Prediksi Stroke Menggunakan Metode NODE dengan Optimasi Hyperparameter

Across the world, stroke contributes substantially to mortality and disability, making the development of early detection systems based on clinical data highly important. This study aims to build a stroke risk prediction model using the Neural Oblivious Decision Ensembles (NODE) method with three hyperparameter optimization techniques: Grid Search, Random Search, and Bayesian Optimization. The dataset used is the Stroke Prediction Dataset from Kaggle, consisting of 5,110 patient records. All data underwent pre-processing, feature selection using the Pearson correlation method, and class balancing with SMOTE applied to the training set to ensure a more balanced data distribution.
This study tested various combinations of node count, layer dimension, and learning rate to determine the most optimal configuration of the NODE model. Evaluation was carried out using the metrics accuracy, AUC (Area Under Curve), confusion matrix, and execution time. Experimental results show that the NODE model optimized with Random Search achieved the best performance, with 90.53% accuracy and an AUC of 76.20%, followed by Grid Search (85.26%; 76.46%) and Bayesian Optimization (87.27%; 78.29%). In terms of efficiency, Bayesian Optimization achieved the fastest execution time, averaging 810.65 seconds, compared to 940.70 seconds for Grid Search. These results demonstrate that applying the NODE method with an appropriate hyperparameter optimization strategy can significantly improve both the accuracy and efficiency of stroke prediction. The Random Search approach proved most effective in identifying the optimal parameter combination, while Bayesian Optimization excelled in computational efficiency. This study is expected to provide benefits to the development of more adaptive and accurate deep learning models for predicting medical conditions using tabular data.