Skin Disease Classification Using Ensemble Learning and Deep Feature Extraction

Skin conditions represent a significant global health burden, necessitating accurate and accessible diagnostic tools. This paper presents a methodological study on an automated classification system that combines deep feature extraction with an ensemble of classical machine learning classifiers. The proposed approach utilizes a pre-trained EfficientNetB0 model to extract high-dimensional feature representations from a dataset of 1,169 dermoscopic images, which spans eight distinct disease categories and presents a notable class imbalance. These features are subsequently classified using a soft-voting ensemble of four diverse classifiers: Support Vector Machine (SVM), Random Forest (RF), Linear Discriminant Analysis (LDA), and Naive Bayes (NB). To mitigate the effects of class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) is applied to the feature set before training. The system achieved a training accuracy of 96.40% and a test accuracy of 82.00%. This paper details the framework, analyzes its performance, and discusses the implications of using such a hybrid approach, particularly in the context of limited and imbalanced datasets. The results highlight the potential of this methodology for integration into diagnostic support tools while also underscoring key challenges, such as model generalization, that warrant further investigation.