An AI model applied to HEp-2 ANA indirect immunofluorescence (IIF) images was developed and temporally validated to predict seven specific autoantibodies (anti-dsDNA, Sm, RNP, SSA, SSB, Scl-70, centromere) in ANA-positive patients (titer ≥1:80) across a development cohort (Dec 2016–Dec 2022, n=47,890) and a prospective validation cohort (Jan 2023–Aug 2024, n=21,427).
Models for CMA, dsDNA, Sm, SSA, and SSB achieved AUROC ≥0.80; CMA performed best (AUROC 0.95, accuracy 94.2%, specificity 95.3%). AI outperformed IIF pattern comparators in ruling out antibody-negative cases for dsDNA (88.2% vs. 54.7% correctly classified negative) and Sm (82.6% vs. 60.3%); RNP (AUROC 0.68) and Scl-70 (AUROC 0.59) performed poorly. dsDNA discrimination remained stable across assay transitions (AUROC ~0.85).
- RNP and Scl-70 models showed poor discriminative performance (AUROC 0.68 and 0.59), limiting clinical utility for these antibodies. - Study is single-institution, which may affect generalizability to labs with different IIF platforms or staining protocols. - Pattern-based comparators used were simplified, potentially underestimating skilled human reader performance.
AI analysis of ANA IIF images can reliably predict centromere, dsDNA, Sm, SSA, and SSB antibodies, and may reduce the need for confirmatory reflex testing in many patients. Clinicians should be cautious about relying on AI IIF analysis alone for RNP or Scl-70, where model performance remains insufficient.
Explore related topics
