- Design
- Post hoc risk-prediction model derivation from a randomised trial
- Population
- 649 PEXIVAS participants with severe ANCA-associated vasculitis in remission
- Primary outcome
- First relapse within 2 years of remission
- Effect
- 100 relapses; optimism-corrected AUC 0.709; calibration slope 0.99 (95% CI 0.65–1.34)
PEXIVAS randomised patients with severe ANCA-associated vasculitis — eGFR below 50 or diffuse alveolar haemorrhage — to plasma exchange and to reduced or standard glucocorticoid dosing. This post hoc analysis took the 649 participants who reached remission, of whom 100 relapsed within two years, and built a model to estimate each person's risk of first relapse.
Six features were retained: sex, induction with oral cyclophosphamide or rituximab, ANCA subtype, non-haemorrhagic respiratory involvement, mucous membrane or eye involvement at presentation, and eGFR at remission. After correction for over-optimism, discrimination was fair (AUC 0.71) and calibration was close to ideal (slope 0.99).
Relapse risk drives decisions about how long to continue maintenance therapy and how closely to monitor. A model using information already in the notes at remission gives that discussion a structure. It was derived in severe, trial-enrolled disease, has not been externally validated, and with an AUC of 0.71 it will misclassify a fair number of patients — so it informs, rather than decides, maintenance.
- Record ANCA subtype, induction agent, organ involvement at presentation and eGFR at remission in every vasculitis patient.
- Consider a structured relapse risk estimate when planning maintenance duration and monitoring.
- Monitor more closely in the first two years after remission, when relapse was concentrated.
- Do not stop maintenance on a low estimate alone; the model is not yet externally validated.
- Note that the model was derived in severe disease with kidney or lung involvement.
Why it matters
It turns a general sense of who relapses into an estimate that can be discussed and recorded.
Don't overread it
The model is derived from one trial's participants and is not externally validated.
The statistics, in plain English
An AUC of 0.71 means the model ranks a patient who will relapse above one who will not about 71% of the time — useful but imperfect. A calibration slope of 0.99 means predicted risks matched observed risks well in this dataset; that usually worsens in new populations.
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.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for rheumatology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free