Penggunaan Model Bahasa Besar (LLM) Untuk Dukungan Kesehatan Mental Pada Pelajar

Student mental health has become an increasingly critical issue due to academic pressure, social challenges, and extensive exposure to digital media, which contribute to rising levels of stress, anxiety, and depression. Limited access to conventional mental health services highlights the need for data-driven and scalable technological approaches to support early detection. The core focus of this study is a comparative analysis between traditional machine learning models and large language models (LLMs) for classifying student stress levels. This research aims to evaluate the performance of the XGBoost machine learning model on tabular student mental health data, assess the effectiveness of LLMs through tabular-to-text serialization using few-shot learning and fine-tuning, and compare the strengths, limitations, and adaptability of both approaches. The proposed methodology involves data preprocessing and normalization, stress-level classification using XGBoost on structured tabular data, and transformation of tabular records into narrative text for classification using LLMs, including DeepSeek, LLaMA, and Gemma. The dataset used is a publicly available Kaggle dataset consisting of 1,100 student records with 21 attributes. Model performance is evaluated using accuracy, precision, recall, F1-score, ROC-AUC, confusion matrix, and log-loss metrics. The results indicate that XGBoost achieved 100% classification accuracy on tabular data. Among the LLMs, Gemma also achieved 100% accuracy after fine-tuning, while LLaMA and DeepSeek reached accuracies of 96% and 91%, respectively. The discussion highlights that XGBoost excels in efficiency and interpretability for structured data, whereas LLMs demonstrate superior contextual understanding and the ability to generate empathetic, language-based recommendations. In conclusion, integrating machine learning and LLM-based approaches offers a comprehensive and adaptive solution for student mental health stress detection.