Scoring systems in dermatology

Abstract

Three scoring systems were considered and described for dermatological applications in which malignant and benign skin lesions have to be recognized: two models are derived from logistic regression and naïve Bayes rule by rounding model parameters to their nearest integer values; the third approach defines the scoring system by a direct stepwise adding of the most significant binary risk factors. An application example of a direct score model was then developed to illustrate important aspects of its design in dermoscopy. The results show that, having many variables available, score models combine simplicity, practicality, high accuracy and good control of overfitting. Also they can incorporate different diagnostic styles of experienced dermatologists, introducing into the model subjective binary variables, some of which also assessed with a significant degree of disagreement.