Background/Purpose: Detection of relevant autoantibodies is key in the identification of autoimmune connective tissue diseases (CTD). The evaluation of multiple autoantibodies for extended serological profiling may improve the diagnosis of these conditions. We evaluated the diagnostic utility, in patients with CTD and disease controls, of machine learning classifiers based on the 15-autoantibody profile performed by a novel, single-use, multiplexed microarray immunoassay, used with its fully automated high-throughput proprietary system for simultaneous the detection of IgG autoantibodies directed to dsDNA, SS-A 60, TRIM21 (SS-A 52), SS-B, Sm, Sm/RNP, U1RNP, Jo-1, Scl-70, Centromere B, Chromatin, Ribosomal P, DFS70, RNAP III and CCP2.

Methods: De-identified sera from 475 patients diagnosed with autoimmune CTD in accordance with current guidelines [127 patients with systemic lupus erythematosus (SLE), 74 with systemic sclerosis, 76 with Sjögren’s syndrome (SjS), 71 with idiopathic inflammatory myopathies, 54 with mixed CTD, 73 with rheumatoid arthritis] and 652 patients with other disorders, who served as disease controls were analyzed using the investigational MosaiQ AiPlex® CTDplus (AliveDx, Switzerland) assay. Classification models were developed using all 15 autoantibodies or a selected subset, employing the RandomForest algorithm (XGBoost and Logistic L1 L2 were also used). Diagnostic performance was evaluated by receiver operating characteristic (ROC) curve analysis.

Results: A RandomForest classifier incorporating all 15 autoantibodies demonstrated robust performance in predicting SLE, achieving an area under the curve (AUC) of 0.92. In comparison, the individual SLE-specific markers dsDNA and Sm yielded lower AUCs of 0.68 and 0.60, respectively. For SjS, the 15-plex RandomForest classifier achieved an AUC of 0.83, outperforming a 3-plex RandomForest classifier based on SS-A 60, TRIM21, and SS-B autoantibodies, which had an AUC of 0.62. The individual AUCs for these markers were 0.63, 0.59, and 0.58, respectively. Similarly, for other CTDs, 15-plex classifiers consistently outperformed the individual disease-specific markers.

Conclusion: Multiplex autoantibody testing combined with machine learning algorithms has the potential to improve the diagnosis of autoimmune CTD.

Figure 1. Model Pipeline

Figure 2. ROC Curves and AUCs – SLE: All 15 analytes vs. 9 Outcome-related analytes*

Figure 3. ROC Curves and AUCs – SjS: All 15 analytes vs. 3 Outcome-related analytes*

ACR Meeting Abstracts - https://acrabstracts.org/abstract/a-machine-learning-algorithm-based-on-a-15-autoantibody-profile-by-a-novel-fully-automated-multiplexed-microarray-immunoassay-for-the-diagnosis-of-autoimmune-connective-tissue-diseases/