Detecting Fluoride Misinformation on Social Media Using Machine Learning

Objectives: This study aimed to detect fluoride-related misinformation on social media using artificial intelligence (AI)-based models.
Methods: Initially, a single expert collected 49,960 English tweets utilizing the Twitter API within a Python 3 interface, employing the keyword “fluoride” for the period between September 2017 and September 2022. Subsequently, five independent, trained, and calibrated investigators meticulously labeled a subset of 15,000 tweets to categorize them as either fluoride-related misinformation or non-misinformation using thematic content analysis. These human-labeled tweets underwent rigorous data cleaning and preprocessing before serving as the foundation for training various AI models, including a) Support Vector Machine (SVM), b) Logistic Regression, c) Long Short-Term Memory (LSTM), d) Bidirectional LSTM, e) XGBoost, f) Convolutional Neural Networks (CNN), and g) Bidirectional Encoder Representations from Transformers (BERT). The selection of the most proficient models was based on key performance metrics such as accuracy, precision, recall, and F1-score.
Results: The accuracy values of the models ranged from 77.00% to 82.00%, with both SVM and BERT demonstrating equally high performances in correctly classifying instances of misinformation. Furthermore, both SVM and BERT exhibited superior precision, recall, and F1-score compared to other AI models. Specifically, SVM demonstrated a higher precision rate (0.83 vs. 0.80), while BERT achieved a superior recall rate (0.81 vs. 0.80). Remarkably, both models attained an equivalent F1-score of 0.82, highlighting their balanced performance in misinformation detection.
Conclusions: Hence, SVM and BERT models demonstrated outstanding performance metrics in detecting fluoride-related misinformation within the domain of social media, outperforming other AI models. These findings have the potential to contribute to the development of a digital ecosystem designed to identify and combat false narratives surrounding fluoride on online platforms, ultimately benefiting public health initiatives. Future research should explore the adaptability and generalizability of AI-based models in detecting a broader spectrum of oral health-related falsehoods across various social media platforms.