An Effective Methodology for Identifying Adverse Drug Reactions using Firefly Algorithm

An individual’s unwanted and unpleasant reactions that arise from regular drug use are referred to as adverse drug reactions, or ADRs. Mild side effects to serious, potentially fatal diseases can be caused by these reactions. In order to guarantee the security and effectiveness of pharmacological interventions, it is imperative to track and comprehend ADRs. This paper presents an approach in pharmacovigilance to classify drug reactions caused by different drugs by combining firefly algorithm with different classifiers. The firefly algorithm assigns firefly to each subsets of features and uses the objective function in order to calculate the distance and making cluster of firefly. In the end it gives us the most optimal set of features which we further use in our classification. We also have experimented on three types of classifiers which are: Random Forest, K-Nearest Neighbour, Decision Tree. In the end we have compared the accuracy, Precision, F1 Score of different classifiers and concluded that by using firefly for feature extraction we can increase accuracy, precision and F1-score. We have also compared this Firefly algorithm with Elephant Herding Optimization (EHO) used for feature selection. We dived into the advancements and challenges faced during the prediction of reactions of drugs. There is much scope in this as we can further increase our performance and efficiency by combining multiple classifiers or using different feature extraction techniques. We used a dataset consisting of 1333 entries with 24 features from Kaggle, split into training and testing sets with a 70:30 ratio. The results demonstrate that applying the Firefly Algorithm enhances model performance. Random Forest classifier achieved the highest accuracy of 97.5% with the Firefly Algorithm, compared to 97.2% without it. Similarly, KNN and Decision Tree classifiers also showed improvements in accuracy, with KNN improving from 93.5% to 95.0% and Decision Tree improving from 96.7% to 97.0%. Additionally, precision and F1-scores were consistently enhanced across all models, making the proposed approach highly effective in ADR prediction.