Machine Learning-Based Risk Stratification for Symptomatic Intracranial Hemorrhage and 3-Month Prognosis following Intravenous Thrombolysis

Abstract

Introduction

Intravenous thrombolysis using tissue-type plasminogen activator (tPA) is widely accepted as a fundamental therapy for acute ischemic stroke. However, its clinical benefit is counterbalanced by the risk of symptomatic intracranial hemorrhage (sICH), a serious complication that can substantially worsen patient outcomes and increase mortality. Functional outcome after stroke is most commonly assessed using the modified Rankin Scale at 3 months. In this study, we sought to construct predictive models for sICH and 3-month outcomes after tPA and to identify key prognostic variables that may support individualized treatment decisions.

Methods

We analyzed data from 434 patients with ischemic stroke who received intravenous tPA at a tertiary medical center over a 5.5-year period. Three supervised classification models were constructed and validated using five-fold cross-validation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, recall, and precision, and the results were compared with those of six commonly used clinical scoring systems for predicting post-tPA sICH.

Results

All three machine learning models showed strong discriminatory performance for sICH prediction, with AUC values of 0.87 for logistic regression, 0.82 for random forest, and 0.89 for XGBoost. Each model consistently outperformed the six conventional scoring tools. Among all variables, the 24-h NIHSS score contributed most strongly to the prediction of both sICH and 3-month functional outcomes. In addition, a prior history of stroke and male sex were associated with an increased risk of sICH, whereas older age was closely linked to worse functional outcomes at 3 months.

Conclusion

The proposed machine learning models achieved high predictive performance for post-tPA sICH and 3-month outcomes and identified key clinical variables with substantial prognostic importance. Notably, the 24-h NIHSS score emerged as the most influential predictor across models. Compared with existing sICH scoring systems, these models demonstrated superior performance, supporting their potential role as clinical decision-support tools in post-thrombolysis management.

Introduction

Intravenous thrombolysis (IVT) using tissue-type plasminogen activator (tPA) is a fundamental therapy for acute ischemic stroke within the conventional 4.5-h time window [1, 2]. Advanced imaging techniques, such as multimodal CT or MRI, are used to identify patients who may still benefit from thrombolysis by detecting salvageable brain tissue. However, despite its proven efficacy, IVT carries the risk of symptomatic intracranial hemorrhage (sICH), a complication associated with severe morbidity and mortality [3]. Hypertension (HTN) has long been recognized as a major risk factor for hemorrhagic stroke because of chronic vascular injury and impaired cerebrovascular autoregulation, contributing to greater cerebrovascular vulnerability [4]. Functional outcome after ischemic stroke is most commonly evaluated using the modified Rankin Scale (mRS) at 3 months, which reflects patients’ level of independence in daily activities. Numerous clinical and demographic factors have been reported to influence 3-month outcomes [5]. In contrast, the relationship between sICH and longer term functional outcome is less straightforward. Many patients who develop sICH already have extensive baseline infarction and severe neurological deficits, which inherently predispose them to poor prognosis regardless of hemorrhagic complications [6]. A number of risk prediction tools have been proposed to estimate the likelihood of sICH after IVT, including SEDAN, HAT, MSS, GRASPS, SITS, and SPAN-100 ([7–12]). Although helpful for risk stratification, these scoring systems generally demonstrate only moderate predictive accuracy, highlighting the need for improved and data-driven approaches. In this study, we sought to develop models to predict outcomes for ischemic stroke patients treated with tPA, specifically focusing on sICH and 3-month mRS prognosis and to identify the key variables associated with these outcomes.

Method

This study analyzed data from 434 ischemic stroke patients who received tPA treatment at a tertiary medical center during the 5.5-year period. The study was approved by the hospital’s Institutional Review Board, and patient data were retrospectively collected. Baseline demographic data included age, height, body weight, body mass index, gender, and initial systolic and diastolic blood pressure on arrival. Comorbidities were categorized based on patients' medical history, including HTN, diabetes mellitus, previous stroke, heart disease (coronary artery disease or valvular heart disease), uremia, dyslipidemia, heart failure, and malignancy. Treatment-related variables included the time from symptom onset to tPA administration (symptom-onset-to-needle time) and the administered tPA dose (mg/kg). Laboratory parameters were collected as comprehensively as possible, including hemoglobin, white blood cell count, neutrophils, lymphocytes, prothrombin time, international normalized ratio, activated partial thromboplastin time, fibrinogen, D-dimer, blood glucose, platelet count, HbA1c, creatinine, total cholesterol, triglycerides, high-density lipoprotein, and low-density lipoprotein. Imaging examinations included brain CT and brain MRI. Brain CT was used to assess hypodensity and early infarction signs, and brain MRI was routinely performed the following day to evaluate ischemic injury, hemorrhagic transformation, and related complications. Medications prior to stroke were recorded as antiplatelet use or anticoagulation use. Neurological severity was documented using the National Institutes of Health Stroke Scale (NIHSS) at the time of arrival in the emergency department and again 24 h after hospital admission. sICH was defined according to the American Heart Association and American Stroke Association criteria. This definition is rooted in the ECASS III framework and requires both radiographic evidence of intracranial hemorrhage and clinical deterioration, defined as an increase in NIHSS score of at least 4 points or death within 36 h, where hemorrhage is considered the most likely cause. Importantly, the NIHSS variable used as a predictor in the models was the absolute 24-h NIHSS score, not the NIHSS increase used in the diagnostic criteria, which avoided circularity in prediction. Among the 434 patients included, laboratory variables contained different degrees of missingness ranging from zero to two percent depending on the test. No patients were excluded solely because of missing laboratory parameters. Only patients lacking 24-h NIHSS evaluation or 3-month mRS assessment were excluded. To minimize selection bias, missing values were addressed using imputation rather than case exclusion. Imputation of continuous variables used the median and imputation of categorical variables used the mode. This process was performed after the train–test split, and parameters were calculated only from the training set to prevent information leakage into the test set. Three supervised classification models were trained to predict sICH and 3-month mRS outcomes: Logistic Regression, Random Forest, and XGBoost. All numerical variables were standardized and categorical variables were encoded when needed. To reduce potential instability in performance estimates, stratified five-fold cross-validation was applied to preserve the original ratio of sICH to non-sICH cases across folds. Model performance was evaluated using area under the receiver operating characteristic (ROC) curve (AUC), recall (sensitivity), specificity, accuracy, and precision. Model performance was compared with six established scoring tools (SEDAN, HAT, MSS, GRASPS, SITS, and SPAN-100) for predicting post-tPA sICH. To assess model interpretability, Shapley Additive Explanations (SHAP) were computed to quantify the contribution of each feature to prediction. These values represent associations rather than causal effects.

Result

A total of 7,457 ischemic stroke cases were identified from the stroke registry, as shown in Figure 1. After excluding 6,736 patients who did not receive IVT, 721 patients remained eligible for screening. Of these, 275 patients treated with combined endovascular therapy and 12 patients with unavailable or incomplete follow-up data were excluded. Ultimately, 434 patients who received IV tPA alone and had complete 24-h NIHSS and 3-month mRS data were included in the final analysis.

Fig. 1.

Patient flow diagram.

Baseline clinical and demographic characteristics of patients with and without sICH following thrombolytic therapy are shown in Table 1. HTN was observed in 25 patients with sICH, compared to 342 patients without sICH. The difference was statistically significant (p = 0.020), indicating that HTN was significantly more frequent in the sICH group and may be associated with increased risk. Age was slightly higher in the sICH group (70.88 vs. 68.44 years), although this difference did not reach statistical significance (p = 0.403). The mean symptom-onset-to-needle time was shorter in the sICH group (116.84 vs. 130.55 min), a pattern that suggested a possible association but lacked statistical significance (p = 0.168). Heart disease and dyslipidemia appeared more common in the non-sICH group, although neither difference was statistically significant. All p values in Table 1 are exploratory and not inferential for feature selection in the predictive models.

Table 1.

Baseline characteristics of patients with and without sICH

Variable	sICH		p value
	no, N = 409 (94.24%)	yes, N = 25 (5.76%)
Age, mean±SD	68.45±14.22	70.88±12.49	0.403
BMI, mean±SD	24.93±4.07	25.07±3.20	0.863
Gender	​	​	0.184
Male	257 (62.84)	19 (76.00)	​
Female	152 (37.16)	6 (24.00)	​
Initial SBP, mean±SD	156.7±28.82	164.2±27.12	0.206
Initial DBP, mean±SD	92.05±18.50	94.72±22.55	0.490
HTN	342 (83.62)	25 (100.0)	0.020
DM	145 (35.45)	9 (36.00)	0.955
CVA	86 (21.03)	7 (28.00)	0.409
Heart disease	188 (45.97)	16 (64.00)	0.079
Dyslipidemia	308 (75.31)	16 (64.00)	0.207
HF	48 (11.74)	2 (8.00)	0.754
Malignancy	16 (3.91)	2 (8.00)	0.277
Onset to tPA (min), mean±SD	130.6±47.78	116.8±54.85	0.168
tPA dose (mg/kg), mean±SD	0.76±0.15	0.78±0.15	0.465
Antiplatelet	70 (17.11)	5 (20.00)	0.784
Anticoagulation	23 (5.62)	2 (8.00)	0.647

Values are presented as n (%) unless otherwise indicated. BMI, body mass index; DM, diabetes mellitus; HF, heart failure.

The ROC curves comparing the performance of three machine learning models and six conventional scoring tools for predicting sICH are shown in Figure 2. All three machine learning models demonstrated strong discriminatory performance: Logistic Regression (AUC = 0.87), Random Forest (AUC = 0.82), and XGBoost (AUC = 0.89), evaluated using stratified 5-fold cross-validation. XGBoost achieved the highest AUC. In contrast, the six traditional scoring tools showed lower predictive accuracy, including SEDAN (AUC = 0.55), MSS (AUC = 0.63), HAT (AUC = 0.75), GRASPS (AUC = 0.71), SITS (AUC = 0.69), and SPAN-100 (AUC = 0.51).

Fig. 2.

ROC curves of machine learning models vs. scoring tools for sICH prediction.

The SHAP summary plot in Figure 3 illustrates the top 20 most influential features contributing to sICH prediction in the XGBoost model. The most determinant feature is the 24-h NIHSS score, where higher values were strongly associated with increased predicted sICH risk. Although shorter onset-to-treatment times tended to be associated with higher sICH risk, many longer times also demonstrated high SHAP values, reflecting considerable heterogeneity and underscoring the limited consistency of this predictor. Previous stroke and sex also showed moderate influence, with male sex associated with higher predicted sICH risk. Predictive importance does not imply a mechanistic causal relationship, and all SHAP findings should be interpreted as associations that improve model performance. These findings highlight the non-linear nature of risk contribution across patients.

Fig. 3.

SHAP summary plot of feature importance for sICH prediction using the XGBoost model.

The DeLong test comparing AUC values between the XGBoost model and the other scoring tools and machine learning models is shown in Table 2. XGBoost significantly outperformed all six traditional tools. In contrast, the differences between XGBoost and the other two machine learning models, Logistic Regression (ΔAUC = 0.024, p = 0.579) and Random Forest (ΔAUC = 0.024, p = 0.403), were minimal and not statistically significant.

Table 2.

AUC comparison of XGBoost vs. clinical scores and ML models for sICH prediction

Tools	AUC Diff (XGBoost – Target)	p value
SEDAN	0.344	0
MSS	0.259	<0.001
HAT	0.142	0.014
GRASPS	0.180	0.006
SITS	0.199	0.002
SPAN-100	0.379	0
Logistic	0.024	0.579
RandomForest	0.024	0.403

ROC curves for predicting 3-month outcomes (good: mRS 0–2; poor: mRS 3–6) based on the three machine learning models are shown in Figure 4. Logistic Regression and Random Forest each achieved an AUC of 0.85 ± 0.03, while XGBoost achieved an AUC of 0.84 ± 0.03, indicating consistently high predictive capability across models.

Fig. 4.

ROC curves of three machine-learning models for predicting 3-month mRS outcomes.

The forest plot in Figure 5 illustrates the odds ratios and their 95% confidence intervals for predictors of poor 3-month functional outcome based on logistic regression. The 24-h NIHSS score demonstrated the highest odds ratio with a narrow confidence interval that did not cross 1, indicating a strong and statistically significant association with poor outcomes. Age also showed a significant association, with a relatively narrow confidence interval, suggesting that older age negatively affects recovery.

Fig. 5.

Forest plot of predictors of poor 3-month mRS outcome.

Discussion

Our models effectively predicted both sICH and 3-month mRS prognosis and highlighted the variables with the greatest influence on these outcomes. The 24-h NIHSS showed the highest predictive influence for both sICH and 3-month mRS outcomes, although its role should be interpreted as associative rather than causative. Higher 24-h NIHSS scores were strongly associated with both sICH and poor functional recovery [13]. This finding is consistent with numerous studies demonstrating that baseline NIHSS is linked to hemorrhagic risk and can predict 3-month outcomes [14], and prior research has reported that both ΔNIHSS (24-h NIHSS minus baseline NIHSS) and 24-h NIHSS alone are useful for outcome prediction [15, 16]. Some studies have also used the 24-h NIHSS score alone to predict 90-day prognosis [17]. HTN, a well-recognized risk factor for hemorrhagic stroke [4, 18], showed a significant difference between groups in the univariate analysis, but its relative importance was low in the machine-learning model. This difference likely arises from the analytical framework used. While Table 1 summarizes unadjusted group-level associations, SHAP evaluates variable importance within a multivariable model that accounts for interactions among predictors. Consequently, some variables that are significant in univariate analyses may play a smaller role once dominant predictors are considered together. Previous stroke and male sex also emerged as relevant predictors in the SHAP analysis. The association between previous stroke and sICH aligns with clinical guidelines that recommend caution in administering IV tPA soon after a previous ischemic event [19]. Male sex displayed higher SHAP values in our cohort, suggesting a greater contribution to sICH risk [12, 20]; although literature on gender differences is mixed [21], biological and vascular factors may contribute to this vulnerability. Increasing age was associated with poor 3-month prognosis [19, 22], consistent with evidence showing that older individuals are more susceptible to complications and delayed functional recovery. Several clinical risk scores have been proposed to estimate the risk of sICH after thrombolytic therapy and to support bedside decision-making. In practice, however, most of these tools achieve only modest discriminative performance, with reported AUC values generally in the range of 0.65 to 0.75. In contrast, the machine-learning models applied in our cohort showed higher discriminatory ability, with AUCs of 0.87 for logistic regression, 0.82 for random forest, and 0.89 for XGBoost, exceeding the performance of established scoring systems. While these findings suggest that machine-learning approaches may offer improved early risk stratification, their translation into routine clinical use should be approached cautiously in light of the study’s design and inherent limitations. One possible reason for this difference in performance relates to the underlying analytical approach. Conventional regression-based scoring systems are typically built on a limited set of preselected variables and assume linear relationships between predictors and outcomes. Such assumptions may restrict their ability to reflect the complex and interdependent factors that contribute to hemorrhagic complications and post-stroke recovery. Machine learning, by contrast, can model high-dimensional and non-linear relationships automatically, allowing the model to identify multivariable patterns that classical tools may overlook. This may be one of the factors contributing to the superior performance observed in many machine-learning–based prediction studies compared with traditional scoring systems. [23–25]. These methodological advantages likely explain the performance gap between machine-learning models and conventional scoring tools. This study has limitations. It was retrospective and single-center with a modest sample size and low sICH event rate, which may affect model robustness. External validation was not performed, and therefore, generalizability is uncertain. Although median/mode imputation minimized missingness, residual selection bias is possible because patients lacking 24-h NIHSS or 3-month outcomes were excluded. ASPECTS was unavailable in a standardized format, and the 24-h NIHSS reflects post-treatment status; future studies incorporating only pretreatment variables are warranted.

Conclusion

Our machine-learning models accurately predicted both sICH and 3-month mRS outcomes after IV tPA and outperformed conventional scoring tools. The 24-h NIHSS score was the strongest predictor of early hemorrhagic risk and long-term prognosis, followed by HTN, previous stroke, age, and sex. These results support the potential of machine-learning–based prediction to enhance individualized risk assessment after thrombolysis. External multi-center validation is needed before clinical implementation.

Acknowledgments

We are grateful to the clinicians and staff at China Medical University Hospital, Taiwan, who contributed to the medical intervention and data collection for this study. The authors also used ChatGPT (OpenAI) solely for English language editing and improvement of readability. All scientific content, interpretation, and conclusions were developed by the authors, who take full responsibility for the manuscript.

Statement of Ethics

This study was approved by the Research Ethics Committee I of China Medical University & Hospital, Taichung, Taiwan (Protocol No. CMUH111-REC1-199(CR-3)). Consent to participate statement: Written informed consent was obtained from all individual participants included in the study.

Conflict of Interest Statement

The authors have no conflicts of interest to declare.

Funding Sources

This study was supported by China Medical University Hospital (Grant No. DMR-115-072) and Asia University (Grant No. ASIA-111-CMUH-11).

Author Contributions

C.-W.L.: conceptualization, methodology, formal analysis, data curation, and writing – original draft. W.-C.W.: investigation and writing – review and editing. J.-L.H. and C.-Y.L. (Chien-Yu Liu): investigation and data curation. S.-H.W.: methodology and formal analysis. C.-Y.L. (Chun-Yuan Lin): supervision, conceptualization, and writing – review and editing. C.C.N.W.: supervision, methodology, and writing – review and editing. All authors approved the final version of the manuscript and agree to be accountable for all aspects of the work.

Funding Statement

This study was supported by China Medical University Hospital (Grant No. DMR-115-072) and Asia University (Grant No. ASIA-111-CMUH-11).

Data Availability Statement

The data that support the findings of this study are not publicly available due to privacy and ethical restrictions related to patient information but are available from the corresponding author upon reasonable request.