Deep Learning Based Bi-Directional LSTM for Sentiment Analysis of Health App Reviews

Health and wellness applications are used by many people to help with daily routines and long term habit formation. Users often share their experiences through app reviews, and these comments can be valuable for understanding what works well and where the design falls short. In this study, we examined these reviews using a deep learning approach built around a manually annotated set of 21,322 English entries. Before modelling, the text was cleaned, standardized, and balanced through random oversampling so that all sentiment categories were represented fairly. Deep learning models like CNN, RNN, LSTM, Bi- LSTM, and attention-based models were investigated in this study. A stacked Bi-LSTM with embedding dimensions, L2 regularized dropout found to be the best model attaining an accuracy of 93.55% with an F1- score of 94.00%. Earlier models relied on shallow or pre-trained embeddings whereas here the new model that has been proposed relies on data level balancing with specific targeted architectural refinements meant to improve upon health related generalization. This is realistically reflected by means of subtle sentiment expression capture to help in evaluation of health and wellness practical foundations.