Deep learning-based method for sentiment analysis for patients' drug reviews.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Pharmacovigilance

Background:

• Patient drug reviews offer valuable insights into medication effectiveness and side effects.
• Analyzing sentiment in these reviews is crucial for understanding patient experiences and improving drug safety.
• Traditional methods often struggle with the nuances and complexities of unstructured text data.

Purpose of the Study:

• To evaluate deep learning models for sentiment analysis of patient drug reviews.
• To compare the performance of bidirectional long-short-term memory (LSTM) and a hybrid bidirectional LSTM-CNN model.
• To assess the impact of GloVe word embeddings on sentiment classification accuracy.

Main Methods:

• Implementation and evaluation of bidirectional LSTM and bidirectional LSTM-CNN models.
• Utilizing GloVe word embeddings for feature representation.
• Training and testing models on two distinct drug review datasets.
• Sentiment classification based on review text, medical conditions, and rating scores.

Main Results:

• The Bi-LSTM-CNN model achieved 96% accuracy (Model A).
• The Bi-LSTM model achieved 87% accuracy (Model B).
• GloVe word embeddings demonstrably improved model performance, confirmed by Cohen's Kappa coefficient.

Conclusions:

• Deep learning approaches, specifically bidirectional LSTM and Bi-LSTM-CNN, are highly effective for sentiment analysis of patient drug reviews.
• The integration of GloVe word embeddings enhances the accuracy and reliability of sentiment classification.
• These findings support the use of advanced NLP techniques for extracting meaningful insights from patient-generated health data.