An Extreme Outlier Handled Stochastic Random Under-Sampling Technique To Handle Class Imbalance In VAERS Dataset To Enhance Adverse Events Prediction

Abstract

The rapid development and deployment of vaccination drives have been crucial in controlling the pandemic. However, monitoring and predicting adverse reactions following vaccination is essential to ensure public safety and maintain vaccine confidence. This study utilizes the Vaccine Adverse Event Reporting System (VAERS) dataset to predict the likelihood of mortality following COVID-19 vaccination, addressing the challenge posed by the highly imbalanced nature of the dataset. Given the disproportionate ratio of mortality outcomes in the data, traditional classification models show significant predictive bias towards the majority class. To mitigate this, under sampling technique was applied to balance the classes before model training, aiming to enhance the performance of the predictive models towards the minority class (Fatal event). After evaluation of random under sampler and Near-Miss under sampler and it was found that Random under sampler is simple and effective solution on  larger datasets. Hence, as a further enhancement, a multi-objective stochastic search method combined with random under sampling was proposed to achieve better F1 score and class distribution. As the performance of prediction was affected by extreme outliers in the data, the proposed method was further enhanced with extreme outlier handling method with window of (0.05 to 0.25). The proposed method achieved an accuracy and F1 score in the range of (92-97%) with extreme outlier window (0.05). This work validates the hypothesis that presence of under-represented patterns would contributes to misclassification and affects the performance of the model confirming that our integrated outlier-handling approach effectively optimizes model robustness.