B-140 Performance and Workflow Impact of Automated Image Analysis in High-Volume ANA HEp-2 IIF Screening

Abstract

Indirect immunofluorescence (IIF) on HEp-2 cells remains the reference standard for ANA screening, providing semi-quantitative titers and clinically relevant fluorescence patterns. Despite its high sensitivity, it lacks specificity, and interpretation is influenced by population characteristics, analytical variability, and observer-dependent subjectivity. Manual microscopy is labor-intensive and prone to inter-observer variability, particularly in pattern attribution and titer estimation, prompting increasing adoption of computer-aided and AI-assisted platforms. Although prior studies have shown good agreement between some automated systems and expert visual interpretation for overall positivity and titers, incomplete pattern concordance and calibration limitations persist, underscoring the need for real-world comparison of automated versus manual ANA HEp-2 IIF interpretation. A retrospective cross-sectional method-comparison study included consecutive routine ANA HEp-2 IIF tests performed over a two-month period in 2025 at a reference laboratory in Fortaleza, Brazil. Specimens were processed using the SPRINTER XL (EUROIMMUN) and analyzed through the EUROPattern Microscope Live (EPA Live) workflow. Automated outputs—including ANA positive/negative classification, nuclear pattern grouping, and titer category—were reviewed in routine practice by an expert analyst with access to both the automated interpretation and digitized images, with modifications permitted prior to final sign-out. Preliminary automated results were subsequently extracted and compared with the final expert-reported interpretations, which served as the reference standard. Among 5,721 routine ANA HEp-2 IIF evaluations, using expert manual interpretation as the reference standard, the automated classifier yielded 2,469 true positives (TP), 2,684 true negatives (TN), 110 false positives (FP), and 458 false negatives (FN), corresponding to a sensitivity of 0.8435 (Wilson 95% CI, 0.8299–0.8562), specificity of 0.9606 (0.9528–0.9672), positive predictive value of 0.9573, negative predictive value of 0.8542, and overall accuracy of 0.9007. Agreement beyond chance was high (Cohen’s ?=0.8019), consistent with substantial-to-near-perfect concordance. Discordances were predominantly driven by false negatives (458 vs 110 false positives), which clustered at low titers—primarily 1:80 and 1:160—and occurred mainly within the nuclear pattern group, with a smaller contribution from cytoplasmic patterns. The three patterns most frequently associated with false negatives were AC-4 nuclear fine speckled at 1:80 (98/458; 21.4%), AC-2 dense fine speckled at 1:80 (54/458; 11.8%), and AC-2 dense fine speckled at 1:160 (39/458; 8.5%). False-positive classifications were uncommon and, by definition, lacked expert-confirmed pattern annotation. Automated false positives were primarily low-titer (predominantly 1:80) and were frequently associated with nucleolar or “other” automated pattern assignments. In a high-volume routine ANA HEp-2 IIF setting, automated classification demonstrated substantial agreement with expert manual interpretation and excellent specificity, supporting its role as a reliable high-throughput screening adjunct. Discordances were primarily driven by false negatives concentrated at low titers (=1:160), particularly within nuclear speckled phenotypes (AC-4 and AC-2). These findings identify borderline low-titer cases as the principal optimization target and support a risk-based review strategy prioritizing expert adjudication of low-intensity and pattern-sensitive results. Refinement of fluorescence thresholds and algorithm calibration may further enhance sensitivity while preserving specificity and workflow efficiency.