A Bimodal Artificial Intelligence Model for Diagnosis and Risk Stratification of Generalized Pustular Psoriasis

Abstract

Introduction: Generalized pustular psoriasis (GPP) is a rare, burdensome psoriasis subtype. Difficulty recognizing pustules may delay diagnosis and treatment. Objectives: In this study, we aimed to develop an artificial intelligence (AI) model simulating dermatologists’ diagnostic workflow for GPP. Methods: Our model combined visual analysis with a clinical questionnaire module, using the ImageNet-22k pre-trained Swin Transformer-Tiny as its backbone. For cases classified as GPP, a questionnaire based on expert clinical experience was triggered for risk stratification. The model was evaluated against dermatologists and other AI models and deployed as a WeChat prototype. Results: A total of 367 patients and 937 images were included, comprising 148 healthy individuals, 72 with psoriasis vulgaris, and 147 with GPP. The model yielded an overall accuracy of 79.7%, a sensitivity of 79.5%, a specificity of 91.1%, and a micro-averaged AUC of 0.848, delivering superior image recognition performance across all evaluated architectures. We further conducted a human-machine comparison study using the WeChat prototype. The model achieved 60.0% accuracy, 62.5% sensitivity, and 50.0% specificity in visual GPP screening, showing a relatively balanced performance. In the final risk assessment, the AI model achieved 20.0% accuracy, representing a lower exact agreement with the reference standard than that of dermatologists. Its quadratic weighted κ was 0.351 (95%CI: 0.102-0.599, p=0.003 ), a value numerically comparable to that of some participating dermatologists. Conclusions: This multimodal framework provides a proof of concept for AI-assisted GPP recognition and risk assessment. Further validation and optimization are needed to expand its clinical applicability.