Tree-Based Reinforcement Learning for Identifying Optimal Personalized Treatment Decisions for Hand Deformity in Rheumatoid Arthritis

Silicone metacarpophalangeal (MCP) joint arthroplasty (SMPA) can reconstruct MCP joint deformities in the rheumatoid hand, but patient selection criteria for the procedure remains unclear. The authors used statistical learning to elucidate patient selection criteria that will enhance long-term patient-reported and functional outcomes in patients with severe hand rheumatoid arthritis (RA). This was a secondary analysis of a prospective study of 169 adults with severe hand RA (average combined ulnar deviation and extensor lag at the MCP joint ≥50 degrees, per finger) with 1-year follow-up, conducted at 3 centers in the United States and England from January 1, 2004, to December 31, 2011. Primary outcomes were Michigan Hand Outcomes Questionnaire pain subscore, changes in extensor lag and ulnar deviation, and Arthritis Impact Measurement Scale score. A tree-based reinforcement learning model was used to estimate clinical decision rules for treatment. A total of 132 patients (mean ± SD age, 61 ± 9 years; 108 female patients [72%]) were included in the SMPA ( n = 50) and non-SMPA ( n = 82) cohorts. To minimize ulnar deviation and extensor lag, patients should undergo SMPA. To minimize pain, patients older than 55 years should undergo SMPA. To increase hand-related quality-of-life (QOL), patients with grip strength less than 12 kg should undergo SMPA. Estimations with imputed missing data were similar, aside from a lower grip strength (