A machine learning algorithm based on trichoscopic features for the assessment of patchy alopecia areata activity

Introduction Recently, the application of artificial intelligence in dermatology has expanded at an enormous rate, especially in the field of dermoscopy. Regarding alopecia areata (AA), however, main publications focus on the quantification of disease extension, but no study has yet been published about AA activity assessment. Objectives This study aims to elaborate a machine learning algorithm, based on trichoscopic features, capable of providing an estimate of patchy alopecia areata activity. Methods We analyzed 210 trichoscopic images of AA patches from a total of 115 patients. A 20x magnification image of the border of each patch was taken by an experienced operator. The following trichoscopic features were counted for each patch: total number of black dots, broken hairs, exclamation mark hairs, tapered hairs, hairs presenting Pohl-Pinkus constrictions (altogether named as “black dots”, BD); number of yellow dots (YD); total number of short vellus hairs, upright regrowing hairs and pigtail hairs (altogether named as “vellus hairs”, VH). We developed a machine learning model, based on these three parameters, able to classify each patch as either progressive, stable or remitting. Results We evaluated 70 progressive patches, 70 stable patches and 70 remitting patches. We found a significant higher number of BD in progressive patches, YD in stable patches and VH in remitting ones. Confusion matrix scores were ≥80% for all stages, both for train and test datasets. VH resulted as the most important parameter for the decisional algorithm (43.5%), while YD as the least relevant one (20.4%). Model performance metrics were all above 90%, with an accuracy of 93.8%. Conclusions The machine learning model we elaborated aims to offer additional support for dermatologists, providing automatic activity assessment for patchy AA based on trichoscopic patterns, although the absence of a software for automated counting of trichoscopic features remains the main obstacle for future applications.