<![CDATA[Data imbalance and scale differences between features are often the main factors that reduce the performance of neural network-based classification models. This study aims to analyze the effect of data preprocessing and hyperparameter tuning on the performance of Backpropagation Neural Network (BPNN) in stroke classification. This study used a stroke dataset from the Kaggle platform consisting of 5,110 patient data with 10 clinical features. The evaluation was conducted using five schemes and consisted of several data balancing techniques. These techniques include no balancing, SMOTE, and ADASYN. In addition, the evaluation also involved data normalization including no normalization, MinMaxScaler, and Z-Score. The BPNN model used has an architecture of 19 input neurons, 29 neurons in the hidden layer, and 1 output neuron. Hyperparameter tuning was performed by finding the best learning rate and number of epochs. The evaluation results showed that the model in scheme one has limitations. This limitation is most visible in identifying stroke classes. The application of SMOTE and MinMaxScaler in scheme two proved that the results were better and its performance increased significantly. On the other hand, the combination of ADASYN and Z-Score in scheme three showed more stable performance and was able to detect stroke cases more accurately. The hyperparameter tuning process in schemes four and five also proved to improve performance. The best results were obtained in scheme five, with an accuracy of 96.47%, a precision of 97.34%, a recall of 95.62%, and an F1-score of 96.47%. These findings indicate that the combination of adaptive balancing techniques, distribution-based normalization, and optimal parameter tuning is very effective in improving the accuracy and stability of BPNN for stroke classification.]]>
Copyrights © 2026
