Stroke is a devastating global health burden and remains one of the leading causes of death and long-term disability across all age groups [1, 2]. The World Health Organization estimates that over 15 million people experience a stroke each year; approximately 5 million die and 5 million are left with permanent disability [3]. A major challenge in effective stroke management is the significant risk of recurrence. Epidemiological evidence indicates that about 5.7–51.3% of patients experience a second stroke within the first year after the initial event, and the risk can persist for years [4]. Recurrent stroke often results in more severe neurological impairment, increased healthcare costs, and a significant reduction in quality of life for patients and their families [5]. Therefore, early and accurate identification of individuals at high risk of recurrence is not merely a clinical priority but also a critical public health need, enabling individualized secondary prevention strategies to mitigate this risk.

Traditional approaches for predicting the risk of stroke recurrence, such as the Essen Stroke Risk Score (ESRS), the Stroke Prognostic Instrument (SPI), and the ABCD² (Age, Blood pressure, Clinical features, Duration of symptoms, Diabetes) score, are widely used in routine clinical care [6]. These models generally rely on a limited set of readily available clinical variables, including age, history of hypertension, diabetes mellitus, atrial fibrillation, and a previous transient ischemic attack (TIA) [7]. While they provide a convenient approach to risk stratification, their predictive performance is often moderate, with validation studies demonstrating area under the curve (AUC) values of 0.6 to 0.7 [8]. This modest accuracy indicates, in part, the limited ability of these strategies to capture the complex, multifactorial biology of stroke, which involves interactions between clinical features, biochemical pathways, and structural brain changes. Moreover, many of these models often do not incorporate detailed neuroimaging information that can provide insights into the severity and anatomical distribution of cerebral damage, all of which are important determinants of recurrence risk.

In recent years, machine learning (ML) has revolutionized various fields of medicine, including diagnostic imaging, prognostic prediction modeling, and assessment of treatment response [9]. By processing high-dimensional data, identifying non-linear relationships, and extracting complex patterns from large datasets, ML approaches offer a promising alternative to traditional statistical methods for predicting stroke recurrence [10]. Unlike conventional approaches, ML models can integrate diverse data sources, including routine clinical variables, laboratory results, and imaging-derived features, enabling the development of more comprehensive and more accurate prediction tools [11].

Neuroimaging, in particular, holds significant potential for enhancing the prediction of recurrent stroke risk. Computed tomography (CT) and magnetic resonance imaging (MRI) can characterize infarct size and location and detect associated pathologies such as leukoaraiosis, cerebral microbleeds, and carotid artery stenosis [12]. These imaging features can reflect the underlying vascular pathology, the severity of cerebral ischemia, and the burden of silent cerebrovascular disease, all of which are strongly linked to stroke recurrence. For example, larger infarct sizes have consistently been associated with a higher recurrence risk [13], likely indicating more extensive vascular injury and a greater likelihood of unstable atherosclerotic plaques. Similarly, leukoaraiosis, a marker of cerebral small-vessel disease, has been established as an independent predictor of recurrent vascular events [14].

Despite growing interest in applying ML in stroke research, limited studies have performed systematic comparisons of various ML algorithms for predicting stroke recurrence using a combination of routine clinical variables and imaging features. Most published studies have assessed only a single algorithm or have used one data modality alone (e.g., clinical data without imaging, or imaging without detailed clinical data), which limits our understanding of which algorithm and which data integration approach yields the best predictive performance. Additionally, prioritizing and interpreting the most influential predictors of recurrence within an integrated dataset remains crucial, both to enhance model transparency and to generate mechanistic insights that could inform the development of more effective secondary preventive strategies.

Therefore, this study aims to address these gaps by evaluating the performance of four commonly used ML approaches: logistic regression, random forest, support vector machine (SVM), and extreme gradient boosting (XGBoost). Using an integrated dataset that combines routine clinical data with detailed imaging features, the study seeks to determine which algorithm achieves the highest predictive performance for stroke recurrence. Furthermore, the study will identify the most influential predictors of recurrence within the integrated dataset and assess the generalizability of the optimal model across clinically relevant subgroups, such as patients with cortical versus subcortical infarcts. Overall, the findings may support the development of more accurate and clinically useful tools for recurrence risk stratification, enabling more individualized secondary prevention and improved patient outcomes.

The present study systematically compared the performance of four machine learning algorithms for predicting stroke recurrence using an integrated set of routine clinical variables and imaging features. Among them, the XGBoost model demonstrated the strongest predictive performance, achieving an AUC of 0.86. The findings underscore the potential of ML-based approaches to enhance risk stratification for stroke recurrence and to address key limitations of traditional prediction models that rely on a narrow set of clinical variables.

The superior performance of XGBoost compared with logistic regression, random forest, and SVM aligns with previous findings highlighting that gradient-boosting frameworks are well-suited to complex, high-dimensional clinical datasets [21]. A possible explanation for its superior performance is XGBoost’s capability to model non-linear relationships and higher-order interactions among variables, such as the synergistic effect of infarct size and hypertension. For instance, while large infarcts are associated with higher recurrence risk, this effect may be significantly amplified in patients with poorly controlled hypertension, a relationship that linear models such as logistic regression may not capture adequately. This capability is particularly relevant in stroke research, where recurrence risk is determined by a complex interaction of vascular, metabolic, and neuroimaging-related factors.

Integrating imaging-derived features into the predictive models represents a key strength of this study. Traditional models often overlook neuroimaging data because of its analytical complexity and the need for specialized interpretation; however, our results indicate that imaging features, particularly infarct size, contribute significantly to recurrence prediction. Infarct size, ranked as a crucial predictor in the XGBoost model, consistent with previous evidence linking larger infarcts to higher recurrence risk [22]. Larger infarcts usually reflect more severe arterial occlusion, greater ischemic injury, and a higher likelihood of underlying vasculopathy, which all together increase the risk of subsequent cerebrovascular events [23]. Additionally, incorporating markers such as leukoaraiosis and carotid artery stenosis captures the contributions of small-vessel disease and large-artery atherosclerosis, respectively, thereby enhancing the clinical relevance of risk stratification [24].

The identification of hypertension and fasting blood glucose as key predictors reinforces the crucial role of metabolic and vascular risk management in secondary prevention. Hypertension, a well-established driver of stroke pathogenesis, promotes arteriosclerosis, disrupts endothelial function, and increases susceptibility to small vessel occlusion [25]. Similarly, elevated fasting blood glucose levels, even among individuals without a diagnosis of diabetes, may indicate insulin resistance and systemic inflammation, both of which contribute to vascular injury and thrombus formation [26]. Notably, lifestyle-based interventions can significantly improve these metabolic parameters [27]. These findings support current clinical guidelines that emphasize tight blood pressure and glycemic management after stroke, while also highlighting how ML-based models may help identify high-risk individuals who could benefit from more aggressive intervention.

Subgroup analyses revealed that the XGBoost model maintained strong predictive performance across patients with cortical, subcortical, and posterior circulation infarcts, suggesting good generalizability in distinct stroke subtypes with varying etiologies (e.g., large-artery atherosclerosis for cortical, small-vessel disease for subcortical, and vertebrobasilar pathology for posterior circulation). This result is clinically relevant because cortical and subcortical strokes often have distinct etiologies, such as large-artery atherosclerosis and small-vessel disease, and may therefore require tailored preventive strategies [15]. The consistent performance of the model across these subgroups supports its potential ability as a flexible and broadly applicable approach in clinical risk stratification.

Our results also highlight the limitations of traditional risk scores. For example, the ESRS, which relies on variables such as age, hypertension, and diabetes, typically achieves an AUC of about 0.65–0.70 for predicting recurrence [28]. In contrast, the XGBoost model yielded an AUC of 0.86, representing a meaningful improvement in predictive accuracy that could improve identification of high-risk patients. However, ML-based models should be used to complement, not replace, clinical decision-making. While the XGBoost model provides a quantitative risk estimation, clinicians should interpret these findings alongside patient-specific factors, including adherence to medication and lifestyle factors, to guide tailored management.

Several limitations of the study should be considered before interpreting these results. First, the single-center, retrospective design may limit the generalizability of the findings. Variations in clinical practice patterns, imaging acquisition and interpretation, and follow-up procedures across institutions could affect model performance, emphasizing the need for external validation in multicenter cohorts. Second, the study focused on recurrence within the first year of stroke, and longer follow-up is needed to assess how well these models predict late recurrent events. Third, several potentially informative predictors, including genetic markers, lifestyle factors (e.g., smoking status and physical activity), and detailed data on medication adherence, were not included due to unavailability in electronic medical records. Incorporating these variables in future studies may further improve predictive accuracy. Fourth, while the XGBoost model demonstrated strong performance, the restricted interpretability typical of “black box” models may hinder clinical acceptance without robust explanation frameworks and prospective assessment. Fifth, and importantly, antithrombotic medications were inadequately characterized. The “antiplatelet therapy” was captured only as a binary variable and did not distinguish between single or dual regimens. Crucially, anticoagulant use, which is a critical determinant of recurrence prevention in patients with atrial fibrillation, was not consistently available. The absence of this key confounder likely affected the model’s performance and should be addressed in future studies.

Despite these limitations, this study advances our understanding of ML-based stroke recurrence prediction by demonstrating the benefit of integrating routine clinical variables with imaging-derived data. The XGBoost model demonstrated high discriminative performance and consistent outcomes across subgroups, indicating potential application for supporting personalized secondary prevention strategies. However, the single-center, retrospective design and the lack of external validation remain significant limitations and may restrict generalizability. The lack of external validation in diverse, multi-center cohorts represents a significant limitation, potentially affecting the generalizability of our model. Future studies should prioritize external validation to ensure robustness across different patient populations, imaging protocols, and clinical workflows. Furthermore, restricting outcomes to a 1-year recurrence window does not capture late recurrent events, and longer follow-up would strengthen the clinical relevance of the model. Future studies should focus on external validation, incorporating additional predictive variables (such as lifestyle, adherence, and other biologically informative predictors), and develop practical, user-friendly tools to facilitate implementation in routine clinical care.

In summary, machine learning algorithms that integrate routine clinical variables with imaging-derived features can effectively predict stroke recurrence risk, with the XGBoost model offering the highest overall performance. Infarct size, hypertension, and fasting blood glucose were identified as most influential predictors, underscoring the importance of structural neuroimaging and rigorous management of metabolic and vascular risk factors in secondary prevention. These findings support the use of ML-based models as adjuncts to clinical decision-making, with the potential to improve outcomes by facilitating more targeted risk reduction approaches.

This study demonstrates that machine learning algorithms integrating routine clinical data and imaging features can predict stroke recurrence risk effectively, with the XGBoost model achieving the highest overall performance. The key predictors, particularly infarct size and a history of hypertension, underscore the significance of structural brain injury and vascular-metabolic dysregulation in driving recurrence risk. Robust performance across cortical, subcortical, and posterior circulation infarct subgroups further supports the model’s potential clinical utility in diverse stroke subtypes with distinct pathophysiological mechanisms.

• Machine learning models, particularly XGBoost, that integrate both routine clinical and imaging-derived features demonstrate a higher predictive performance for stroke recurrence risk than traditional models.

• Infarct size, a history of hypertension, and fasting blood glucose levels were identified as the most influential predictors of recurrence.

• The XGBoost model maintained robust predictive performance across different stroke subtypes defined by infarct location.

• This study highlights the potential of applying advanced analytical methods and multimodal data for enhancing risk stratification and supporting personalized secondary prevention strategies in stroke survivors.

The datasets analyzed during the current study are available from the corresponding author on reasonable request.

LG designed the study. STW and JLL analyzed the data. MKZ performed the study. LG drafted the manuscript. All authors contributed to important editorial changes in the manuscript. All authors read and approved the final version of the manuscript. All authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

This study was conducted in accordance with the Declaration of Helsinki. The protocol was approved by the Ethics Committee of The Fifth People’s Hospital of Jinan (Approval No. 25-5-16). Informed consent was waived because the study used retrospective, de-identified data from electronic medical records, which involved no intervention or risk to patients. This meets the criteria for waiving informed consent as specified in Article 39 of The Regulations of Ethical Reviews of Biomedical Research Involving Human Subjects of China, which states that retrospective studies using anonymized data with minimal risk to privacy do not require informed consent.

Not applicable.

This research received no external funding.

The authors declare no conflict of interest.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/BJHM50394.