Transformer based on the prediction of psoriasis severity treatment response

Abstract

Psoriasis, a chronic inflammatory skin disease, must be continuously monitored to prevent recurrence and increase treatment effectiveness. However, time-series analysis studies considering the chronic characteristics of psoriasis are rare. The psoriasis area and severity index (PASI) score, a therapeutic decision-making index for psoriasis, has limitations in detecting disease variability during treatment. For successful treatment, an imagingbased approach that can intuitively observe lesion changes is required. We propose a novel deep learning-based evaluation method that identifies psoriasis severity characteristics and predicts changes in severity during treatment. The proposed method consists of two steps: extraction of deep features of psoriasis and prediction of disease severity in time series using the deep features. The proposed deep-feature model extracted global and deformable convolution layers V2-based local features using the hierarchical information of RegNetY-1.6G and fused the features to predict disease severity. After severity classification using fusion features, the f1-score was 0.92. Next, we constructed a short-term time-series disease dataset and extracted deep features using the proposed model. Among six time-series analysis models (LSTM, GRU, GRU-FCN, ResNet, InceptionTime, and Transformer), Transformer demonstrated the best prediction performance, with a root mean squared log error (RMSLE) of 0.44 and a symmetric mean absolute percentage error (SMAPE) of 0.34. From the Student's t-test and regression analysis, the predictions for reducing disease severity for each patient and body part were similar to the actual values. Our study verified the possibility of tracking dynamic changes in the disease and personalized treatment using a psoriasis treatment response prediction method.