Abstract
Detection of vaccine adverse events is crucial to the discovery and improvement of problematic vaccines. To achieve it, traditionally formal reporting systems like VAERS support accurate but delayed surveillance, while recently social media have been mined for timely but noisy observations. Utilizing the complementary strengths of these two domains to boost the detection performance looks good but cannot be effectively achieved by existing methods due to significant differences between their data characteristics, including: 1) formal language v.s. informal language, 2) single-message per user v.s. multi-messages per user, and 3) one class v.s. binary class. In this paper, we propose a novel generic framework named Multi-instance Domain Adaptation (MIDA) to maximize the synergy between these two domains in the vaccine adverse event detection task for social media users. Specifically, we propose a generalized Maximum Mean Discrepancy (MMD) criterion to measure the semantic distances between the heterogeneous messages from these two domains in their shared latent semantic space. Then these message-level generalized MMD distances are synthesized by newly proposed mixed instance kernels to user-level distances. We finally minimize the distances between the samples of the partially-matched classes from these two domains. In order to solve the non-convex optimization problem, an efficient Alternating Direction Method of Multipliers (ADMM) based algorithm combined with the Convex-Concave Procedure (CCP) is developed to optimize parameters accurately. Extensive experiments demonstrated that our model outperformed the baselines by a large margin under six metrics. Case studies showed that formal reports and extracted adverse-relevant tweets by MIDA shared a similarity of keyword and description patterns.
Introduction
Monitoring vaccine safety is a critical public health priority. Traditional systems like the Vaccine Adverse Event Reporting System (VAERS) provide high-quality, verified data but often suffer from reporting delays. Conversely, social media platforms like Twitter offer real-time insights but are plagued by noise and informal language. Bridging these two domains presents three primary challenges: the language gap between medical jargon and colloquialisms, structural differences (single reports vs. "bags" of tweets per user), and class imbalance (VAERS is mostly positive cases, while Twitter is mostly negative).
To address these, the researchers developed the MIDA (Multi-instance Domain Adaptation) framework. The study utilized 2,500 formal reports from VAERS and a custom Twitter dataset comprising 41,438 tweets from 1,572 users to validate the approach.
Methodology
The MIDA framework is designed to align the source domain (formal reports) with the target domain (social media users) in a shared latent semantic space.
Key technical innovations include:
- Generalized Maximum Mean Discrepancy (MMD): This criterion measures semantic distances between heterogeneous messages, effectively "translating" between medical terminology and informal tweets.
- Mixed Instance Kernels: Since VAERS data consists of single instances and Twitter data consists of "bags" of instances, these kernels synthesize message-level distances into user-level distances.
- Optimization via ADMM and CCP: To handle the non-convex nature of the optimization problem, the authors employed the Alternating Direction Method of Multipliers (ADMM) combined with the Convex-Concave Procedure (CCP), ensuring efficient and accurate parameter estimation.
A core insight of MIDA is treating Twitter users as "bags" of instances (Multi-instance Learning). This allows the model to identify the specific tweet indicating an adverse event while ignoring irrelevant content, such as personal anecdotes or unrelated status updates.
Experimental Results
MIDA was evaluated against several baseline models, including multi-instance learning (miFV, miVLAD) and transfer learning (JDA, ARTL) approaches.
The results demonstrate MIDA's superior performance across multiple metrics:
- Accuracy (ACC): 0.7767
- AUC: 0.8530
- Precision: 0.7735
The high precision is particularly noteworthy for public health applications, as it reduces the risk of false alarms. Furthermore, the ADMM-based optimization proved scalable, maintaining linear complexity relative to the number of users and keywords.
Conclusion and Future Work
MIDA successfully bridges the gap between formal medical reporting and informal social media data, providing a robust tool for timely vaccine safety monitoring. While the current model relies on predefined keywords and is computationally more intensive than simpler baselines, future iterations could leverage deep learning architectures like BERT for automated feature extraction. Additionally, extending the framework to incorporate multiple source domains, such as clinical trial data, remains a promising avenue for research.
Summary updated