Background: Antinuclear antibodies (ANAs) play a crucial role in diagnosing systemic autoimmune rheumatic diseases, particularly systemic lupus erythematosus. The recommended standard for ANA detection is indirect immunofluorescence testing (IIFT) using human epithelial (HEp-2) cells. Since visual interpretation (VI) of IIFT images is time-consuming and labor-intensive, research is focusing on automated interpretation systems that use artificial intelligence (AI). Methods: Consecutive serum samples (number of sera = 143) from routine clinical care were collected from patients visiting our tertiary rheumatology center. ANA were detected by IIFT with visual interpretation and compared with IIFT using the AI-based interpretation system akiron® NEO (Medipan, 15827 Blankenfelde-Mahlow, Germany). ANA titer levels and patterns were analyzed according to the Competent Level of the International Consensus on ANA Pattern classification. Results: Agreement of positive/negative ANA discrimination between AI-aided and VI-IIFT at the recommended cut-off of 80 was good (Cohen’s kappa κ 0.69) but significantly different (McNemar test, p < 0.0001). At a cut-off of ≥1/80, the agreement was improved (κ 0.76) and the difference between both methods was non-significant (p = 1.0000). The ANA pattern recognition agreement between both approaches was moderate (κ = 0.54). The direct comparison using only the akiron® NEO HEp-2 cell ANA assay revealed a good agreement (0.67), which improved to very good (κ = 0.80) when differences between ANA patterns anti-cell (AC)4/5 and AC2 were neglected. Notably, titer levels in the automated evaluations were frequently assessed at higher values than in the gold standard interpretation. Conclusions: Our study demonstrates a good agreement for positive/negative ANA discrimination. ANA pattern recognition by AI-aided interpretation showed moderate to very good agreement with VI. Further research and algorithm refinement (e.g., improved pattern recognition and titer calibration) are necessary to support its future implementation as a reliable screening method.
Mayr et al. (Tue,) studied this question.