Abstract
Vaccine hesitancy is a complex, context-dependent behavior influenced by socioeconomic status, trust in institutions, and cultural background. While machine learning models are effective at predicting vaccine refusal, they often lack the interpretability required to understand the underlying drivers across diverse demographic groups. This paper presents VACADOPT, a culture-aware machine learning framework designed to analyze COVID-19 vaccine hesitancy. By integrating ensemble classifiers with Explainable AI (XAI) techniques—specifically SHAP and a novel Tabular-PGM approach—the study identifies distinct drivers of hesitancy among Hispanic, African American, and Asian communities in the United States.
Introduction
Public health efforts to combat COVID-19 have been significantly hindered by vaccine hesitancy. Unlike a simple binary choice, hesitancy is a spectrum of behavior shaped by a multitude of factors, from religious beliefs to concerns about side effects. Traditional "black-box" machine learning models can predict whether an individual might refuse a vaccine, but they rarely explain why. This lack of transparency is particularly problematic when addressing minority communities, where historical context and cultural nuances play a critical role in health-related decision-making.
The VACADOPT framework seeks to bridge this gap by providing a model that is both highly accurate and culturally interpretable. By identifying specific drivers for different ethnic groups, the framework enables public health officials to move beyond one-size-fits-all communication strategies and design targeted, effective vaccination campaigns.
Methodology
The VACADOPT framework utilizes a three-stage approach to analyze vaccine hesitancy:
- Predictive Modeling: The researchers evaluated several machine learning classifiers, including eXtreme Gradient Boosting (XGB), Random Forests (RF), and TabNet (a deep learning architecture for tabular data). A Random Forest model utilizing a hybrid encoding schema (combining label and one-hot encoding) was found to provide the most robust performance.
- Individual Feature Analysis (SHAP): To interpret individual predictions, the framework employs SHAP (Shapley Additive Explanations). This allows for the identification of specific factors—such as trust in the FDA or income level—that influence the model's decision for a single participant. This stage is crucial for distinguishing between "rational" concerns (e.g., side effects) and "irrational" drivers (e.g., conspiracy theories).
- Composite Feature Analysis (Tabular-PGM): Recognizing that individual survey questions often relate to broader themes, the authors adapted the PGM-Explainer tool. Originally intended for Graph Neural Networks, it was modified into Tabular-PGM to analyze how clusters of related features (e.g., "Cultural Identity" or "Media Consumption") impact the final prediction.
Data and Demographics
The study utilized a comprehensive benchmark dataset derived from a national online survey conducted via Qualtrics in the Fall of 2021. The dataset includes 125 possible factors per participant, covering media consumption, trust in health authorities (CDC, FDA), political affiliation, and specific cultural/religious identifiers.
To ensure the model was statistically grounded in cultural awareness, the sample was designed to reflect U.S. Census data. Specific quotas were maintained for Hispanic (18.1%), African American (13.4%), and Asian (5.9%) participants, allowing for a granular analysis of how hesitancy drivers diverge across these populations.
Results and Discussion
Model Performance
The Random Forest classifier with hybrid encoding achieved an accuracy of 88.8% and an F1-score of 93.2%, outperforming both XGBoost and SVMs. Notably, the specialized deep neural network TabNet performed the worst among the tested models, reinforcing the observation that ensemble methods remain the gold standard for structured, survey-based tabular data.
Cultural Nuances in Hesitancy
The most significant finding of the study is that vaccine hesitancy is not monolithic; its drivers vary significantly by ethnicity:
- Asian Communities: Decisions are largely "product-centric." Hesitancy in this group is primarily driven by vaccine-specific factors, such as trust in the vaccine's approval status and its perceived necessity for healthy individuals.
- African American Communities: Drivers are more "identity-centric." Hesitancy is most strongly influenced by cultural factors, including social identity and religiosity.
- Hispanic Communities: This group is most impacted by COVID-related information and misinformation (Group D) sourced from social media platforms.
Rational vs. Irrational Drivers
SHAP analysis revealed a clear divide in the nature of hesitancy. "Rational-based" hesitancy, often rooted in concerns about the speed of vaccine development or potential side effects, may be addressable through transparent data sharing. In contrast, "irrational-based" hesitancy, strongly associated with conspiracy theories and political affiliation, is more likely to lead to outright vaccine rejection.
Conclusion
The VACADOPT framework demonstrates that cultural awareness is essential for understanding and addressing vaccine hesitancy. By moving beyond simple predictive modeling to a multi-layered XAI approach, researchers can pinpoint the specific social, cultural, and informational drivers that influence different communities. Future work will focus on developing strategies to mitigate social media misinformation in the Hispanic community and using these insights to build tailored, trust-increasing resources for diverse ethnic groups.
Summary updated