In recent years, the administration of recombinant tissue plasminogen activator (rt-PA) for intravenous thrombolysis has become a cornerstone in the acute management of ischemic stroke (1). Numerous studies have highlighted the importance of early neurological improvement (ENI) as a predictor of long-term outcomes and functional independence (2, 3). Research has shown that patients who exhibit rapid improvement in their neurological status within the first few hours post-treatment are more likely to achieve favorable outcomes (4). The mechanisms underlying ENI are complex and multifactorial, involving reperfusion of ischemic but viable brain tissue, reduction of infarct size, and preservation of the blood–brain barrier (5, 6). Moreover, several blood biomarkers are being explored to identify patients most likely to benefit from rt-PA and to predict early response to treatment.

Machine learning (ML) algorithms have emerged as powerful tools in the construction of predictive models for adverse events following acute ischemic stroke (AIS), offering significant advantages over traditional statistical approaches (7, 8). By leveraging complex, non-linear relationships within large and diverse datasets, ML algorithms can identify subtle patterns and risk factors that may not be apparent through conventional analysis, thereby enhancing the accuracy and robustness of prediction. Techniques such as logistic regression, decision trees (DT), random forests (RF), support vector machines (SVM), and neural networks have been applied to forecast complications like hemorrhagic transformation, recurrent stroke, and mortality (9–11). The use of ML also supports real-time predictions and personalized medicine, enabling timely interventions to mitigate adverse outcomes.

To date, several studies reported machine-learning based prediction of future outcome in stroke patients. For instance, Wen et al. (12) reported that the model constructed by two machine-learning served as robust tools for predicting early neurological deterioration in acute ischemic stroke patients following thrombolysis. Moreover, Fan et al. (13) used four ML methods to screen and recombine the features for construction of prognostic model, and found that this model offers improved prediction accuracy that may reduce rates of misdiagnosis and missed diagnosis in patients with AIS. Regarding ENI in AIS patients undergoing rt-PA treatment, although three studies (14–16) revealed that several clinical indexes, such as diabetes mellitus history, kynurenic acid and kynurenine aminotransferase were associated with the ENI, no studies using ML algorithms to construct predictive model for ENI in AIS patients undergoing rt-PA treatment. In addition, these previous models largely rely on a limited number of variables and traditional statistical methods (e.g., logistic regression), which may not adequately capture the complex, non-linear relationships among multiple prognostic factors. Therefore, our study addresses this gap by incorporating a comprehensive set of clinical, laboratory variables and applying multiple ML algorithms to better model these interactions, thereby enhancing predictive accuracy.

Currently, limited evidence is available for the prediction of ENI in AIS patients undergoing rt-PA treatment. The present study conducted a comprehensive analysis by a larger sample of patients to identify significant differences in various clinical and biochemical parameters between ENI and non-ENI groups. The ENI group, consisting of 97 patients, exhibited lower levels of Hemorrhage, ONT, MCV, NEU, NEU%, NLR, and Fasting GLU, while higher levels of Weight, BMI, LYM, LYM%, LMR, A/G ratio, PA, CHE, and APTT were observed compared to the non-ENI group. These findings suggest that these indices are closely associated with ENI, whereas resting indexes did not significantly differ between the two groups, indicating their limited impact on ENI development.

To further explore the predictive capacity of ML models for ENI, we employed four ML algorithms, including MLP, RF, SSVM, and XGBoost, using the common clinical data. By dividing the patients into an 8:2 training-to-testing ratio, the MLP model demonstrated the highest predictive performance with an AUC of 0.77 in the testing set, outperforming RF (0.72), SVM (0.63), and XGBoost (0.68). Subsequently, by intersecting the critical parameters selected by all four ML methods, we identified six common parameters (APTT, ALT/AST, ONT, MCHC, Weight, and NLR) that were then used to construct a logistic regression model. This refined model achieved an AUC of 0.74, indicating its robustness in predicting ENI. Notably, the nomogram based on these six parameters showed a markedly improved predictive performance compared to individual parameters, underscoring the value of this composite approach.

In the present study, we chose the intersection of top-ranked features across multiple ML algorithms as our primary feature selection strategy for several methodological and clinical reasons. This is because different ML algorithms have distinct biases in feature importance estimation. This consensus approach enhances reproducibility. In addition, methods like LASSO are sensitive to multicollinearity and may arbitrarily select one variable from a correlated group. SHAP values, while interpretable, can be computationally intensive and sensitive to model choice. Our intersection method provides a model-agnostic consensus, reducing dependency on any single algorithm’s output. Finally, we provide the Venn diagram to visually justify the selection, enhancing interpretability for clinicians.

The six predictors in our nomogram exhibit strong pathophysiological plausibility. APTT reflects intrinsic coagulation pathway activity; prolonged APTT may indicate impaired clot lysis or re-occlusion post-thrombolysis, increasing ENI risk. NLR is a well-established marker of systemic inflammation, which exacerbates blood–brain barrier disruption and cerebral edema after ischemic stroke. ONT is a critical determinant of tissue viability; delays beyond 4.5 h are associated with reduced reperfusion success and higher complication rates. ALT/AST ratio may reflect hepatic metabolic capacity and redox state, potentially influencing drug metabolism and oxidative stress. MCHC and weight may serve as proxies for nutritional status and comorbidity burden, which are known to affect stroke outcomes. Notably, we used the ALT/AST ratio rather than the more conventionally reported AST/ALT (De Ritis) ratio. While these ratios are mathematically reciprocal, their interpretability in predictive modeling differs. In our machine learning framework, the ALT/AST ratio demonstrated higher feature importance and better discrimination for ENI compared to the AST/ALT ratio. A lower ALT/AST ratio reflects relatively elevated AST levels, which may indicate subclinical hepatic dysfunction, increased oxidative stress, or systemic inflammation, the conditions known to impair neurovascular recovery after ischemic stroke (18, 19). Emerging evidence suggests that an elevated De Ritis ratio (low ALT/AST) is associated with increased infarct volume, hemorrhagic transformation, and poor functional outcomes in AIS (18, 19). This aligns with our finding that a lower ALT/AST ratio is negatively associated with ENI, reinforcing its role as a biomarker of metabolic vulnerability. Furthermore, ALT is predominantly expressed in hepatocytes, while AST is present in multiple tissues including brain, heart, and skeletal muscle; thus, a shift in this ratio may reflect multi-organ stress responses that modulate post-stroke recovery (20).

At present, the clinical application of this ML-based predictive model is substantial to the clinicians. It enables healthcare providers to identify AIS patients who are more likely to experience ENI after rt-PA treatment, thereby facilitating personalized care plans and timely interventions (21, 22). By leveraging the predictive power of the identified parameters, clinicians can optimize patient selection for thrombolysis, enhance monitoring strategies, and potentially improve outcomes. Moreover, the model’s ability to predict ENI may contribute to reducing the risk of adverse events and improving resource allocation in clinical settings (23, 24). Addition, the nomogram in our study could be used in a clinical setting to aid in decision-making and patient counseling. For example, clinicians can input this patient’s specific variables, such as APTT, ALT/AST, ONT, MCHC, Weight, and NLR, to generate a personalized probability of outcome. Suppose the nomogram-predicted risk is 75%. This high estimated risk may prompt earlier initiation of aggressive therapy or enrollment in a clinical trial, whereas a predicted risk of 20% might support a strategy of active surveillance. In patient counseling, this visual and quantitative tool can help clinicians clearly communicate individual risk, facilitating informed discussions about the potential benefits and harms of different management options. Future studies should focus on validating the model across diverse populations and integrating it into clinical decision support systems to maximize its utility in real-world practice.

Previously, there were studies using ML methods to construct predictive model of stroke outcomes (25, 26), but no study using ML methods to construct predictive model for ENI. In addition, although some previous studies have identified several clinical variables associated with ENI (27, 28), to knowledge, our study was firstly using multiple ML algorithms to construct predictive for ENI in AIS patients undergoing rt-PA treatment. In addition, our results revealed that this model showed a moderate predictive performance, which was more accuracy than previous studies that only present the variables associated with ENI. More importantly, unlike some expensive tests, the clinical indexes used to construct the model are common and cheap in clinical practice, thus it is easy for the clinical doctors to construct the predictive model, and it also did not add the addition burden on the patients. Finally, the nomogram enables the clinician to easy distinguish the patients with high risk of ENI. Therefore, our results hold promise for the precision medicine approaches in AIS patients undergoing rt-PA treatment.

To reduce the risk of overfitting of the model, our study used multiple, complementary strategies throughout the modeling pipeline to mitigate this risk. To reduce optimism in performance estimates, we implemented a strict train-test split (80%: 20%) and reported performance only on the held-out test set. The significant drop in AUC from training (XGBoost: 0.99) to testing (0.68) clearly indicates overfitting in some models, which is why we selected the MLP model (AUC 0.77), the most stable performer across training and test sets, as our primary ML model. In addition, rather than using all the variables in the final model, we drastically reduced dimensionality by selecting only 6 overlapping features from the top 20 of four diverse ML algorithms. This conservative selection was designed precisely to combat overfitting. Third, although we used a final logistic regression model, the intersection-based feature selection acts as a form of implicit regularization by selecting only features consistently ranked high across multiple algorithms, reducing the inclusion of spurious associations. Finally, the nomogram’s C-index (0.817) was bootstrap-corrected, meaning it was adjusted for overfitting using internal validation with 1,000 bootstrap resamples. This provides a more realistic estimate of model performance on new data.

In our study, the VIF value of the six parameters were all less than 2 (range: 1.28–1.45), indicating no significant multicollinearity. We note that although some original laboratory parameters (such as AST, ALT, NEU, LYM) are biologically related, our feature selection strategy prioritized composite indices (including ALT/AST ratio, NLR) over individual components, thereby reducing redundancy and enhancing model stability.

Nevertheless, our study also has many limitations. First, while our model demonstrates acceptable discrimination and calibration in internal validation, the retrospective design, single-center setting, and lack of external validation limit its generalizability. Second, the algorithm was built from the input features, and some hidden relationships may have been ignored because unknown or neglected features were not evaluated by physicians. Third, the patient’s long-term prognosis results were not collected. Fourth, the ML algorithms have its own limitation, which can suffer from overfitting, where models perform well on training data but fail to generalize to new, unseen data. Additionally, these methods often lack transparency, making it difficult to interpret the decision-making process, which can be a significant barrier in clinical applications where explainability is crucial (29, 30). Finally, despite our efforts to minimize overfitting, the relatively small sample size (n = 217) and high-dimensional feature space pose a risk of overfitting, a common challenge in clinical ML studies. Our findings require external validation in larger, multicenter cohorts to ensure generalizability. Therefore, future study is warranted to verify our results and address the above issues.

We developed and internally validated a machine learning-based nomogram that shows promising performance in predicting ENI after thrombolysis. The model, incorporating six clinically accessible variables, may serve as a potential tool to support clinical decision-making. However, Future research should focus on external validation and integration of this model into clinical practice to maximize its utility in clinical settings.