Artificial intelligence for the diagnosis of erythematous-squamous dermatological diseases: technological contributions to primary care

Abstract

Background  Accurate diagnoses in dermatology can be challenging for general practitioners. In this context, the support of artificial intelligence tools can be beneficial in the Brazilian primary care setting.

Objectives  To develop an interpretable machine-learning algorithm capable of assisting in the diagnosis of erythematous-squamous dermatological diseases through clinical data, without histopathological support.

Methods  The random-forest algorithm was trained with the public Dermatology database of 366 patients diagnosed with: chronic dermatitis, lichen planus, pityriasis rosea, pityriasis rubra pilaris, psoriasis, or seborrheic dermatitis. The model was evaluated by performance metrics and interpretability techniques.

Results  The model showed good predictive performance, with ROC-AUC ranging from 0.89 to 1.00, and overall accuracy of 0.86. The best results were for the diagnosis of pityriasis rubra pilaris (f1-score: 1.00) and the worst for chronic and seborrheic dermatitis (f1-score: 0.77 and 0.76, respectively). The clinical characteristics that most influenced the model's decision were, in decreasing order: involvement of knees and elbows, involvement of scalp, Koebner phenomenon, polygonal papules, and involvement of oral mucosa.

Study limitations  The model was not validated with Brazilian data.

Conclusion  The developed technology obtained good predictive performance and clinical coherence. There is a need for adaptation for implementation, using national data. The results indicate the potential for similar models to be improved and adapted to clinical practice for the benefit of the Unified Heath System.

Keywords
Artificial intelligence; Dermatology; Diagnosis, differential; Expert systems; Machine learning