The study, published in Frontiers in Neurology by researchers at Zhujiang Hospital of Southern Medical University in Guangzhou, drew on 16 years of clinical records, from 2008 to 2024. After screening 2,773 hospital files, the team assembled a cohort of 2,095 patients hospitalized with acute stroke. Roughly half of them, 1,050 individuals, went on to develop a documented post-stroke functional impairment after completing acute treatment, while the remaining 1,045 showed no such event during their observed follow-up and were statistically censored. The median follow-up stretched across 1,175 days, giving the researchers an unusually long window into each patient’s trajectory.

Post-stroke sequelae are notoriously difficult to anticipate. Traditional clinical assessments lean heavily on measures such as the NIH Stroke Scale, age, and language or cognitive testing — tools that carry substantial subjective judgment and vary between clinicians. The mechanisms behind lasting impairment are equally tangled: irreversible brain tissue injury interacts with patient-related factors like advanced age, lesion location, and pre-existing comorbidities, alongside biochemical storms of excitatory neurotransmitter release, calcium overload, oxidative stress, neuroinflammation, and blood-brain barrier dysfunction. The Guangzhou team set out to replace this partly subjective picture with an objective, data-driven risk estimate computable from the first blood draw.

Their starting point was ambitious: 37 candidate variables captured at admission, mostly from laboratory measurements taken within the first 24 hours. To decide which variables actually mattered, the researchers ran four independent feature-selection methods in parallel — univariate Cox regression, multivariable Cox regression, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation, and random survival forest variable-importance ranking. Only variables endorsed by all four methods survived the filter, a deliberately conservative strategy designed to eliminate statistical noise. Eight predictors emerged: homocysteine (HCY), creatine kinase-MB (CK-MB), lactate dehydrogenase (LDH), gamma-glutamyl transferase (GGT), high-density lipoprotein cholesterol (HDL-C), venous glucose, direct bilirubin (DBIL), and the CK-MB to total creatine kinase ratio.

With the predictors fixed, the team then pitted 101 different survival-model pipelines against one another, comparing their performance through resampling within the training set. The winner was a hybrid approach that chained a forward stepwise Cox regression — a classical statistical technique for time-to-event data — with a random survival forest, a machine learning method that grows 1,000 decision trees over survival outcomes using log-rank splitting rules. The winning pipeline was locked before it ever saw the evaluation data, a design choice that guards against the quiet optimism that creeps in when models are tuned and tested on the same patients.

The performance figures are striking for a model built from ordinary blood chemistry. The Harrell concordance index, which measures how well the model ranks patients by risk, reached 0.87 in the training set of 1,467 patients and 0.85 in the held-out internal test set of 628. Time-dependent areas under the receiver operating characteristic curve in the test set were 0.800 at one year, 0.830 at three years, and 0.820 at five years, with precision-recall areas of 0.748, 0.772, and 0.761 respectively. Brier scores, which penalize both discrimination errors and miscalibration, stayed between 0.140 and 0.180 across all three horizons. Calibration slopes of 0.96, 0.93, and 0.90 indicate the model’s predicted probabilities tracked observed event rates closely rather than systematically over- or under-promising.

Perhaps most intriguing are the biomarkers themselves. Elevated LDH, an enzyme released when tissue is damaged, correlated with stroke severity and poor functional recovery, suggesting that metabolic stress leaves a measurable signature in peripheral blood. Direct bilirubin, long viewed through the lens of liver function, has recently been linked in meta-analyses to stroke severity, hemorrhagic transformation, and adverse outcomes. Admission hyperglycemia is well established as a harbinger of poor recovery in ischemic stroke. Homocysteine may impair vascular endothelial function and reduce nitric oxide availability, stiffening vessels and limiting perfusion. And the CK-MB to CK ratio points toward a fascinating frontier: the so-called brain-heart axis, in which neurological injury triggers cardiac enzyme release that may itself reflect the depth of the brain insult.

To open the black box, the researchers applied SHAP (SHapley Additive exPlanations) analysis, a technique from explainable AI that decomposes each individual prediction into contributions from every feature. Global importance rankings confirmed LDH, HCY, CK-MB, and DBIL as leading drivers of the model’s output, while beeswarm plots revealed how each variable pushed individual patients toward higher or lower risk. The team also distilled the full eight-variable model into a simplified nomogram using only LDH, CK-MB, and DBIL — the three markers shared by the top rankings of both interpretation methods. The three-variable version gave up some accuracy, with test-set AUROCs of 0.750 to 0.780, though paired bootstrap comparisons suggested the difference was not statistically significant. Both the full model and the simplified nomogram have been deployed as a free interactive web application where clinicians can enter a patient’s values and receive individualized one-, three-, and five-year probabilities of remaining sequelae-free.

The clinical implications could be significant. Decision curve analysis showed the model delivered higher net benefit than both treat-all and treat-none strategies across clinically relevant threshold ranges, meaning that in practical terms it could help triage limited rehabilitation and surveillance resources toward the patients who need them most. Kaplan-Meier curves separated predicted high-risk and low-risk groups dramatically in both the training and test cohorts, with log-rank p-values below 0.001. Because the predictors are drawn from tests ordered routinely at admission, the framework requires no new infrastructure, no specialized imaging, and no additional cost per patient — an attribute that matters enormously for health systems in low- and middle-income countries where the stroke burden is heaviest.

The authors are careful, appropriately, to temper enthusiasm. This was a retrospective, single-center study, internally validated only; the 16-year recruitment window may span changes in laboratory assays, treatment pathways, and follow-up practice that a random train-test split cannot fully account for. Selecting one winner from 101 candidate pipelines introduces a residual risk of selection optimism despite careful resampling discipline, and the correlated CK-MB and CK-MB/CK predictors warrant further scrutiny. External validation in independent, geographically diverse, and prospectively recruited cohorts — with standardized adjudication of post-stroke sequelae — remains the essential next gate before any clinical deployment. Still, if those validations hold, the study points toward a future in which the first vial of blood drawn in the emergency room quietly forecasts not just whether a patient will survive a stroke, but what kind of life awaits on the other side.

Subject of Research: Machine learning prediction of post-stroke sequelae risk using clinical and biochemical variables

Article Title: Post-stroke sequelae risk score (PSSRS): a 16-year retrospective study developing an internally validated multi-algorithm model for post-stroke sequelae

Article References: Dong, Y., Lu, R., Zhang, Y., Huang, Y., Wang, Z., Sun, J., Li, S., Jia, Y., Tian, Y., Hu, H., Li, X., Liu, Y., & Duan, C. (2026). Post-stroke sequelae risk score (PSSRS): a 16-year retrospective study developing an internally validated multi-algorithm model for post-stroke sequelae. Frontiers in Neurology, 17, Article 1833445. https://doi.org/10.3389/fneur.2026.1833445

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1833445

Keywords: stroke, post-stroke sequelae, machine learning, random survival forest, biomarkers, risk prediction model, LDH, CK-MB, direct bilirubin, survival analysis, clinical prediction, neurology