Authors - Aarthi R, Aniketha Prasad, Dhamini Manoj, Manasvi G, Meghaa Sunil Abstract - Early and accurate diagnosis of dermatological disorders remains one of the main issues in clinical dermatology, especially with regard to diseases with similar appearances. Despite the achievements of deep learning methodologies in the classification of cutaneous lesions with the help of images, structured clinical metadata is not used to the fullest, despite its significant diagnostic potential. In a practical clinical setting, dermatologists do not solely use visual evaluation of the case but also use patient-specific metadata, which includes age, lesion progression, pruritus, hemorrhage, anatomic location, prior biopsy, and family history. The current study presents a fully explainable, metadata based, multi-class classification of skin diseases, using the PAD-UFES-20 database, and concentrated on 6 distinct diagnostic categories. Although the dataset is dermoscopic, the predictive quality of formal metadata variables are mainly under consideration in the present work. The explainability analyses revealed that biopsy status, elevation, itch, region and age are attributes that have significant effects on the classification results. However, empirical evidence shows that the reduced model consisting of the premier five features lowers accuracy, which highlights the importance of a thorough combination of metadata features to determine skin disease rather than limited combination. Comparative studies indicate that the Multi-Layer Perceptron shows an improvement in a model performance with a corresponding increase of the number of selected features. The suggested framework thus highlights interpretability in line with predictive efficacy thus enhancing the importance of transparent artificial intelligence systems in medical decision-making.
Open Zoom