AI11 Harnessing artificial intelligence-based image analysis and self-reported data for the assessment of psoriasis severity

Abstract Psoriasis severity assessment currently requires in-person evaluation using scoring tools such as Psoriasis Area and Severity Index (PASI), which have limitations including inter- and intraobserver variability. Artificial intelligence (AI)-based analyses of images offer opportunities for more objective and efficient disease severity assessments, but the prediction accuracy of these AI tools relies on high-quality, diverse image datasets. Challenges to accruing such datasets include data privacy concerns, poor representation of real-world populations, and limited annotation of images. We developed a novel vision transformer (ViT) called ‘MultiViT’ for AI-based automation of PASI assessments using a real-world image dataset that was prospectively acquired in a UK national specialized psoriasis service following informed consent (April 2022 to May 2024, REC ref 20/YH/0135). The dataset comprised 1109 professional studio and self-taken photographs from 152 participants, with matched detailed clinical and demographic data. Most participants were male (63%) and had Fitzpatrick skin type (FST) I–IV (84%). Their median age was 45.5 years (interquartile range 33.8–57.0) and the median body mass index (BMI) was 29.7 kg m−2. MultiViT was trained and evaluated over 300 epochs, with an 80 : 20 training–validation split of the image dataset. The mean absolute error (MAE) was calculated to assess the PASI prediction accuracy of MultiViT compared with a standard ViT. MultiViT achieved a lower MAE (3.88) for PASI prediction than the standard ViT (MAE 5.13). The performance of MultiViT improved when assessing images of individuals with FST I–IV (MAE 2.80) compared with FST V–VI (MAE 8.17). Model performance was also greater for assessment of less severe disease: MAE 1.97 for Physician’s Global Assessment (PGA) ≤ 3, vs. MAE 9.13 for PGA > 3. To expand our resource of self-taken images for enhanced MultiViT performance, mySkin (myskin.org), an online codesigned patient self-report platform, was launched in July 2023. Patient-reported outcomes are collected quarterly, alongside standardized self-taken images. As of 31 January 2025, 900 individuals with psoriasis have been recruited to mySkin, with 1181 self-taken images from 271 individuals collected to date. Their demographics are 58.7% female, 87.1% White ethnicity, 74.9% FST II–III; mean age 51.3 years (SD 13.5), average age of psoriasis onset 25.2 years (SD 15.8) and mean BMI 29.2 kg m−2 (SD 6.9). Our findings from a limited image dataset indicate the potential of MultiViT for automating psoriasis severity assessments from skin images. A larger and more diverse image dataset (severity, skin type), enabled by mySkin, will further advance model performance, enhancing generalizability for real-world applications. MultiViT has potential for future integration into a smartphone-based patient-centred tool to enable remote monitoring and personalized care for people with psoriasis.