Data-driven discovery of candidate predictors of future rheumatoid arthritis diagnosis in the UK biobank.

BACKGROUND: Rheumatoid arthritis (RA) is an inflammatory autoimmune disease in which the immune system attacks joint tissues. Identifying predictors of RA diagnoses may enable earlier detection and opportunities for prevention. METHODS: We applied a hypothesis-free data-driven machine learning approach using 445,515 UK Biobank participants, including 4,510 incident cases (8.4 median years follow-up, IQR 5.3 - 10.9). From 2,898 baseline input features, those identified as potentially important by the machine learning model were taken forward to logistic regression analyses, adjusting for known confounders. We used a Bonferroni corrected p-value of p<3.0 × 10-4. RESULTS: The model identified 200 features as potentially important for prediction of RA diagnosis. Epidemiological analyses confirmed associations with known RA risk factors (older age, female sex, smoking, and physical inactivity) and some indicators of low socioeconomic status and psychosocial well-being (e.g. highest vs lowest neuroticism score OR 1.44, 95% CI 1.24-1.67). History of joint disorder, osteoporosis, hypothyroidism, diverticular disease, and emphysema/chronic bronchitis were each associated with a 48% to 230% higher odds of RA. Markers of inflammation (e.g. C-reactive protein Q5 vs Q1 OR 1.92, 95% CI 1.71-2.16), altered liver and kidney function (e.g. cystatin C Q5 vs Q1 OR 1.55, 95% CI 1.38-1.74), and a range of blood cell markers were also found to associate with the odds of RA. CONCLUSION: Our data-driven analyses identified lifestyle, psychosocial, clinical and biomarker features predictive of RA years before diagnosis. While external validation is required, many identified candidate predictors likely reflect prodromal RA and may aid earlier disease detection, while modifiable lifestyles could support prevention. Further studies are required to confirm causality and clinical relevance.