Improving Differential Diagnosis of Pathologically Similar Dermatological Conditions to Minimize Invasive Procedures

The erythemato-squamous dermatological diseases share many clinical and histopathological attributes making their differential diagnosis a challenging task for physicians. In this paper, we develop a classification algorithm based on previous clinical data to assist dermatologists in differentiating among these conditions. The final embodiment of the model was an optimized SVM with a Guassian kernel (γ = 2-8) combined with principal component analysis (PCA), achieving a maximum classification accuracy of 97.8%, an average sensitivity and specificity of 97.64% and 99.57%, respectively. Feature selection revealed a subset of features with relatively high positive predictive impact (e.g., itchiness of affected area) and relatively negative predictive impact (e.g., age). Optimized models trained on only non-invasive feature subsets (features not requiring invasive biopsy) yielded an average sensitivity of 81.56% and a specificity of 96.78%, suggesting the utility of our model as a first-line screening tool for dermatologists to rule-out the more severe diseases, and thus minimize subsequent invasive testing.