominates