A stroke risk score built for rheumatoid arthritis rather than borrowed

Design

cross-sectional risk model development with LASSO variable selection and external validation in a separate cohort

Population

1,366 participants from NHANES 2011-2020 for derivation; 774 patients with rheumatoid arthritis from a Beijing hospital cohort for validation and enhancement

Primary outcome

discrimination for prevalent stroke, measured as area under the receiver operating characteristic curve

Effect

basic model AUC 0.712 (0.665-0.759), external validation 0.716 (0.674-0.758); enhanced model 0.829 (0.794-0.864), sensitivity 70.0%, specificity 81.6%; Framingham 0.611

General-population cardiovascular calculators underestimate risk in rheumatoid arthritis, and the usual workaround — multiplying the answer by a fixed factor — is crude. This study built a risk identification model in two steps: LASSO regression on NHANES 2011-2020 data (1,366 participants) to select traditional factors, then external testing in a Beijing Tiantan Hospital rheumatoid arthritis cohort of 774, then a second model adding five disease-specific variables.

Eleven traditional factors were retained, with neutrophil count, hypertension and coronary heart disease carrying the strongest associations. The basic model reached an area under the curve of 0.712, holding at 0.716 on external validation. Adding DAS28-CRP, anti-CCP antibody, rheumatoid factor, methotrexate use and disease duration lifted it to 0.829, with sensitivity 70.0% and specificity 81.6%. It beat the Framingham Risk Score, which managed 0.611 in this population.

The design is the limitation and it is a real one: this is cross-sectional, so the model identifies patients who have had a stroke rather than predicting who will. It has not been tested prospectively, and the cohorts are American survey participants and a single Chinese hospital, neither of which is an Indian rheumatology clinic. What is nonetheless usable today is the variable list. Disease activity, serology and disease duration carried real independent information about stroke, which is an argument for treating uncontrolled inflammation as a vascular risk factor and recording it alongside blood pressure and lipids rather than in a separate part of the notes.

• Do not rely on Framingham or a general-population calculator alone in rheumatoid arthritis
• Record blood pressure, lipids, smoking and DAS28-CRP at the same review — they belong to one risk assessment
• Persistent disease activity is a modifiable vascular risk factor, so say so when discussing treatment escalation
• Note the neutrophil count: it was among the strongest traditional associations in this analysis
• The model is cross-sectional and unvalidated prospectively — use the variables, not the score

The statistics, in plain English

An area under the curve of 0.829 means that given one patient who had a stroke and one who did not, the model ranks them correctly about 83% of the time; 0.611 for Framingham is barely better than a coin toss weighted by age. The jump from 0.716 to 0.829 on adding disease-specific variables is large and statistically supported by both likelihood ratio and DeLong tests. The decisive caveat is that a cross-sectional model is fitted to strokes that have already happened, and variables such as methotrexate use and disease activity may be consequences of the event or of the care that followed it, which inflates apparent performance relative to genuine prediction.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.