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

Abstract

Neuropsychiatric systemic lupus erythematosus (NPSLE) remains challenging to diagnose because of heterogeneous clinical presentations, nonspecific findings, and the absence of definitive biomarkers. Artificial intelligence (AI) methods have been increasingly explored using neuroimaging and other biologically informative data to support identification of neuropsychiatric involvement in systemic lupus erythematosus (SLE). However, the reported performance and methodological robustness of these approaches have not been systematically characterized. To perform an exploratory meta-analysis describing reported diagnostic performance, heterogeneity, and methodological characteristics of AI-based models using neuroimaging and multimodal biomarkers for detecting neuropsychiatric involvement in SLE. We conducted a PRISMA-compliant systematic review of studies applying machine learning or deep learning models to neuroimaging or biologically informative modalities relevant to central nervous system involvement, including structural or functional MRI, magnetic resonance spectroscopy, spectroscopy-based molecular fingerprints, and CSF or serum biomarkers. PubMed, Scopus, and Web of Science were searched through August 2025. Given substantial heterogeneity in study design, model objectives, input modalities, and validation strategies, analyses were undertaken within an exploratory framework. Random-effects models were used to summarize reported area under the curve (AUC), accuracy, sensitivity, and specificity. Subgroup and leave-one-out sensitivity analyses were performed. Fourteen studies involving more than 800 participants were included. Most studies used neuroimaging, particularly resting-state functional MRI, while others incorporated non-imaging biomarkers. Reported performance metrics were generally high (pooled AUC 0.86; accuracy 0.87), but between-study heterogeneity was substantial. Sensitivity analyses demonstrated that pooled estimates were unstable and influenced by individual studies. No clear performance differences were observed between classical machine learning and deep learning approaches. External validation and formal explainable AI methods were uncommon. This exploratory synthesis indicates that AI-based models applied to neuroimaging and multimodal biomarkers have shown promising reported performance in NPSLE. However, marked heterogeneity, limited robustness, and poor interpretability currently preclude firm conclusions regarding clinical applicability. More standardized, externally validated, and interpretable studies are needed before translation into clinical practice.