Artificial Intelligence-based detection of neuropsychiatric lupus: an exploratory meta-analysis of neuroimaging and multimodal biomarker models.

Comprehensive Summary

Nouroozi et al. conducted a PRISMA-compliant systematic review and exploratory meta-analysis examining how well AI-based models using neuroimaging and multimodal biomarkers can detect neuropsychiatric involvement in systemic lupus erythematosus (NPSLE). The authors searched PubMed, Scopus, and Web of Science through August 2025, ultimately including 14 studies with a combined total of more than 800 participants, and applied random-effects meta-analytic models to pool reported AUC, accuracy, sensitivity, and specificity across those studies. Pooled diagnostic performance was generally high, AUC 0.86 (95% CI: 0.84–0.87), accuracy 0.87 (95% CI: 0.86–0.89), sensitivity 0.87 (95% CI: 0.86–0.88), and specificity 0.82, though between-study heterogeneity was substantial (I² ≥ 94% across metrics). Subgroup analysis found no statistically significant difference in accuracy between classical machine learning and deep learning models (Kruskal-Wallis p = 0.55), and a leave-one-out sensitivity analysis revealed the pooled AUC was heavily influenced by a single study (Simos et al.), with exclusion of that study shifting the pooled AUC from 0.86 to 0.96, and exclusion of any other single study dropping it to approximately 0.75. The authors conclude that while AI-based neuroimaging in NPSLE shows promise, marked heterogeneity, absence of external validation, and a near-complete lack of formal explainable AI (XAI) methods across all 14 studies preclude firm conclusions about clinical readiness.

Outcomes and Implications

NPSLE is notoriously difficult to diagnose; standard imaging and blood tests often come back normal even when the brain is affected, making AI-assisted detection a genuinely exciting prospect for rheumatologists and neurologists. However, the promising accuracy numbers reported here should be interpreted with caution: the studies underlying them were small, tested mostly on the same patients used to train the models, and none explained which brain features actually drove their predictions, a major barrier to clinician trust. The authors are clear that these tools are not ready for the clinic yet, and that larger, multi-site studies with standardized methods and transparent AI pipelines will need to come first before patients can realistically benefit.