Evaluating the Accuracy of Machine Learning Models for Pharmaceutical Image Classification in Indonesia

In Indonesia, drug image classification has significant obstacles owing to discrepancies in packaging and variable lighting conditions. Numerous machine learning models created in industrialized nations have not been successfully implemented in Indonesia, mostly owing to the absence of representative local datasets. This research seeks to assess the precision of a drug classification model using a Convolutional Neural Network (CNN) trained on a local dataset of 30 medication kinds. The model attained 59.92% accuracy on training and validation datasets; however, its accuracy plummeted to 5.65% on external data, indicating an overfitting issue. The study of the confusion matrix reveals that the model struggles with correct classification, as seen by poor accuracy, recall, and $\mathbf{F 1}$-score values, with elevated classification mistakes, including false positives and false negatives. The imbalance of data across classes leads to unsatisfactory performance, as the model preferentially aligns with dominant classes and neglects the distinctive characteristics of each medication class. Suggestions for enhancement include data balancing, implementation of weighting methods, and improved feature extraction approaches to augment the model’s capacity to distinguish patterns within classes, consequently facilitating more dependable medication categorization within the pharmaceutical sector in Indonesia.