To aid etiology and treatment research of the very heterogeneous rheumatoid arthritis (RA) population, we aimed to identify phenotypically distinct RA subsets using baseline clinical data. We collected baseline numerical- (hematology work-up & age) and categorical variables (serology, joint location & sex) from the Electronic Health records (EHR) repository of the Leiden University Medical Center, comprising 1,387 unique first visits to the outpatient clinic. We used deep learning and graph clustering to identify phenotypically distinct RA subsets. To ensure the robustness of our findings, we tested a) cluster stability (1000 fold) b) physician confounding, c) association with remission and methotrexate failure, d) generalizability to a second different data set (the Leiden Early Arthritis clinic; n=769). In total we identified four subsets (C1-C4) of patients with rheumatoid arthritis that were delineated on the following characteristics: C1) arthritis in feet, C2) seropositive oligo-articular disease, C3) seronegative hand arthritis, C4) polyarthritis. Our validity analyses showed high stability (mean 78%-91%), no physician confounding, and a significant difference in methotrexate failure(P-value=6.1e-4) and occurrence of remission(P-value=7.4e-3), and generalizability to a second dataset. The hand-cluster (III) had the most favorable outcomes (HR remission=1.65 (95%CI:1.20-2.29), HR methotrexate=0.48 (95%CI:0.35-0.77)), particularly the ACPA-positive patients in this cluster, while in the other clusters the ACPA-negative patients did best. The clusters outperformed standard clinical variables, which were attributed to the hand and feet differentiation. In conclusion, we discovered four phenotypically distinct subgroups of rheumatoid arthritis at baseline that associate with clinical outcomes. Furthermore, our study provides evidence for the presence of separate hand and foot subgroups in RA.