This article is a narrative review that summarizes and critically appraises published applications of AI, ML, and DL in major autoimmune diseases (SLE, RA, IBD, MS, and T1DM). The review focuses on clinically relevant tasks (diagnosis, risk stratification, flare prediction, treatment response, and monitoring) and highlights translational requirements for safe deployment.

We searched PubMed/MEDLINE, Scopus, and Google Scholar from database inception to December 31, 2025 (final search date: December 31, 2025). The PubMed search string was: (“autoimmune” OR “autoimmune disease” OR “rheumatoid arthritis” OR “systemic lupus erythematosus” OR “multiple sclerosis” OR “inflammatory bowel disease” OR “ulcerative colitis” OR “Crohn” OR “type 1 diabetes”) AND (“machine learning” OR “deep learning” OR “artificial intelligence” OR “neural network” OR “support vector machine” OR “random forest” OR “gradient boosting” OR “XGBoost” OR “natural language processing”). Equivalent adaptations were used for Scopus and Google Scholar. Reference lists of included papers and relevant reviews were hand-searched for additional eligible studies.

We included peer-reviewed original studies that developed, validated, or evaluated AI/ML/DL models using human data in autoimmune diseases and reported measurable model performance (e.g., AUC, sensitivity/specificity, accuracy, F1, or calibration metrics). We excluded editorials, commentaries, purely theoretical/simulation studies without human data, and abstracts without sufficient methodological details.

Two reviewers independently screened titles/abstracts and then full texts; disagreements were resolved by discussion. We extracted: disease indication, clinical task, data modality (imaging, EHR/NLP, omics, wearable/proxies, multimodal), cohort size and setting (single vs multicenter), model family, feature engineering, validation approach (internal split or cross-validation, external validation, prospective testing), key discrimination metrics, and whether calibration and explainability were reported.

To strengthen methodological transparency, each included predictive model study was appraised using a structured tool suitable for clinical prediction models (e.g., PROBAST for risk of bias and applicability), with additional ML-specific checks (data leakage prevention, proper separation of training/validation/test sets, handling of class imbalance, and reporting of hyperparameter tuning). We summarize common risks (single-center development, limited external validation, and overreliance on accuracy) and their impact on clinical readiness.

Given heterogeneity in diseases, endpoints, and validation strategies, we performed a qualitative synthesis. We organized evidence by disease and clinical task and provided standardized comparative tables of modalities, model types, validation, and performance. Where claims of very high performance are reported, we explicitly note whether external validation, calibration, and prospective evaluation were performed.

Tables 1–3 present study characteristics, model/validation features, and performance/readiness indicators across diseases.

Fundamentals of ML in Autoimmune Disease: In the field of autoimmune disease, AI/ML has contributed significantly to diagnosis and management. Considering ML, it is a subset of AI and primarily learns and improves itself from data rather than programming. In the autoimmune disease, ML supports diagnosis, biomarker selection, and correlation with EHRs, imaging repositories, and multi-omic records.³ ML can thus distinguish between SLE and RA more accurately than the traditional approach and analyze the image to predict the probability of bone involvement in RA in advance. ML is classified as supervised, unsupervised, semi-supervised, and reinforcement learning, as shown in Figure 2. In supervised learning, models are trained on a particular set of data/samples, where each sample has a known outcome.⁵ In other words, this model trains from input to output, producing an output for new or unseen data. In autoimmune disease, this model is used for disease classification and risk prediction. Also, for predicting diagnoses based on genetic or laboratory data, classifiers, primarily support vector machines (SVMs), gradient boosting machines, or random forests (RFs), are commonly used.⁵

In unsupervised learning, the algorithm is designed to precisely identify patterns or produce outputs without predefined training or label datasets. This model can be used to identify molecular subgroups using previously unexplored genetic data.⁶ This model is primarily used for exploratory analysis and the identification of novel biomarkers and is not commonly used to predict disease outcomes or inform treatment decisions.⁶ Reinforcement learning primarily learns from environmental factors. In other words, this model, upon action, is either rewarded or penalized and, over time, is trained, leading to the production of the best or most precise outcome.^5,6 In autoimmune disease, its use is nascent, but over time, it may be better optimized and contribute to the field.

Recently, there has been growing interest in hybrid approaches that combine DL with ensemble methods and structured clinical variables. Deep neural networks such as convolutional neural networks (CNNs) remain central for imaging-heavy tasks (radiology, histology, MRI, endoscopy), while tree-based ensembles (e.g., gradient boosting, XGBoost) and RFs are often competitive for tabular clinical, laboratory, and multi-omic data. Natural language processing can be applied to EHR narratives to extract phenotypes, flares, and treatment response signals. The choice of model should be driven by data modality, sample size, interpretability requirements, and validation strategy rather than by a single “best” algorithm.

A summary of the reported predictive performance, explainability, and readiness of the included ML models to the real world is presented in Table 3. Most of the models are in an experimental or pilot phase, although internal performance measures are high because of the lack of external validation, calibration studies, and future clinical testing.

SLE is a systemic autoimmune disease where healthy cells are attacked by the hyperactivated immune system, and the prevalence of SLE is 1.5–11 in 100,000 patients per year, with higher prevalence in Black, Asian, and women. ML has been used in the diagnosis and management of therapy. In a systematic review and meta-analysis, Zhou et al.⁹ used ML, and the sensitivity and specificity were 0.90 and 0.89, respectively. Li et al.¹⁰ used the 2021 ML (multimodal-based algorithm) and analyzed six proteins to diagnose SLE. Adamichou et al.¹¹ used RFs on 802 subjects with SLE and rheumatological disease. The outcome showed an accuracy of 94% and a sensitivity of 93.8% against the SLE risk probability index. The LASSO-LR model was used by Han et al.¹² to predict progression from SLE to SLE-SS. A SVM model was used by Tan et al.¹³ to diagnose NPSLE, and the results showed accuracy, sensitivities, and specificities of 94.9%, 91.3%, and near-100% (single dataset; requires external validation), respectively. Choi et al.¹⁴ used the longitudinal clustering technique on 805 patients, and the findings showed that SLE patients have a 10% higher risk of CVD. Hence, ML could be a promising tool for early SLE detection, risk prediction, and treatment decisions.

Rheumatoid arthritis is one of the common systemic arthritides typically associated with chronic inflammation, leading to extraarticular organs and joints. Despite advances in the diagnosis and management of RA, there is still a need for greater specificity and accuracy. A CNN model was developed by Ahalya et al.¹⁵ for RA identification using X-ray images, with sensitivities and specificities of 95% and 94%, respectively. Lim et al. (2023) used ML to detect single-nucleotide polymorphisms in the training data of RA patients, and in another study, Guo et al.¹⁶ used an ML algorithm, network analysis, and RFs to explore biomarkers for RA from the GEO database. The results identified POLE4, AKR1C3, and MCEE as potential biomarkers. Mehta et al.¹⁷ used a RF algorithm to differentiate osteoarthritis from RA using H&E-stained synovial tissue sections. The outcome showed 82% accuracy (micro-AUC 0.87±0.04). In one study, 240 thermal images of the hands of RA and healthy individuals were used, and k-means clustering was employed for hot spot segmentation. Among three classifier models, the LogitBoost classifier showed the best accuracy (93.75%), followed by the quantum SVM (92.7%).¹⁸ Gossec et al. reported flares in patients with RA and axial spondyloarthritis in the ActConnect study, with average sensitivities and specificities of 96% and 97%, respectively.

IBD is among the most common gastrointestinal autoimmune diseases, characterized by inflammation, and affects more than 7 million patients worldwide. Diagnosis and management of IBD are critical, as they involve an integrated approach that includes imaging, laboratory findings, and histological analysis. Additionally, its management is challenging due to its complex etiology. A study by Takenaka et al.²¹ using CNNs to assess endoscopic and histological images reported 90% and 93% accuracy for histological and endoscopic data, respectively. Quénéhervé et al.²² used AI-based laser endomicroscopy for mucosal healing, and sensitivity and specificity were near-100% (single dataset; requires external validation). When natural language processing (NLP) was used in EHRs to predict disease flare and treatment response, the results were promising. However, AI/ML has not yet been thoroughly studied in IBD, but evidence to date suggests immense potential for diagnosis, differential diagnosis, and management.

T1DM is an autoimmune disease that causes β-cell destruction and significantly affects patients’ quality of life. In recent times, ML has been used for diagnosis, management, and monitoring of disease using genetic, metabolic, or immunological data. Cheheltani et al.²³ used ML to differentiate T2TD and T1DM. Oviedo et al.²⁴ used ML to predict hypoglycemia, and the outcome showed a 37% reduction in hypoglycemia. Cederblad et al.²⁵ used ML to confirm the causes of hypoglycemia, and the findings showed a profound role of basal insulin pressure and bolus, at 44% and 27%, respectively. Indeed, ML has demonstrated a promising impact in the diagnosis and management of T1DM. However, the lack of prospective validation limits its widespread application. In the future, integrating genomics, proteomics, metabolomics, and EHR may improve early diagnosis and management of T1DM.

MS is a chronic autoimmune disease of the CNS and is primarily characterized by neurodegeneration and demyelination. The clinical manifestation of autoimmune disease causes disabilities and considerable socioeconomic pressure. MRI is primarily used for diagnosis, and accurate diagnoses are often missed. AI/ML plays a critical role in the diagnosis, management, and monitoring of MS. ML models, such as SVM, RF, and neural nets, are used, along with brain MRI data, to diagnose or score the disease (EDSS score).²⁶ In one published study, the use of ML for diagnosing MS achieved a sensitivity and specificity of 60%–80%.²⁷ Moreover, when imaging data, laboratory reports, genomics, and MRI details, such as lesion location or volume, are integrated with ML, accuracy is significantly improved.

Apart from the above-discussed autoimmune disease and the role of ML, there are other autoimmune diseases, where the role of ML is critical. Graves’ disease and Hashimoto’s thyroiditis are among such diseases. In one study, ML was used to differentiate Graves’ ophthalmopathy from healthy individuals. SVM, k-nearest neighbor, and generalized regression neural network were used in the analysis, and sensitivity and specificity of more than 90% were achieved. In another study, ML was used to predict prognosis and identify responders to first-line glucocorticoid therapy among patients with Graves’ ophthalmopathy. The result of an AI application project might personalize treatment options for patients with moderate-to-severe, active Graves’ ophthalmopathy. For the differential diagnosis of thyrotoxicosis, ML algorithms were used in combination with EHR data, yielding more reliable reports of RF. Celiac disease is another autoimmune disease that primarily affects the small intestine and is caused by sensitivity to gluten in the diet. The prevalence is 1 in 100, and approximately 22% of first-degree relatives have an increased risk. Notably, undiagnosed cases may range from 80% to more than 80%. In one of the early studies, capsule endoscopy was used to train a deep-CNN (GoogLeNet) and to evaluate pathological involvement. The finding showed near-100% (single dataset; requires external validation) accuracy in diagnosing celiac disease. To correlate pathology and genetic expression in primary biliary cholangitis, a cDNA microarray was used.