Rare autoimmune diseases often present diagnostic challenges, leading to significant delays that negatively impact patient prognosis and treatment efficacy. While advanced AI models are highly adept at identifying intricate patterns in complex multi-omics datasets, their 'black-box' nature impedes clinical trust and adoption. Introducing explainable AI (XAI) techniques into clinical decision support (CDS) systems is paramount not only for enhancing diagnostic accuracy but also for providing clinicians with transparent, interpretable reasoning, thus fostering confidence and facilitating meaningful human-AI collaboration. This work addresses a critical gap in clinical translational AI.
The aim of the study.
To develop and validate an explainable artificial intelligence (XAI) framework for the diagnosis of rare autoimmune diseases using integrated multi-omics data, and to assess clinician perception of the interpretability and utility of the generated explanations for clinical decision support.
Material and methods.
This retrospective study included integrated multi-omics data (Whole Exome Sequencing, RNA Sequencing, Mass Spectrometry-based Proteomics) from 500 patients diagnosed with one of three challenging rare autoimmune diseases (200 Systemic Lupus Erythematosus, 150 Sjogren's Syndrome, 150 Systemic Scleroderma) and 500 age/sex-matched healthy controls. Data was collected from a biobank specializing in autoimmune disorders. A Multimodal Transformer architecture was developed to process and integrate different omics data types, trained on 70% of the dataset and validated on the remaining 30%. Post-hoc XAI techniques, including SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), were applied to generate patient-specific explanations. A questionnaire-based survey involving 25 rheumatologists and clinical geneticists assessed the interpretability of explanations (comprehensibility, clinical relevance) and their utility in clinical decision-making on a 5-point Likert scale (1=Not helpful to 5=Very helpful).
The Multimodal Transformer model achieved a high overall diagnostic accuracy of 92.5% (95% CI: 90.1-94.6) across the three rare autoimmune diseases. The macro-averaged F1-score was 0.93. For individual diseases, accuracy was 93.1% for SLE, 91.8% for Sjogren's, and 92.2% for Scleroderma. The XAI framework successfully generated patient-specific explanations, consistently highlighting genomic variants (e.g., in BLK, STAT4 for SLE), differentially expressed genes (e.g., IFIT1, MX1 signatures for Sjogren's), and dysregulated protein pathways (e.g., complement pathway for SLE, collagen synthesis for Scleroderma) that contributed most significantly to each AI prediction. In the clinician survey, 85% (21/25) of participants rated the explanations as 'helpful' or 'very helpful' for understanding the AI's reasoning, and 76% (19/25) reported that the explanations would increase their trust in AI-driven diagnostic recommendations. Specifically, targetable pathways and genes were frequently cited as being highly useful.
This study demonstrates the successful development and validation of an explainable AI framework that not only accurately diagnoses rare autoimmune diseases from complex multi-omics data but also provides transparent and clinically interpretable justifications for its predictions. The strong clinician endorsement of the XAI explanations underscores their potential to bridge the gap between advanced AI capabilities and their practical integration into clinical decision support systems. By offering high diagnostic accuracy coupled with transparency, this framework promises to accelerate diagnosis, enhance clinician confidence, and ultimately improve patient outcomes in autoimmune medicine. Future work will involve prospective clinical validation and integration into a real-time diagnostic pipeline.
1. Lundberg SM, Lee SI. A Unified Approach to Interpreting Model Predictions. Adv Neural Inf Process Syst. 2017;30:4765-4774.
2. Ribeiro MT, et al. 'Why Should I Trust You?': Explaining the Predictions of Any Classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016:1135-1144.
3. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44-56.
4. Aibar J, et al. Multi-omics data integration for precision medicine: an overview of computational methods. Brief Bioinform. 2021;22(2):bbab023.
5. Fan J, et al. Artificial intelligence in rheumatology: a systematic review. Autoimmun Rev. 2021;20(9):102872.