Home > Rheumatology > ACR 2024 > Rheumatoid Arthritis > Machine learning to aid evaluation of ANA pattern and titer

Machine learning to aid evaluation of ANA pattern and titer

Presented by
Dr May Choi, University of Calgary, Canada
Conference
ACR 2024
Doi
https://doi.org/10.55788/d6029be1
A Canadian group of researchers developed 8 machine-learning models to evaluate anti-nuclear antibody (ANA) patterns and titers, identifying 1 model with the best performance for ANA pattern identification. They conclude that machine learning could become a highly effective and efficient aid to reduce variability, increasing laboratory accuracy and efficiency.

The presence of ANAs is a classification criterion for RA and other systemic autoimmune rheumatic diseases (SARD), making ANA immunofluorescence (IFA) patterns and titers a key part of diagnostics. Dr May Choi (University of Calgary, Canada) and colleagues developed 8 machine-learning models to aid laboratories in ANA pattern and titer interpretation to reduce the considerable variability within and between laboratories that occurs with manual interpretation [1]. One model specifically assessed the nuclear dense fine-speckled (DFS) ANA pattern (AC-2), a rare pattern that decreases the likelihood of SARD.

To compare the performance of the 8 machine-learning models in recognising 13 ANA patterns, the researchers fed their models 13,671 ANA images from patients enrolled in the Systemic Lupus International Collaborating Clinics Inception Cohort (SLICC, n=2,825 images), non-SLE subjects enrolled in the Ontario Health Study (OHS, n=10,639 images), and the International Consensus on ANA Patterns (ICAP, n=207 images). As the reference standard, a highly experienced laboratory technologist interpreted ANA patterns and titers for each image.

The ANA Reader© model had the best performance of all 8 machine-learning techniques. It had the highest area under the curve (AUC) score of 83.4%, albeit with modest performance in other metrics: weighted accuracy was 68.4%, precision 67.1%, sensitivity 70.1%, and F1 score 67.2%. The ANA patterns with the best performance were centromere (AUC 0.97) and pleomorphic patterns (AUC 0.97).

All 4 developed convolutional neural network (CNN) models performed similarly, with high AUC scores (96.5%–97.1%). These models were meant to differentiate AC-2, which is not associated with SARD, from 2 similar patterns (AC-4 and AC-30), which are associated with SARDs.

There was a strong correlation between titers reported by the identified model and the technologist’s interpretation (Spearman rank 0.93; P<0.0001). In most cases (96.0%), reported titers were identical or differed by only one dilution.

Dr Choi expects the performance of the ANA Reader© to improve further as they continue to train it with more images. The AC-2 model can potentially speed up the differentiation of people who are and are not at risk of SARDs. The group is working on external validation studies and new machine-learning models.

  1. Choi M, et al. Rheumatology diagnostics utilizing artificial intelligence (ANA Reader©) for ANA pattern identification and titer quantification. Abstract L11, ACR Convergence 2024, 14–19 November, Washington DC, USA.

Copyright ©2025 Medicom Medical Publishers



Posted on