Machine learning model sharpens diagnostic separation in suspected systemic lupus erythematosus

Systemic lupus erythematosus remains one of the most difficult autoimmune diseases to identify cleanly at the bedside. Its symptoms overlap with rheumatoid arthritis, Sjögren’s syndrome, undifferentiated connective tissue disease, and even nonspecific inflammatory syndromes, while routine laboratory testing rarely provides a single decisive answer. In that setting, a diagnostic tool that combines clinical data with immunoinflammatory markers has obvious appeal, especially when it can separate confirmed lupus from patients who look similar on first presentation.

This retrospective study built exactly that kind of model. Investigators assembled 1,950 symptomatic adults from a broader pool of 2,240 hospital patients evaluated for suspected lupus, then divided them into 1,000 confirmed systemic lupus erythematosus cases and 950 non-SLE controls. The control group was not a trivial comparator: it included patients with other autoimmune diseases, undifferentiated connective tissue disease, and individuals with suggestive symptoms but no confirmed autoimmune diagnosis. That design mirrors real diagnostic ambiguity more closely than a clean healthy-control comparison would.

Seven markers carried the final diagnostic signal

The modeling pipeline began with a wide panel of demographic, laboratory, comorbidity, and treatment variables, then narrowed the list through a dual feature-selection strategy using LASSO and Boruta random forest screening. Seven variables survived that process: antinuclear antibodies, C-reactive protein, interferon-gamma inducible protein 10, tumor necrosis factor-alpha, interleukin-6, B-cell activating factor, and anti-double-stranded DNA. Each marker fits the biology of lupus in a different way. Autoantibodies capture the loss of immune tolerance, while cytokines and chemokines reflect the inflammatory amplification that drives tissue injury and symptom burden.

The final model that performed best was logistic regression, chosen not because it was the most complex option but because it balanced performance with interpretability. In a disease where clinicians need to justify why one patient is more likely to have lupus than another, a transparent scoring architecture has real value. A black-box classifier may look impressive on paper, yet it can be hard to trust in a mixed autoimmune clinic where treatment decisions often hinge on subtle distinctions.

The internal validation cohort showed the integrated model reached 81.0% specificity, 83.0% sensitivity, 82.0% accuracy, and an area under the receiver operating characteristic curve of 0.851. Those results outperformed individual predictors and also surpassed more complex algorithms that were more vulnerable to structural overfitting. In practical terms, the model separated lupus from its mimics with a level of consistency that is difficult to obtain from a single antibody assay alone.

A cytokine-only panel stayed useful when autoantibodies were removed

One of the more clinically interesting findings came from the autoantibody-independent analysis. The cytokine-only panel, built from BAFF, IL-6, TNF-α, IP-10, and CRP, still performed well, producing an AUC of 0.820 in internal validation. It also outperformed the conventional ANA and anti-dsDNA baseline, which reached an AUC of 0.765. In plain clinical terms, inflammatory biology still carried useful diagnostic information even after the classic autoantibody signals were removed.

This matters because antibody testing can be confounded by timing, phenotype, and the reality that not every symptomatic patient presents with a textbook serologic profile. A cytokine-based profile may help when serologies are incomplete, equivocal, or lagging behind the clinical picture. The inflammatory axis in lupus is not decorative; it is part of the disease engine, and markers such as BAFF and IL-6 are plausible readouts of that biology.

Decision curve analysis also favored the integrated approach over standard diagnostic baselines across a wide range of threshold probabilities. That finding suggests the model was not only statistically stronger, but also more likely to deliver net clinical benefit when used to guide diagnostic decisions. For clinicians faced with a patient who has fatigue, arthralgia, rash, or nonspecific laboratory abnormalities, that kind of incremental certainty can shorten the path to definitive evaluation.

Why the model may fit bedside decision-making

The study population was drawn from routine hospital visits between January 2019 and June 2024, and the serum samples were analyzed in a single batch using the same ELISA kits. Missing data were sparse and handled with k-nearest neighbors imputation after Little’s test suggested missingness was completely at random. The investigators also tried to reduce data leakage by confining feature selection to the training cohort and keeping the validation set isolated until the end. Those design choices strengthen confidence that the reported performance was not inflated by sloppy modeling.

The model was also translated into a graphical scoring index, which is important because diagnostic tools tend to fail when they are too cumbersome to use in real clinics. Logistic regression lends itself to that kind of translation. A physician can more easily understand a weighted contribution from ANA, anti-dsDNA, BAFF, IL-6, TNF-α, IP-10, and CRP than from an opaque ensemble output. In autoimmune practice, interpretability often determines whether a model becomes part of workflow or remains an academic exercise.

External validation remains the next clinical test

The study’s strongest limitation is also its most obvious one: the results come from a retrospective, single-setting cohort and still need prospective multicenter validation. The control group was carefully constructed, but all participants came from a symptomatic population already under evaluation for suspected lupus, so the model has not yet been tested across broader referral patterns or community settings. Generalizability will depend on whether the same markers perform with similar reliability in different laboratories, different patient mixes, and different pretest probabilities.

Another practical issue is incorporation bias. The investigators addressed this by evaluating an autoantibody-independent cytokine panel, but routine lupus diagnosis in everyday practice still often leans heavily on ANA and anti-dsDNA. Whether the biomarker combination can improve diagnostic confidence without adding cost or turnaround burden will determine how useful it becomes outside a research environment.

The DOI for the paper is 10.3389/fimmu.2026.1997874. The work suggests that a machine learning-guided panel built from immune and inflammatory biomarkers can support differential diagnosis in suspected lupus, but the real test now is whether the same performance holds in prospective external cohorts where the clinical noise is higher and the diagnostic stakes are just as unforgiving.


Reporting by Dr. A. Nikhil Krishnan, MBBS for Medicine World Foundation, with additional data provided by Frontiers in Immunology.
PRIMARY CLINICAL EVIDENCE & REGISTRY DATA