From Text to Therapy: A Continuous Sentiment Analysis Framework for Real-Time Mental Health Monitoring in Mobile Applications

Abstract

In the recent years, the adoption of mobile mental health applications has increased substantially, and with that, there’s a real need for solid, real-time sentiment analysis tools that can actually adapt effectively, tools that track user emotional states and enable timely intervention. But most current models are limited in effectiveness. They struggle with limited high-quality labelled data, miss the subtleties in people’s emotions, and do not adapt effectively because they rely on static Machine Learning (ML) models. A Continuous Sentiment Analysis for Mental Health Monitoring (CSA-MH) framework is introduced in this study to bring together semi-supervised learning, Bidirectional Encoder Representations from Transformers (BERT)-based contextual embeddings, and lexicon-driven emotional features. This integration boosts both the quality of labelling and the depth of emotional understanding. In addition, it uses a hybrid model that combines Long Short-Term Memory (LSTM) and Gradient Boosting Machine (GBM) model, and it keeps it up to date through an online learning sliding window. This way actually grows the model, changes along with the user, and always stays relevant. Practically, when tested on a handpicked slice of the Sentiment140 dataset, 50,000 tweets filtered for mental health keywords, the framework achieved: 92.3% accuracy and an F1-score of 0.91. That’s way ahead of baseline models like BERT, LSTM, and Support Vector Machine (SVM). Case studies show it can spot emotional changes over time and even send out alerts when something shifts. Bottom line, CSA-MH isn’t just another tool; it’s a scalable, adaptable, and effective way to keep tabs on mental health in real time. In such fields, this kind of tech pushes digital mental health care toward something much more personal.