Evaluation of a machine learning tool for the early identification of patients with undiagnosed psoriatic arthritis – A retrospective population-based study

• Published in: Journal of translational autoimmunity (Online) Vol. 7; p. 100207
• Main Authors: Shapiro, J., Getz, B., Cohen, S.B., Jenudi, Y., Underberger, D., Dreyfuss, M., Ber, T.I., Steinberg-Koch, S., Ben-Tov, A., Shoenfeld, Y., Shovman, O.
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.12.2023; Elsevier
• Subjects: Artificial intelligence; Early diagnosis; Machine learning; Psoriasis; Psoriatic arthritis; Research paper; Early diagnosis; Psoriasis; Artificial intelligence; Machine learning; Psoriatic arthritis
• DOI: 10.1016/j.jtauto.2023.100207
• ISSN: 2589-9090

Psoriatic arthritis (PsA), an immune-mediated chronic inflammatory skin and joint disease, affects approximately 0.27% of the adult population, and 20% of patients with psoriasis. Up to 10% of psoriasis patients are estimated for having undiagnosed PsA. Early diagnosis and treatment can prevent irreversible joint damage, disability and deformity. Questionnaires for screening to identify undiagnosed PsA patients require patient and physician involvement. To evaluate a proprietary machine learning tool (PredictAI™) developed for identification of undiagnosed PsA patients 1–4 years prior to the first time that they were suspected of having PsA (reference event). This retrospective study analyzed data of the adult population from Maccabi Healthcare Service between 2008 and 2020. We created 2 cohorts: The general adult population (“GP Cohort”) including patients with and without psoriasis and the Psoriasis cohort (“PsO Cohort”) including psoriasis patients only. Each cohort was divided into two non-overlapping train and test sets. The PredictAI™ model was trained and evaluated with 3 years of data predating the reference event by at least one year. Receiver operating characteristic (ROC) analysis was used to investigate the performance of the model, built using gradient boosted trees, at different specificity levels. Overall, 2096 patients met the criteria for PsA. Undiagnosed PsA patients in the PsO cohort were identified with a specificity of 90% one and four years before the reference event, with a sensitivity of 51% and 38%, and a PPV of 36.1% and 29.6%, respectively. In the GP cohort and with a specificity of 99% and for the same time windows, the model achieved a sensitivity of 43% and 32% and a PPV of 10.6% and 8.1%, respectively. The presented machine learning tool may aid in the early identification of undiagnosed PsA patients, and thereby promote earlier intervention and improve patient outcomes.