Mental health is an essential aspect for university students, as undetected mental health disorders can have a significant impact on students' academic performance and well-being. This study contributes by evaluating Synthetic Minority Oversampling Technique (SMOTE)'s role in improving classification models' performance. Despite the increasing use of machine learning in mental health detection, limited research has addressed the challenges posed by imbalanced datasets, particularly in smaller student populations. This research aims to develop a mental health early detection system based on student data from Multi Data University Palembang using the Mental Health Scale (SKM)-12 mental health measurement. The system aims to remind students' awareness of the importance of mental health. To improve accuracy, this research compares the performance of three models, namely Support Vector Machine, Random Forest, and Logistic Regression, both with and without using SMOTE. The dataset obtained is 78 students, and SKM-12 consists of several groups, namely optimal mental health profile with symbol (+-), maximum mental illness profile with symbol (++), minimum mental illness profile with symbol (--), and minimal mental health profile with symbol (-+). The results of this study using the Logistic Regression method using SMOTE obtained better model performance compared to other methods, with an accuracy of 89.28%, an average class precision of 89.5%, an average class recall of 89.75%, and an average F1 - class score of 88.5%. This research shows that overcoming class imbalance using SMOTE can significantly improve the performance of mental health classification models.