Advancing Tabular Stroke Modelling Through a Novel Hybrid Architecture and Feature-Selection Synergy

Technical Summary: Advancing Tabular Stroke Modelling Through a Novel Hybrid Architecture and Feature-Selection Synergy

Problem Statement
Brain stroke remains a leading cause of global mortality and disability, yet existing predictive models utilizing tabular data frequently fail to surpass the 95% accuracy threshold, limiting their real-world clinical utility. Current challenges include the inherent complexity of stroke pathophysiology, where risk factors interact in non-linear ways that traditional linear models (e.g., logistic regression) struggle to capture. Furthermore, stroke datasets suffer from severe class imbalance (typically ~5% positive cases), missing values, outliers, and inconsistent feature selection strategies. These issues often lead to models with poor sensitivity for the minority class (stroke occurrence) and a lack of interpretability, hindering their adoption in diverse healthcare environments.

Methodology
The study proposes a comprehensive, data-driven framework designed to predict stroke risk using a public cohort of 4,981 records containing 11 features (demographic, lifestyle, and clinical variables). The methodology follows a rigorous pipeline:

1. Data Preprocessing:

   • Cleaning: The dataset underwent outlier detection and removal using the Interquartile Range (IQR) method, reducing the dataset from 4,981 to 4,337 records by eliminating extreme values in glucose levels and BMI.
   • Encoding: Categorical variables were transformed using Label Encoding to maintain ordinal relationships where appropriate (e.g., smoking status) while ensuring computational efficiency.
   • Imbalance Correction: The Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training set to balance the class distribution (from 5% stroke/95% non-stroke to 50/50), preventing model bias toward the majority class.

2. Feature Selection:
   Two complementary strategies were employed to identify predictive variables:

   • Correlation-Based: Utilized point-biserial correlation to identify linear relationships, selecting features like age, hypertension, and heart disease.
   • Tree-Based: Employed Random Forest Gini importance to capture non-linear interactions and feature synergies, highlighting average glucose level, BMI, and age as critical predictors.
   • Three distinct feature sets were created for modeling: Full Features (10 variables), Correlation-based Features (7 variables), and Random Forest-based Features (7 variables).

3. Model Development and Optimization:

   • Base Classifiers: Ten diverse algorithms were implemented, including tree ensembles (Decision Tree, Random Forest), boosting methods (Gradient Boosting, XGBoost, LightGBM, AdaBoost), kernel methods (SVC), linear models (Logistic Regression), distance-based methods (KNN), and neural networks (MLP).
   • Hyperparameter Tuning: All models were optimized using GridSearchCV with stratified 5-fold cross-validation, targeting the maximization of ROC-AUC to address class imbalance.
   • Ensemble Architecture: A hybrid stacking ensemble was constructed. Base models (including Random Forest, XGBoost, LightGBM, and SVC) generated probability predictions, which served as meta-features for a Logistic Regression meta-learner. This approach was designed to leverage the complementary inductive biases of different algorithm families.

Key Results
The study evaluated performance across the three feature sets using accuracy, precision, recall, F1-score, and ROC-AUC:

• Individual Models: Among single classifiers, LightGBM consistently outperformed others, achieving a peak accuracy of 91.36% (Full Features) and 91.24% (Correlation-based). Other models like SVM and AdaBoost showed significantly lower performance (below 74%).
• Ensemble Performance: The proposed stacking ensemble significantly outperformed all individual models.
  • Full Features Ensemble: Achieved 97.20% accuracy, a 97.15% F1-score, and a 99.66% ROC-AUC.
  • Correlation-Based Ensemble: Achieved 95.56% accuracy and 98.95% ROC-AUC.
  • Random Forest-Based Ensemble: Achieved 95.65% F1-score.
• Comparison: The proposed model surpassed the best single learner (LightGBM at ~91.4%) and recent state-of-the-art benchmarks (e.g., 96.59% from a dense stacking ensemble in prior work).

Significance and Claims
The authors claim that this study establishes a new benchmark for tabular stroke prediction on widely used open datasets. The primary significance lies in demonstrating that a disciplined pipeline—combining rigorous preprocessing, dual feature-selection strategies, and a heterogeneous stacking ensemble—can convert low-cost, routinely collected tabular data into a "near-clinical-grade" risk assessment tool.

Key takeaways emphasized by the authors include:

• Synergy of Methods: The success of the hybrid architecture validates the hypothesis that combining learners with diverse inductive biases (tree-based, kernel-based, and linear) yields superior generalization compared to single models.
• Clinical Relevance: The model relies solely on demographic, historical, and basic metabolic variables, making it suitable for low-resource settings where advanced imaging is unavailable.
• Interpretability: Despite the complexity of the ensemble, the feature importance rankings aligned with established epidemiological risk factors (age, hypertension, BMI), lending clinical face-validity to the data-driven findings.

Limitations and Future Directions
The paper maintains a modest tone regarding immediate clinical deployment. The authors acknowledge that the study relies on a single, cross-sectional cohort, meaning geographic and ethnic transportability remains untested. They note that the small size of the minority class in the untouched test set raises concerns about calibration drift in rarer subgroups. Additionally, while SMOTE was essential for training, the authors caution that synthetic oversampling may inadvertently amplify noise. Future work is proposed to focus on prospective, multi-center validation, the integration of longitudinal data, and the use of explainability tools (e.g., SHAP) to ensure transparency in clinical decision-making.