The Framingham Risk Score (FRS), rooted in the landmark Framingham Heart Study, one of the most influential longitudinal investigations in cardiovascular epidemiology, offers a foundational statistical model for estimating an individual’s 10-year risk of cardiovascular events, including stroke. This model leverages a combination of clinical and demographic variables such as age, systolic blood pressure, diabetes status, smoking behavior, and body mass index (BMI) to assess risk (29). Its integration of both modifiable and non-modifiable risk factors underscores its utility in guiding stroke prevention efforts.

Over time, the FRS has gained widespread validation and has been embedded within clinical guidelines across diverse healthcare systems (23, 30). Despite its prominence, a central critique of the FRS lies in its reliance on additive linear assumptions, which can obscure the complex and often non-linear interactions among risk factors. This shortcoming is particularly pronounced in heterogeneous populations, where traditional models may fail to capture nuanced relationships, precisely the kind of complexity that machine learning (ML) techniques are well-equipped to address.

In this study, the FRS informs the initial selection of predictive features and establishes a reference point for evaluating the performance of ML-based models. Its inclusion is especially pertinent given the well-documented contributions of variables such as BMI, age, hypertension, and diabetes to stroke risk, variables repeatedly validated across traditional and emerging analytical paradigms (31). Furthermore, the FRS highlights the relevance of occupational stress, which recent research has linked to increased cardiovascular burden, particularly in high-demand work environments (32).

As machine learning models are deployed to augment and refine the predictive power of the FRS, this framework serves not only as a benchmark for comparison but also as a theoretical scaffold. It ensures the clinical validity of selected features while facilitating a deeper understanding of how non-linear modelling approaches can enhance stroke prediction. In this way, the FRS provides a vital conceptual bridge between classical epidemiological methods and the evolving capabilities of contemporary data science.

The Atherosclerosis and Thrombosis Theory offers a critical biomedical framework for understanding the pathogenesis of stroke, focusing on the physiological mechanisms that underlie vascular occlusion. Central to this theory is atherosclerosis, a progressive condition marked by the accumulation of lipids, inflammatory cells, and fibrous elements within arterial walls. Over time, these deposits form plaques that can rupture, triggering thrombotic events and ultimately leading to ischemic stroke, the most common stroke subtype (33).

This theoretical perspective is particularly salient for informing the inclusion of clinical variables such as cholesterol levels, blood glucose, and body mass index (BMI) in predictive modelling. Both hypercholesterolemia and hyperglycaemia are implicated in endothelial dysfunction and plaque destabilization, while obesity promotes systemic inflammation, a driving force in atherogenesis (34). Empirical evidence further supports this connection, demonstrating that elevated low-density lipoprotein (LDL) levels and increased BMI are strongly associated with heightened stroke risk (Gao et al., 2021).

Importantly, the Atherosclerosis and Thrombosis Theory also intersects with behavioral and lifestyle risk factors. Smoking, for instance, has been shown to exacerbate oxidative stress and inflammation, accelerating arterial damage and thrombogenesis (Ambrose and Barua, 2004). Similarly, sedentary behavior contributes to metabolic dysregulation and vascular dysfunction. These biologically plausible pathways enrich the theoretical underpinnings of stroke risk, extending beyond statistical associations to include mechanistic insight, thus complementing the epidemiological focus of the Framingham Risk Score.

Incorporating this theory into the present study strengthens the clinical rationale behind variable selection and lends credibility to the model’s interpretability. By anchoring machine learning outputs in well-established physiological processes, the study ensures that its findings are not only statistically sound but also biomedically coherent. This integration deepens the explanatory power of the predictive framework, bridging the gap between data-driven insights and clinical relevance.

The Metabolic Syndrome Hypothesis conceptualizes stroke because of interrelated metabolic abnormalities, including central obesity, insulin resistance, hypertension, hyperglycaemia, and dyslipidaemia, that collectively elevate cardiovascular risk (35). Its relevance has intensified amid rising urbanization and lifestyle shifts, especially in developing regions where increased metabolic syndrome prevalence parallels stroke incidence (36). Core components such as elevated BMI and impaired glucose regulation, strongly linked to cerebrovascular outcomes (37), are thus essential for predictive modelling. Within the machine learning context, this framework supports modelling strategies that detect complex, non-linear interactions among co-occurring risk factors, capturing high-risk profiles that conventional models may overlook. Integrating this hypothesis into the present study strengthens both the theoretical grounding and the clinical relevance of selected features, while underscoring the value of ML’s capacity to reflect the multifactorial and synergistic nature of stroke etiology.

Complementing biomedical insights, psychosocial and behavioral theories help contextualize the elevated stroke risk among individuals in private or self-employed roles. The Job Strain Model attributes this to chronic occupational stress leading to cardiovascular dysfunction (38). Similarly, the Biopsychosocial Model reinforces the study’s findings by framing stroke risk as an interplay between physiological factors and social environments, including work and lifestyle. Broader public health perspectives, such as the Social Determinants of Health (SDOH) framework, justify the inclusion of variables like residence type and employment status, while the Health Belief Model emphasizes the need for targeted education to modify behaviors in high-risk groups (39–42). Collectively, these frameworks support the multifactorial understanding of stroke presented in this study and highlight the value of integrated, data-driven approaches in guiding prevention strategies and policy interventions.

The integration of the Framingham Risk Score, Atherosclerosis Theory, and Metabolic Syndrome Hypothesis grounds the study in clinical, biological, and systemic insights, enhancing the relevance and interpretability of machine learning in stroke prediction.

Stroke prevention is a multidimensional process that encompasses a continuum of interventions, primordial, primary, and secondary, designed to mitigate cerebrovascular risk across the lifespan. At the foundational level, primordial prevention aims to eliminate the emergence of risk factors within populations through structural, educational, and behavioral health strategies (30, 81). These include reducing tobacco use, encouraging healthy dietary habits, increasing physical activity, and addressing socioeconomic determinants of health. Primary prevention targets individuals with identifiable risk factors such as hypertension, diabetes, and dyslipidaemias but without prior cerebrovascular events, aiming to prevent the initial occurrence of stroke. Meanwhile, secondary prevention is focused on patients who have experienced a stroke or transient ischemic attack (TIA), utilizing pharmacologic and non-pharmacologic strategies to reduce the risk of recurrence (82). These levels are interconnected; effective population-level prevention supports individual risk modification, and tailored clinical management reinforces broader public health goals.

Accurate risk assessment is a prerequisite for effective stroke prevention and management. This involves both population-based tools and individual clinical evaluations. For general cardiovascular risk estimation, tools such as the Framingham Stroke Risk Profile (83) and the ASCVD Risk Estimator (73) are widely validated. These integrate age, sex, systolic blood pressure, diabetes, smoking status, cholesterol levels, and the presence of cardiovascular comorbidities. In specific populations, particularly those with atrial fibrillation (AF), risk stratification models such as the CHA₂DS₂-VASc score are used to determine thromboembolic risk and guide anticoagulation therapy (84). Clinical evaluation must also include routine blood pressure monitoring, lipid profiles, glucose/HbA1c testing, and anthropometric measurements such as body mass index (BMI) and waist-to-hip ratio. The dynamic nature of stroke risk, affected by aging, comorbid conditions, and medication adherence, necessitates periodic reassessment to refine preventive strategies over time.

The effective management of modifiable stroke risk factors remains the most actionable and evidence-based component of stroke prevention, particularly in secondary prevention settings. Hypertension, the most critical determinant of stroke risk, must be addressed aggressively; even modest reductions in blood pressure (e.g., 5–6 mm Hg diastolic) have been shown to reduce stroke incidence by up to 40% (85). Similarly, while strict glycaemic control has not consistently demonstrated reductions in macrovascular events in diabetic patients, its benefits for microvascular outcomes and overall cardiovascular risk justify ongoing treatment (86). Lipid-lowering therapy, particularly with statins, has a strong evidence base, showing a 25–30% reduction in stroke risk among high-risk populations, including those with previous cardiovascular disease (87). Lifestyle interventions, smoking cessation, alcohol moderation, structured physical activity, and dietary changes, are equally critical. Smoking cessation alone halves stroke risk within 5 years (88), quitting smoking after a stroke or TIA markedly lowers the risk of recurrence and mortality compared to continued smoking (89), while moderate-intensity exercise is associated with a significantly reduced risk of ischemic stroke (90). These interventions must be sustained and supported by structured counselling, patient education, and multidisciplinary care models.

Beyond lifestyle and medical risk factor control, targeted pharmacological interventions play a pivotal role in reducing stroke recurrence, particularly in high-risk patients. Antiplatelet agents, including aspirin and clopidogrel, are central to secondary prevention in patients with non-cardioembolic ischemic stroke or TIA. Meta-analyses confirm that aspirin, in doses of 75–325 mg daily, reduces the risk of recurrent stroke by approximately 25% (91). Clopidogrel provides a modest additional benefit and is particularly recommended in aspirin-intolerant individuals (92). In cardioembolic stroke, especially due to atrial fibrillation, anticoagulant therapy is essential. Trials such as EAFT and subsequent analyses show that warfarin reduces stroke risk by 68–70% (84, 93). More recently, direct oral anticoagulants (DOACs) such as apixaban and rivaroxaban have demonstrated superior safety profiles with comparable or greater efficacy (94). In certain cases, procedural interventions, such as carotid endarterectomy (CEA) in asymptomatic patients with significant stenosis, may be appropriate, though only when operative risk is low and life expectancy exceeds 5 years (95). Ultimately, prevention requires an integrated framework of clinical vigilance, evidence-guided decision-making, and patient-centred care.

Despite the substantial advances in traditional stroke prevention and management, including pharmacological therapies, lifestyle interventions, and risk stratification tools such as the Framingham Stroke Risk Profile and CHA₂DS₂-VASc, these approaches remain limited in their ability to capture the complex, dynamic, and multifactorial nature of stroke risk. Traditional models often rely on linear assumptions and static variables, which may inadequately reflect individual patient trajectories, temporal changes in physiology, and nuanced interactions among clinical, genetic, behavioral, and environmental factors (96). Moreover, they underperform in diverse populations and offer limited personalization. In this context, machine learning (ML) presents a transformative opportunity to enhance stroke prediction, diagnosis, and treatment by leveraging large-scale, heterogeneous datasets, such as electronic health records, imaging, genomics, and wearable sensor data, to model non-linear relationships and discover latent risk patterns (97). By continuously learning from new data and adapting to individual-level variation, ML algorithms offer a scalable, data-driven complement to traditional methods, potentially improving accuracy, equity, and clinical decision-making across the stroke care continuum.

LR is widely used in medical research for modelling binary outcomes due to its ability to quantify relationships between outcome variables and various types of predictors, including nominal, ordinal, interval, and ratio-level data (107, 108). Popularly used for binary classification, it estimates the probability of an outcome belonging to a specific class, despite its name suggesting a regression technique (109). It uses the sigmoid (logistic) function to compute probabilities and applies a decision boundary to classify data points (110). Its appeal lies in its interpretability and straightforward implementation. However, LR is limited to modelling linear associations and is susceptible to issues such as multicollinearity and variance inflation, which can reduce predictive accuracy (111, 112). These limitations restrict its effectiveness in capturing the complex, non-linear relationships often present in acute stroke prognosis (108).

The mathematical equation for logistic regression shown in Equation 1 is based on the logistic (sigmoid) function and models the probability that a given input x belongs to a particular class (typically class 1 in binary classification):

Where,

i = 1…N (number of observations).

j = 1…M (number of individual variables).

𝑝𝑖 = predicted probability of a ‘1’ at observation i.

𝛽𝑗= Regression Coefficient.

𝑥𝑖𝑗= The j-th variable at observation i.

Although LR is limited in capturing non-linear interactions, recent studies show it can perform competitively with proper data preprocessing. For instance, one study achieved approximately 86% accuracy using LR by applying data imputation, outlier removal, and the Synthetic Minority Oversampling Technique (SMOTE) to correct class imbalance (113). The model included key predictors such as blood pressure, body mass index (BMI), cardiac history, age, smoking status, glucose levels, and prior stroke events. These findings reinforce that, while ML methods may better capture complex interactions, LR remains a practical, transparent, and effective tool for stroke risk assessment when supported by rigorous data engineering.

Support Vector Machines (SVM) are a robust machine learning method grounded in statistical learning theory, capable of addressing both classification and regression tasks (114). SVMs classify data by identifying an optimal hyperplane that separates data points, represented as n-dimensional vectors, into distinct classes. Through kernel functions, SVMs can effectively manage both linear and non-linear datasets, making them particularly useful for complex medical data (115). In stroke prediction, SVMs have been extensively applied, especially in earlier studies, due to their ability to handle high-dimensional feature spaces and model non-linear relationships. A systematic review by (116) found that SVM was the best-performing model in 10 out of 39 stroke-related studies conducted between 2007 and 2019. These studies demonstrated SVM’s effectiveness in both stroke diagnosis and outcome prediction using clinical and imaging data. While more recent models such as ensemble methods often outperform SVMs in terms of raw predictive accuracy, SVM continues to serve as a strong benchmark model. For instance, Vu et al. (117) included SVM in their comparative analysis of classifiers for stroke incidence prediction in the Suita cohort, where SVM maintained competitive accuracy. Moreover, SVMs are frequently integrated into feature selection workflows due to their capacity to identify support vectors, data points that are most critical to the decision boundary. However, SVM performance is highly sensitive to hyperparameter tuning and data imbalance. As such, strategies like feature scaling, cost-sensitive training, and the use of SMOTE are commonly employed to enhance its utility in stroke datasets (103).

The following terms in Equation 2 can be used to define a support vector classifier:

Where, β0 = bias, S = set of observations, α = model parameters that must be learned.

Support Vector Machines (SVM) handle multiclass classification using one-vs-all or one-vs-one strategies and are widely recognized for their effectiveness in supervised learning tasks, including classification and outlier regression (118).

KNN is a simple, non-parametric, instance-based learning algorithm often used in stroke prediction as a baseline or supporting model. Unlike models that learn a function from the training data, KNN classifies new observations by analysing the majority class among their k nearest neighbors in an n-dimensional feature space, based on similarity metrics such as Euclidean distance (119–121). While its simplicity and intuitiveness make KNN widely accessible, it can perform poorly in high-dimensional or imbalanced datasets due to sensitivity to irrelevant features and noise. Consequently, KNN is rarely the top-performing model in stroke risk prediction but is commonly included in comparative analyses or ensemble methods. For example, (122) used KNN as the final estimator in a stacking ensemble alongside random forest and decision trees, achieving high predictive accuracy (98.6%). In such settings, KNN benefits from the aggregated insights of stronger base learners. Furthermore, techniques like principal component analysis (PCA) can enhance KNN’s performance by reducing feature dimensionality and improving neighbour similarity, mitigating the “curse of dimensionality” often encountered in clinical datasets.

The KNN mathematical equation is described by Equation 3:

The model’s performance will be presented in the results section of this study.

DTs are a supervised learning method that classifies a target variable by recursively learning simple decision rules from input features (123). These rules split variables based on impurity measures (e.g., Gini index, entropy) until a stopping criterion is met (110). Visually, a decision tree resembles an inverted tree, with the root node at the top and branches representing feature-based splits. Decision trees are widely used in stroke prediction for their interpretability, ability to model non-linear relationships, and capacity to handle mixed data types.

Among tree-based models, Random Forests (RF), ensembles of multiple decision trees, have become a dominant approach in stroke risk prediction. RF models improve accuracy by aggregating predictions from many de-correlated trees, reducing overfitting and automatically modelling feature interactions. A systematic review by Asadi et al. (27) found RF to be the top-performing algorithm in 25% of stroke-related ML studies between 2019 and 2023. Studies by Choudhury et al. (2023) and Vu et al. (117) similarly reported RF outperforming models like logistic regression, SVM, and even deep neural networks. Additionally, RF offers insights into feature importance, as demonstrated in (104) explainable-AI study, where SHAP analysis highlighted age, triglyceride levels, and aphasia as key predictors. RF’s explainability and predictive strength make it especially valuable in clinical settings. Gradient boosting variants like XGBoost and LightGBM have also shown strong results, with XGBoost ranking second to RF in multiple studies (104). Collectively, tree-based ensembles, particularly Random Forest, remain a cornerstone of modern stroke prediction research due to their robustness, accuracy, and interpretability.

NB is a probabilistic, supervised learning classifier that assumes conditional independence among features and has been explored in stroke risk prediction for its simplicity and efficiency. Despite its strong independence assumption, NB can perform surprisingly well when features are informative and relatively uncorrelated. In structured health datasets, NB has occasionally outperformed more complex models; for example, a study reviewed in Nature reported NB achieving the highest accuracy (~82%) for early stroke identification, likely due to dataset characteristics and predictor strength (124). Although NB is rarely the leading method in the current era of deep learning, it remains a valuable baseline in comparative studies and is often incorporated into ensemble frameworks (e.g., bagging or voting) to enhance predictive robustness. Its low computational cost and interpretability make it particularly useful in smaller or cleaner clinical datasets (124, 125).

Neural Networks has gained increasing attention in stroke risk prediction, particularly with the availability of large and complex datasets. Models such as Multilayer Perceptron (MLP), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs), including Long Short-Term Memory (LSTM) networks, have been applied to both structured tabular data and heterogeneous sources like clinical records, imaging, and time-series signals. A major strength of deep learning (DL) lies in its capacity to automatically learn feature representations, removing the need for manual feature engineering. Moulaei et al. (126) highlighted DL’s growing role in stroke management, particularly in acute stroke diagnosis and imaging-based detection, due to its speed and effectiveness (104, 127). In stroke risk prediction using clinical data, the debate has centered on whether DL can consistently outperform well-optimized traditional machine learning (ML) models. Until recently, most studies evaluated DL and ML separately, limiting direct comparison (128–131). Addressing this gap, (104) compared eight models (four ML and four DL) on the same dataset, finding that while Random Forest (RF) had the best overall performance, DL models as a group outperformed other ML methods in both accuracy and sensitivity (104). Notably, LSTM networks achieved a sensitivity of 96.15%, and the Feedforward Neural Network (FNN) variant showed the highest specificity and F1-score among DL models (104). Another study cited by Heo et al. (132) reported ~94% accuracy using LSTM on EEG data for stroke prediction, demonstrating DL’s potential with alternative modalities. CNNs, though traditionally used for image analysis, have been adapted for tabular data by encoding clinical features as “image-like” inputs, capturing local feature interactions effectively (133, 134). Despite these advances, DL models are often criticized for their black-box nature, which limits clinical interpretability. To address this, explainability techniques such as SHAP and layer-wise relevance propagation are increasingly applied to interpret predictions (104, 135). Ultimately, DL is emerging as a high-performing approach in stroke risk modelling, particularly when rich and diverse datasets are available.

In the past 5 years, stroke risk prediction has witnessed major advancements through the application of machine learning (ML), with significant improvements in predictive accuracy and methodological sophistication (27). Recent studies have reported model accuracies frequently exceeding 90%, compared to earlier averages around 80%, largely due to enhanced data preprocessing techniques (e.g., SMOTE, SMOTE-ENN), inclusion of broader clinical variables, and the adoption of ensemble and hybrid methods such as stacking (103, 104, 122, 124, 138). Moulaei et al. (104) demonstrated that while Random Forest (RF) remained the top overall performer, deep learning model, particularly LSTMs, achieved superior sensitivity (104). These head-to-head comparisons have provided greater clarity on model selection under standardized conditions. Further, the use of real-time prediction with wearable devices (139), and explorations of transfer and multi-task learning, reflect the field’s shift toward dynamic, personalized risk assessment.

Alongside performance gains, recent ML studies have deepened clinical insight by identifying both established and novel stroke risk factors. Tools such as SHAP have consistently ranked age, hypertension, diabetes, smoking, and atrial fibrillation as primary contributors (117, 124), while also revealing additional predictors like fructosamine, haemoglobin, calcium, skinfold thickness, triglyceride levels, and aphasia, many of which are absent from traditional risk scores (104). This dual role of ML, as a predictive and exploratory tool, has opened new avenues for research and refined risk stratification strategies. Moreover, ML has expanded into specific clinical subdomains such as post-stroke mortality, functional recovery, haemorrhagic complications, cognitive impairment and perioperative stroke risk (122, 132, 140). These applications demonstrate how machine learning methods not only improve general population-level predictions but also support targeted, domain-specific clinical decisions. The growing emphasis on model explainability further strengthens the clinical utility of ML by enhancing interpretability, trust, and translational potential.

In a study by Shoily et al. (141), four ML algorithms, Naive Bayes, J48, K-nearest neighbour, and Random Forest, were applied, with J48, KNN, and Random Forest achieving an impressive 99.8% accuracy, while Naive Bayes reached 85.6%. A social media-based approach in (142) used spectral clustering on tweets, applying Naive Bayes, support vector machine, and probabilistic neural networks (PNN), with PNN performing best at 89.90% accuracy. In (143), a comparative study of nine classifiers found that a boosting model with decision trees achieved the highest recall (99.94%), and Random Forest delivered the best precision (97.33%). The widely used Kaggle stroke dataset was employed in (138) who achieved 82% accuracy using various classifiers. In (144), data from Sugam Multispecialty Hospital, India, showed that ensemble methods and support vector machines provided 91% accuracy, while artificial neural networks trained with stochastic gradient descent exceeded 95%. An EHR-based analysis in Nwosu et al. (145) evaluated neural networks, decision trees, and Random Forests, with accuracies of 75.02, 74.31, and 74.53%, respectively. Lee et al. (129) explored ML for analysing diffusion-weighted imaging (DWI) and fluid-attenuated inversion recovery (FLAIR) images within 24 h of symptom onset, using logistic regression, support vector machines, and Random Forests to estimate stroke onset time, demonstrating comparable or superior performance to human interpretation based on sensitivity and specificity.

Govindarajan et al. achieved 95% accuracy using ANN and SGD on data from 507 patients. Amini et al. used C4.5 and KNN on 807 subjects with 50 risk factors, attaining 95 and 94% accuracy, respectively (146, 147). Cheng et al. (148) employed two ANN models for ischemic stroke prognosis with up to 95% precision. Cheon et al. (128) used a deep neural network and PCA on 15,099 patients, reaching an AUC of 83%. Singh & Choudhary (149) applied decision trees, PCA, and neural networks on the CHS dataset, achieving 97% accuracy. Chin et al. (150) developed a CNN-based system for early ischemic stroke detection using 256 images, achieving 90% accuracy. Sung et al. (151) used linear regression and data mining on 3,577 cases to build a stroke severity index, with KNN yielding the best results. Monteiro et al. (152) predicted ischemic stroke outcomes three months post-admission with an AUC above 90%. Kansadub et al. (153) found Decision Tree and Naive Bayes most accurate among tested models. Adam et al. (154) compared KNN and Decision Tree for classifying ischemic stroke, favoring the latter for clinical use.

Despite significant advancements in machine learning (ML) for stroke risk prediction, several challenges remain unresolved in the literature. Chief among these is class imbalance, limited external validation, and the trade-off between predictive performance and interpretability. Stroke datasets often exhibit a skewed distribution, with relatively few positive cases, leading to biased models if left uncorrected. While recent studies have adopted oversampling strategies, not all rigorously evaluate their impact. The complexity of deep learning models has also raised concerns around explainability and clinical trust, particularly in high-stakes applications like stroke prevention (27, 124).

In response, this study addresses several of these concerns through a multi-faceted approach. To mitigate class imbalance, Random Over-Sampling was employed to increase the representation of stroke cases during model training. This straightforward yet effective technique allowed for balanced learning without introducing synthetic data. This approach aligns with best practices in interpretable ML, emphasizing domain-informed model design. The model ensemble incorporated interpretable algorithms, such as logistic regression and KNN, alongside more complex classifiers like random forests, SVM and gradient boosting, allowing for both performance and transparency. This work also addresses the often-overlooked inclusion of lifestyle-related variables in stroke prediction models. By incorporating features like smoking history and BMI, elements typically missing in models relying solely on clinical variables, our analysis enhances the model’s relevance to real-world prevention strategies. Feature importance was evaluated using built-in model metrics, offering insights into key risk factors without relying on black-box techniques. Importantly, our feature selection was guided by a theoretical framework rooted in clinical and epidemiological evidence, ensuring variables like age, blood pressure, BMI, diabetes, and smoking were chosen for their known relevance to stroke risk. By including clinical, demographic, and behavioral variables, and benchmarking performance across diverse models, this work contributes to the development of robust, interpretable, and clinically meaningful stroke prediction tools (27, 103).

To address these gaps, this study investigates:

The accuracy, the overall correctness of the classifier, measures the proportion of correctly predicted instances (both positives and negatives) among all observations. It is calculated as shown in Equation 4:

where TP = true positives, TN = true negatives, FP = false positives, and FN = false negatives.

While intuitive, accuracy can be misleading in imbalanced datasets (174).

The F1-score is the harmonic mean of precision and recall, providing a balanced metric in the presence of class imbalance. It is computed as shown in Equation 5:

Where Precision (the portion of true stroke prediction among all positive predictions) is shown in Equation 6:

Recall (sensitivity), the model’s ability to identify actual stroke cases, crucial in medical diagnosis is shown in Equation 7:

This measure is especially useful when false negatives and false positives carry similar costs (156).

Specificity (true negative rate) quantifies the proportion of actual negatives correctly classified as described in Equation 8:

It is essential in medical diagnostics to avoid false positives as it carries significant consequences (175).

AUROC evaluates a model’s ability to distinguish between classes at various thresholds. This identifies the actual stroke cases, crucial in medical diagnosis. Values range from 0.5 (no discrimination) to 1 (perfect discrimination), summarizing model performance across all classification thresholds (176).

A confusion matrix is a performance summary table showing true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). It supports the computation of key classification metrics and helps evaluate specific types of errors (177).

These metrics were selected based on their relevance to clinical risk classification, where false negatives (undetected stroke risk) are more critical than false positives.