Abstract
The fundamental issue with drug-drug interactions (DDIs) is that they cannot be ignored or overlooked since negative drug reactions and the use of medical services as a result are detrimental to patients and increase healthcare expenses. Conventional machine learning (ML) applications to DDI extraction, such as logistic regression or support vector machines, have been less successful, as the relationships in biomedical text are difficult to describe comprehensively. Recent advances in deep learning and transformer-based models offer improved contextual insight, but their resource-intensive demands may pose a barrier to scalability. We introduce CNN-DDI, a convolutional neural network model that can extract DDIs in biomedical text efficiently. On the SemEval-2013 dataset, we performed a comparative analysis of the classical models of ML (Logistic Regression, SVM, Random Forest, Naive Bayes, Decision Trees), transformer-based models (BioBERT, RoBERTa, DeBERTa, ELECTRA, DistilBERT), and the designed CNN-DDI. All the models were trained under similar procedures of parameter tuning and preprocessing. When comparing the models, CNN-DDI shows the highest performance with 86.81 percent overall accuracy and 83.81 percent F1-score, outperforming transformer-based models (best F1-score 81.41%, BioBERT-BiLSTM) as well as traditional ML models (best F1-score 77.09 percent, Logistic Regression). CNN-DDI integrates a competitive performance and fewer computer requirements, making it a feasible option in large-scale biomedical text mining.