This study investigates public sentiment on the Mobile JKN application using Logistic Regression enhanced with SMOTE-based statistical feature selection. Unlike prior works that relied solely on conventional feature combinations such as TF-IDF or Word2Vec, this research performs a comparative evaluation of three statistical feature selection techniques: Recursive Feature Elimination (RFE), Chi-Square, and Mutual Information, under both TF-IDF and Word2Vec representations in a low-resource Indonesian language setting. The dataset consists of 2,382 user reviews from the Google Play Store, balanced using SMOTE to mitigate class imbalance. The best configuration, TF-IDF combined with Mutual Information, achieved an accuracy of 73.38% and an F1-score of 50%, indicating a moderate yet consistent performance. A confusion matrix-based error analysis revealed that most misclassifications occurred between neutral and negative classes due to semantic overlap. The relatively low F1-score highlights challenges in sentiment separability, while the superior performance of Mutual Information demonstrates its ability to capture discriminative linguistic features. The superior performance of Mutual Information is attributed to its ability to capture non-linear dependencies between features and sentiment labels, yielding richer discriminative information compared to Chi-Square or RFE. This research establishes a comparative methodological framework that integrates feature selection and data balancing techniques, providing interpretable sentiment classification insights for under-resourced language settings.