Utilization of Clinical Trial Imaging for Accurate Psoriasis Area and Severity Index Prediction

According to a study published in Digital Biomarkers, utilizing image processing and deep learning can potentially translate inaccessible clinical trial data into accurate, extensible machine learning models to assess therapeutic efficacy.

Researchers leverage longitudinal imaging data from the Ultima 2 clinical trial (NCT02684357) involving patients with psoriasis undergoing treatment. This dataset comprised 2700 body images annotated with Psoriasis Area and Severity Index (PASI) scores by uniformly trained dermatologists.

The researchers developed an image-processing workflow, termed the "One-Step PASI" framework, which integrates clinical photos of multiple body regions into a single model pipeline. This framework simultaneously performs body and lesion detection, as well as lesion severity classification. Group-stratified cross-validation was conducted using 145 deep convolutional neural network models combined in an ensemble learning architecture.

Results indicated that the highest-performing model achieved impressive accuracy metrics, including a mean absolute error of 3.3, Lin's concordance correlation coefficient of 0.86, and Pearson correlation coefficient of 0.90 across various PASI scores representing disease severity classifications. Moreover, within-person, time-series analysis demonstrated that the model accurately tracked changes in PASI scores over time, closely mirroring physician assessments from severe to clear skin without systematic over- or underestimation.

“This study demonstrates the potential of image processing and deep learning to translate otherwise inaccessible clinical trial data into accurate, extensible machine learning models to assess therapeutic efficacy,” the authors concluded.

Reference
Xing Y, Zhong S, Aronson SL, et al. Deep learning-based psoriasis assessment: harnessing clinical trial imaging for accurate psoriasis area severity index prediction. Digit Biomark. Published online March 4, 2024. doi:10.1159/000536499