POS1055 IDENTIFICATION OF PsA PHENOTYPES WITH MACHINE LEARNING ANALYTICS USING DATA FROM A PHASE 3 CLINICAL TRIAL PROGRAMME OF GUSELKUMAB IN A BIO-NAÏVE PATIENT POPULATION

Abstract

Background: Psoriatic arthritis (PsA) is typically described by its individual domains or clinical components.1,2 Objectives: This post hoc analysis aimed to identify hypothesis-free phenotype clusters according to patients' clinical features and baseline (BL) characteristics with data from the Phase 3 DISCOVER-1 and-2 guselkumab (GUS) clinical trials. Method(s): Data from bio-naive patients with PsA treated with GUS 100 mg every 4 or 8 weeks in DISCOVER-1 and-2 were retrospectively analysed. Non-negative matrix factorisation was used as an unsupervised machine learning technique to identify clusters of PsA phenotypes, with BL characteristics and clinical observations as input features, according to which clusters were described. Result(s): Data from 661 patients were pooled and 8 distinct clusters of PsA phenotypes identifed (Table 1). Cluster 1 was characterised by lower limb involvement and the lowest rates of severe skin involvement (Figure 1); Cluster 2 by high skin involvement, the lowest proportion of women and highest proportion of overweight patients (body mass index [BMI] 25-<30, 70%); and Cluster 3 by high burden of disease in the hand/wrist. In Cluster 4 all patients had dactylitis and >=3% body surface area (BSA) psoriasis involvement and the second highest proportion of men. Cluster 5 had the highest BL enthesitis rate; large joint involvement was also common. Cluster 6 had a high level of small joint involvement in the hands/feet, but low mean dactylitis score; nail involvement and BL enthesitis were also common. In Cluster 7, all patients had axial involvement at BL, 49.4% had dactylitis, 69.9% had enthesitis and most had BSA >=3% (Figure 1). Cluster 8 had limited joint involvement, extensive skin involvement and the highest proportion of obese patients (BMI >30, 67%). Minimal disease activity (MDA) response rates at Week (W)24 and W52 were highest in Cluster 2 and lowest in Cluster 5. Clusters 3 and 4 had low MDA response rates at W24, increasing at W52. Conclusion(s): Unsupervised machine learning identifed 8 clusters of PsA pheno-types with signifcant differences in demographic and clinical features, including patterns of domain involvement and MDA responses. These clusters differ in their initial vs. later responses to GUS.