Optimasi Prediksi Penyakit Asma Menggunakan Improved LightGBM Berbasis Bayesian Optimization dengan Hybird SMOTE-ENN dan SHAP Feature Selection

Abstract

Asthma is one of the most prevalent chronic respiratory diseases worldwide, affecting more than 300 million people, and its early prediction is essential for timely clinical intervention. A major obstacle in data-driven asthma prediction is the severe class imbalance of large-scale clinical datasets, which biases conventional classifiers toward the majority (non-asthma) class. This study proposes an Improved LightGBM that integrates three components: Hybrid SMOTE-ENN to correct class imbalance and remove noisy boundary samples, SHAP-based feature selection to retain the most informative attributes, and Bayesian Optimization for hyperparameter tuning. A Kaggle-derived asthma dataset (409,216 SMOTE-balanced training records and 59,672 test records over 23 encoded clinical features) was used. Hybrid SMOTE-ENN reduced a 40,000-sample working set to 8,322 cleaned, balanced instances; SHAP selected 13 of 23 features; and Bayesian Optimization produced the optimal configuration (best cross-validation accuracy 92.20%). On the balanced hold-out test set the proposed Improved LightGBM achieved an accuracy of 93.87%, precision of 0.9447, recall of 0.9359, F1-score of 0.9403, and ROC-AUC of 0.9839, clearly surpassing the LightGBM Bayesian-Optimization baseline reported in the main reference (78% accuracy, ROC-AUC 0.975). Evaluation on the original imbalanced test distribution (accuracy 72.83%, ROC-AUC 0.6392) transparently reflects the difficulty of severely imbalanced real-world clinical data. The results show that combining Hybrid SMOTE-ENN, SHAP feature selection, and Bayesian Optimization yields a more accurate, interpretable, and discriminative asthma-prediction model