DeepSide: A Deep Learning Framework for Drug Side Effect Prediction

Abstract

The problem of adverse drug responses is significant in the fields of contemporary medicine as it usually results in severe issues, an increase in the number of hospitalizations, and an increase in expenses of medical care. Therefore, it is relevant to learn about potential drug side effects in time to ensure the patient safety and assist physicians in making judgment. Because of the use of lab tests, clinical trials, and hand analysis, the standard methods of determining whether a drug is safe rely on them. They are time-intensive, costly and not necessarily successful in discovering complicated patterns in huge volumes of biomedical information. To avoid these challenges, we introduce DeepSide, a DL model of automatic prediction of drug side effects on a dataset of drug images. In order to be able to distinguish the images of the drugs, the proposed approach involves the use of sophisticated convolutional neural network architectures, including VGG19 with Batch Normalization, DenseNet121, ResNet18 and ConvNeXt-Tiny. The information is prepared in terms of validating structured data, transforming it and performing controlled train-validation splitting to be used in reliable model training. Performance can be judged by means of standard classification measures such as Accuracy, Precision, Recall, F1- Score, ROC curve analysis and confusion matrix analysis. The DenseNet121 architecture is more effective in predicting the future as compared to the other models that were implemented. Its success rate in all the evaluative measures reached 99.2% of success. The better model is then designed into a Flask-based application to allow you to make predictions through images and view your confidence levels and display chemical structures more understandable. On the whole, the proposed solution is an effective and reliable method of automatic prediction of drug side effects, which results in more safety analysis and intelligent pharmaceutical decision support.