"PCA-ENHANCED ENSEMBLE LEARNING FOR ECZEMA-CENTRED SKIN DISEASE CLASSIFICATION: A COMPARATIVE EVALUATION WITH MOBILENETV2-BASED DEEP LEARNING"

Abstract

Atopic dermatitis (Eczema) is a chronic, inflammatory skin disease that poses significant diagnostic challenges Due to its clinical overlap with other erythematous-squamous skin conditions such as psoriasis, seborrhoeic dermatitis, lichen planus, pityriasis rosea, chronic dermatitis, and pityriasis rubra pilaris, atopic dermatitis (Eczema) is a chronic inflammatory skin disease that presents significant diagnostic challenges. Timely and accurate differential diagnosis is critical for effective treatment, yet remains heavily dependent on clinical expertise and invasive biopsy procedures. The majority of current systems work on either structured clinical records or dermatological images alone, leaving the comparative value and fusion potential of both modalities unexplored. However, recent advances in machine learning and deep learning offer compelling opportunities to automate and support dermatological diagnosis.

In order to classify skin diseases centred on eczema, this research suggests and assesses a dual-pipeline machine learning (ML) framework. The first pipeline processes structured histopathological and clinical data from the UCI Dermatology dataset (366 instances, 34 features, 6 disease classes including eczema) using Principal Component Analysis (PCA) for dimensionality reduction — compressing 34 features into 21 principal components retaining 95.3% of total variance — followed by three classical classifiers: Random Forest, Support Vector Machine (SVM), and XGBoost. The second pipeline applies transfer learning via MobileNetV2, pretrained on ImageNet, for deep visual feature extraction and classification across a skin image dataset of 12,453 training images spanning 23 visual skin condition classes, including eczema.

Results reveal a striking performance contrast between the two modalities. The tabular pipeline delivered near-perfect classification, with Random Forest achieving 97.3% accuracy, a macro F1-score of 0.970, and a macro ROC-AUC of 0.995. SVM and XGBoost followed at 96.2% and 95.6% accuracy, respectively, with all models achieving perfect classification on several disease classes. Feature importance analysis consistently revealed that the most diagnostically discriminative components were PC1, PC2, PC3, and PC4, which correspond to dominant histopathological markers known to differentiate eczema from clinically similar conditions. On the 23-class visual task, however, the image-based MobileNetV2 pipeline only managed 20.8% training accuracy and about 17% validation accuracy. This shows considerable underfitting, which can be explained by high inter-class visual similarity, few fine-tuning epochs, and the semantic gap between ImageNet pre-training and clinical dermatological imagery.

These findings establish that structured histopathological features, when reduced via PCA and modelled with ensemble classifiers, yield superior diagnostic performance for eczema-related skin disease classification compared to image-based deep learning under current data constraints.