Deep Learning-Based Diagnostic Support for Complex Autoimmune Syndromes: A Multi-label Classification Study

Abstract

Autoimmune conditions frequently show similar clinical signs and symptoms, making it difficult to accurately diagnose when there are coexisting conditions. This study proposes a computational framework that aims to simultaneously predict multiple autoimmune labels using routine clinical data and advanced serological tests. The study examined 13,812 patients with 79 diagnostic features, including hematologic tests, inflammatory tests, and specific autoantibodies. The study trained a multi-layer perceptron (MLP) to predict autoimmune conditions such as Systemic Lupus Erythematosus (SLE), Rheumatoid Arthritis, and Graves' Disease. The model achieved strong performance, with validation F1 scores starting at 0.4723 and near-perfect Area Under the Curve (AUC) values ranging from 0.98 to 1.00 for all categories. The 79-parameter approach was highly effective in describing the autoimmunity “mosaic” and identifying polyautoimmunity with high precision compared to contemporary models that use fewer autoantibodies. The study shows that Multi-Label Multi-Layer Perceptron (ML-MLP) are able to process different types of data from lab tests and provide high utility diagnostic support for complex and overlapping autoimmune conditions.