Detecting Medical Misinformation on Social Media Using Multimodal Deep Learning

Zuhui Wang, Zhaozheng Yin, Young Anna Argyris

TL;DR

Researchers have developed a multimodal deep learning framework specifically designed to detect antivaccine misinformation on Instagram by analyzing images, captions, and hashtags simultaneously. By leveraging a novel "semantic- and task-level" attention mechanism and an ensemble of models, the system achieves a remarkable 97.4% accuracy, significantly outperforming traditional text-only detection methods.

What Problem Does This Paper Address?

Vaccine hesitancy is a major public health challenge, often fueled by misinformation on social media. While previous research focused heavily on text-based platforms like Twitter, misinformation has migrated to image-centric platforms like Instagram.

On these platforms, the "antivaccine" message isn't just in the caption; it’s often embedded within the image itself (as overlaid text) or hidden within a string of specific hashtags used to reach like-minded communities. Existing detection systems that only "read" the text or "look" at the image in isolation fail to capture the full context, especially when users intentionally overlay text on images to subvert automated moderation. This paper seeks to close that gap by creating a system that "sees" and "reads" a post just as a human would.

Methods

The authors propose a three-branch multimodal deep learning network:

  1. Hashtag Branch: Uses a fine-tuned fastText embedding to handle the "out-of-vocabulary" nature of hashtags (e.g., #vaccineinjury).
  2. Text Branch: Combines the post's caption with text extracted from the image via Optical Character Recognition (OCR). This combined text is processed through a bidirectional RNN with GRU units.
  3. Image Branch: Uses a fine-tuned VGG19 network to extract visual features, enhanced by a "spatial and channel squeeze-and-excitation" (scSE) block to focus on the most relevant parts of the image.

The Secret Sauce: SeTa Attention The core innovation is the Semantic- and Task-level Attention (SeTa_Att) mechanism. Standard attention usually looks for "meaning" (semantic). SeTa adds a "task" layer that specifically weighs features based on their similarity to known antivaccine markers. For example, it doesn't just look for any hashtag; it prioritizes hashtags that historically correlate with antivaccine sentiment.

Finally, the authors use an Ensemble Method. They take the prediction scores from the three-branch model and three separate single-modal models, feeding them into a Support Vector Machine (SVM) with an RBF kernel to make the final "Antivaccine: Yes/No" call.

Datasets

The researchers curated a new, real-world dataset from Instagram:

  • Size: 31,282 posts collected between January 2016 and October 2019.
  • Composition: 50% antivaccine and 50% non-antivaccine (including pro-vaccine and irrelevant posts).
  • Collection Method: A "snowballing" technique was used, starting with known keywords (#bigpharma, #vaccinesafety) and identifying 32 influential antivaccine accounts to gather more data.
  • Ground Truth: Posts were independently labeled by three trained annotators, with a majority-voting scheme used to resolve disagreements.

Key Insights and Ideas

  • Multimodality is Non-Negotiable: The paper demonstrates that ignoring any one element (image, text, or hashtags) leads to failure. For instance, a post might have a neutral caption but an image of a syringe with "poison" written over it.
  • Text in Images as a Subversion Tactic: The authors highlight that "savvy users" intentionally overlay text on images to dodge automated text filters. By integrating OCR into the text branch, the model effectively neutralizes this tactic.
  • Hashtags as Community Signifiers: Unlike natural language, hashtags are often used as "catchphrases" to summarize a post's main claim. The model treats them as a distinct, high-signal data stream rather than just more text.
  • Text Still Reigns Supreme: While the multimodal approach is best, the ablation studies showed that the "caption-only" model (91.1% accuracy) performed better than the "image-only" model (86.5%), suggesting that textual information remains the most explicit signal for misinformation.

Key Results

The proposed system sets a new benchmark for misinformation detection:

  • Ensemble Model Accuracy: 97.4% (Precision: 0.978, Recall: 0.967).
  • Three-Branch Model (No Ensemble): 96.6% accuracy.
  • Comparison: It significantly outperformed state-of-the-art models like MVAE (92.9%) and EANN (84.8%).
  • Real-World Testing: In a 30-day online experiment (Nov 2019), the model successfully tracked the daily volume of antivaccine posts on Instagram, maintaining high precision and recall in a live environment.

Limitations and Future Work

Despite the high accuracy, the authors identify a few "failure cases":

  • External Knowledge Gaps: The model struggles with posts that require specific context, such as a post protesting a specific bill (e.g., "Veto SB 792") without explicitly mentioning vaccines.
  • OCR Limitations: The system sometimes fails to read small, handwritten, or highly stylized text within images.
  • Lack of "Memory": The current architecture processes each post in isolation. The authors suggest adding "memory cells" in future iterations to help the model remember essential features across different training samples, potentially improving its ability to handle subtle or evolving misinformation tactics.

Summary updated

Abstract

In 2019, outbreaks of vaccine-preventable diseases reached the highest number in the US since 1992. Medical misinformation, such as antivaccine content propagating through social media, is associated with increases in vaccine delay and refusal. Our overall goal is to develop an automatic detector for antivaccine messages to counteract the negative impact that antivaccine messages have on the public health. Very few extant detection systems have considered multimodality of social media posts (images, texts, and hashtags), and instead focus on textual components, despite the rapid growth of photo-sharing applications (e.g., Instagram). As a result, existing systems are not sufficient for detecting antivaccine messages with heavy visual components (e.g., images) posted on these newer platforms. To solve this problem, we propose a deep learning network that leverages both visual and textual information. A new semantic- and task-level attention mechanism was created to help our model to focus on the essential contents of a post that signal antivaccine messages. The proposed model, which consists of three branches, can generate comprehensive fused features for predictions. Moreover, an ensemble method is proposed to further improve the final prediction accuracy. To evaluate the proposed model's performance, a real-world social media dataset that consists of more than 30,000 samples was collected from Instagram between January 2016 and October 2019. Our 30 experiment results demonstrate that the final network achieves above 97% testing accuracy and outperforms other relevant models, demonstrating that it can detect a large amount of antivaccine messages posted daily. The implementation code is available at https://github.com/wzhings/antivaccine_detection.
BibTeX
@article{wang2021detecting,
  author = {Zuhui Wang and Zhaozheng Yin and Young Anna Argyris},
  title = {{Detecting Medical Misinformation on Social Media Using Multimodal Deep Learning}},
  journal = {IEEE J. Biomed. Health Informatics},
  year = {2021},
  url = {https://openreview.net/forum?id=61jXsipjOC},
}
Figure