Thesis Defence: Combating Diabetes Misinformation: A Transformer Model Approach

Linda Okpanachi, supervised by Dr. Ifeoma Adaji, will defend their thesis titled “Combating Diabetes Misinformation: A Transformer Model Approach” in partial fulfillment of the requirements for the degree of Master of Science in Computer Science.

An abstract for Linda Okpanachi’s thesis is included below.

Defences are open to all members of the campus community as well as the general public. Registration is not required for in-person defences.

---

Abstract

Health misinformation is an increasingly serious public health concern, particularly in relation to chronic conditions such as diabetes. The proliferation of unverifiable information on digital platforms, especially social media, makes it more challenging for individuals to find trustworthy information regarding diabetes symptoms, care, and treatment options. This study introduces DiaBERT, a transformer-based tool designed to identify misinformation related to diabetes, with the goal of categorizing health claims as True, False, or Partially True. The tool was trained on a varied dataset that merges formal vetted content and informal online discussions, which presented challenges like inconsistent terminology, shifts in context, and unclear labels. To address domain differences, DiaBERT was constructed using BioBERT and further refined with a domain adaptation strategy using Domain Adversarial Neural Network (DANN), enabling the model to adjust to different linguistic styles. In evaluations, DiaBERT performed better when compared to the machine learning and deep learning models, achieving an accuracy of 79% and demonstrating the ability to detect nuanced claims labeled as Partially True. A content filtering feature was also incorporated to help ensure that the system processes only diabetes-related claims. To enhance transparency and user confidence, the system includes an explainability feature that emphasizes key segments of the input text contributing to the classification. This assists users in understanding the basis of the system’s decision-making and makes the tool more approachable for nonexperts. DiaBERT was launched as a Chrome extension that provides users with classification results, confidence scores, and brief explanations while they explore content online. The user study conducted with 45 participants showed that users found the tool trustworthy, useful, and easy to understand, particularly valuing features like keyword highlighting and explanation clarity. This research helps in addressing diabetes health-related misinformation and promoting more informed choices.