In this study, five machine learning algorithms were implemented in order to evaluate their predictive capability for stroke risk prediction. Each model was trained using its optimized hyperparameters, obtained through the tuning process, to ensure that the comparison between models reflects their best achievable performance on the dataset.

The models used in the experiments include:

I. XGBoost [11] II. LightGBM [12] III. CatBoost [13] IV. Support Vector Machine (SVM) [14] V. Logistic Regression [15] The selected algorithms represent two major categories of machine learning methods. The first category consists of gradient boosting–based ensemble models, including CatBoost, LightGBM, and XGBoost. These models construct predictive functions by sequentially combining multiple weak learners, typically decision trees, allowing them to capture complex nonlinear relationships and feature interactions within structured datasets.

The second category includes classical machine learning algorithms, namely SVM and Logistic Regression. These models were included as baseline methods to provide a comparative perspective and to evaluate whether more advanced ensemble approaches can achieve improved predictive performance in the context of stroke prediction.

The performance of the implemented models was assessed using five commonly used evaluation metrics in classification tasks:

 Accuracy  Precision  Recall  F1-score  ROC–AUC

To obtain reliable performance estimates and reduce the effect of randomness in the training–testing split, the experiments were conducted using Stratified 5-Fold Cross Validation [16]. In this validation strategy, the dataset is divided into five folds while preserving the distribution of the target classes. Each fold is used once as a validation set while the remaining folds are used for training.

The final performance values reported in the study correspond to the average results obtained across the five folds.

The performance results obtained from the 5-fold cross-validation experiments are presented in Table 1. This table summarizes the average values of Accuracy, Precision, Recall, F1-score, and AUC for all evaluated models [17].

As illustrated in Table 1, the three gradient boosting algorithms achieved consistently higher performance across nearly all evaluation metrics compared with the traditional classification models.

The experimental results presented in Table 1 reveal several important observations regarding the comparative performance of the evaluated machine learning models.

First, the boosting-based models (CatBoost, LightGBM, and XGBoost) clearly outperform the classical algorithms. These models achieved significantly higher values in terms of Accuracy, Precision, F1-score, and AUC, demonstrating their strong capability to capture complex relationships within the dataset.

Among all evaluated algorithms, LightGBM achieved the highest Accuracy (0.985185) and F1-score (0.985402). These results indicate that LightGBM provides the most balanced predictive performance among the tested models.

The CatBoost model achieved the highest AUC value (0.999646), indicating an almost perfect ability to discriminate between stroke and non-stroke instances. Such a high AUC reflects the strong discriminative capacity of the model in distinguishing between the two classes across different decision thresholds.

Similarly, XGBoost demonstrated competitive performance, achieving results very close to those of LightGBM and CatBoost. The consistency of high scores across all metrics confirms the effectiveness of gradient boosting algorithms for structured medical datasets.

Another significant observation is that the Recall value for all boosting models equals 1.000000. This indicates that all actual stroke cases in the validation folds were successfully identified by these models. In the context of medical prediction systems, this behavior is particularly desirable because it minimizes the risk of failing to detect patients who may suffer from stroke.

In contrast, SVM and Logistic Regression showed noticeably weaker performance. Although SVM achieved moderate results, its accuracy and AUC values remain substantially lower than those of the boosting models. Logistic Regression produced the lowest performance among all evaluated models, suggesting that linear modeling approaches may not adequately capture the complex patterns present in the dataset.

To further evaluate the effectiveness of the proposed framework, the obtained results were compared with several recent studies in the literature. The comparison is summarized in Table 2, which presents the methods and performance metrics reported by previous research works alongside the results achieved in this study. As shown in Table 2, the proposed framework achieves highly competitive results compared with previously published methods. In particular, the proposed approach achieved 100% recall, indicating that all stroke cases were correctly detected in the experimental evaluation.

Although some previous works report slightly higher accuracy values, the proposed model demonstrates a strong balance between precision, recall, and F1-score, while maintaining perfect recall for the minority class.

The primary contribution of this research lies in the development of a complete predictive framework for stroke risk assessment, integrating several stages including data preprocessing, class balancing, feature selection, and machine learning modeling.

The proposed methodology introduces a hybrid feature selection strategy designed to identify the most informative and stable features. Instead of relying on a single importance measure, the proposed method computes a final feature score based on the combination of three complementary criteria:

I. SHAP feature importance II. Statistical correlation with the target variable III. Feature stability across cross-validation folds [8].

These criteria are combined into a unified score:

Final⁡Feature⁡Score⁡=⁡α⁡×⁡SHAP⁡+⁡β⁡×⁡Correlation⁡+⁡γ⁡×⁡Stability.

The optimal weights assigned to these three components are automatically determined using the Optuna optimization framework, which searches for the combination that maximizes the cross-validated predictive performance of the models.

Unlike traditional feature selection techniques that rely on a single metric, this hybrid approach simultaneously considers model interpretability, statistical relevance, and robustness across validation folds. As a result, the selected feature subset provides both stable and highly informative predictors for stroke classification

In this study, a comprehensive machine learning framework was developed for stroke risk prediction using structured clinical data. The proposed pipeline integrates systematic data preprocessing, class balancing, hybrid feature selection, and a comparative evaluation of multiple machine learning models.

The experimental results demonstrate that gradient boosting algorithms—particularly LightGBM and CatBoost—outperform classical models such as SVM and Logistic Regression. Due to the nonlinear relationships among clinical variables and the presence of class imbalance in medical datasets, gradient boosting models are better suited to capture complex feature interactions and learn predictive patterns effectively.

A key outcome of this research is achieving 100% Recall in identifying stroke patients. Although some previous studies have reported slightly higher overall accuracy values, the proposed model prioritizes Recall, which is critically important in medical decision‑making. High recall indicates the model’s strong ability to correctly identify actual stroke cases and significantly reduce false negatives, which is essential in healthcare applications where missing a true patient may lead to severe consequences.

In addition to its perfect recall, the proposed model maintains a high Precision (97.12%) and achieves a strong F1‑score, demonstrating a well-balanced trade‑off between correctly identifying stroke patients and minimizing false alarms.

Furthermore, the proposed hybrid feature selection framework, optimized using Optuna [24], improves model robustness by selecting features that are not only highly informative but also stable across different cross‑validation folds. This approach enhances both the reliability and generalization capability of the predictive model.

Overall, the findings suggest that the proposed methodology—combining intelligent feature selection with advanced gradient boosting techniques—provides an effective and reliable approach for stroke risk prediction systems and has strong potential for supporting clinical decision‑making in healthcare environments.