Enhancing Stroke Prediction Using LightGBM With SMOTE-ENN and Fine-Tuning: A Comprehensive Analysis

Introduction: Accurately classifying stroke cases is a significant challenge in health care, as early detection can reduce severe complications and improve outcomes. Stroke datasets are usually imbalanced, with non-stroke cases in the majority, which poses a challenge to traditional machine learning algorithms and usually results in low stroke detection rates. This research proposes an advanced approach using Light Gradient-Boosting Machine (LightGBM) with Synthetic Minority Over-sampling Technique-Edited Nearest Neighbors (SMOTE-ENN) to address this imbalance. Further optimization was performed using RandomizedSearchCV with LightGBM, achieving an area under the curve (AUC) of 0.946, higher than any of the baselines. Methods: Class imbalance needed to be addressed first using SMOTE-ENN, which combines SMOTE with ENN for the creation of a balanced training set to be generalized well. We then used the LightGBM algorithm, which works quite efficiently with large datasets, and optimized it with RandomizedSearchCV. Results: The highest performance of LightGBM was achieved with an AUC of 0.946, improving precision, recall, F1 score, and receiver operating characteristic curve performance. Therefore, LightGBM demonstrates high sensitivity and specificity for stroke detection compared to baseline models. Conclusion: The integration of SMOTE-ENN with LightGBM and extensive hyperparameter optimization provides a robust framework for predicting stroke in imbalanced datasets. This approach not only enhances model performance but also proves to be a potential solution for other medical prediction problems significantly affected by class imbalance.