Optimization of Imbalanced Tuberculosis Data Classification Using Cost-Sensitive Binary Logistic Regression

Abstract

Tuberculosis (TB) remains a major public health challenge in Indonesia, particularly in urban areas. This study aims to optimize the classification of TB case predictions by comparing three binary logistic regression approachesard binary logistic regression, cost-sensitive binary logistic regression, and SMOTE-based binary logistic regression. The dataset consists of 5,180 patient samples obtained from a health foundation. Initial analysis reveals a significant class imbalance, with TB negative cases dominating the data, while TB-positive cases are relatively scarce. The standard binary logistic regression model demonstrates weak predictive performance for positive cases; out of 195 TB-positive cases, only 4 were correctly identified, while 191 were misclassified as negative, posing a high risk in real-world implementation.

Conversely, the cost-sensitive binary logistic regression approach assigns higher weights to the minority class to reduce bias caused by class imbalance. The class weights are determined based on the inverse class frequency using the formula Based on the distribution of the training dataset, which consists of 3.175 negative cases and 451 positive cases, the resulting weights are approximately and The application of this weighting scheme improves the model's ability to detect positive cases, with 76 cases correctly classified, particularly in the context of low public disclosure regarding health conditions. The SMOTE-based binary logistic regression model achieves a higher recall, detecting 82 positive cases; however, the use of synthetic data introduces potential concerns regarding predictive validity. Overall, the cost-sensitive model achieved a recall of 39%, an F1-score of 32%, and an overall accuracy of 79%, with higher AUC-ROC and AUC-PR values compared to the baseline model. Although the improvement in recall remains moderate at 39%, the cost-sensitive approach shows potential in enhancing the model’s ability to detect positive cases. Therefore, this approach may be considered as a supporting method in efforts to improve more targeted TB control strategies in Indonesia.