JMIR Publications

ABSTRACT

Background:

Early diagnosis of systemic lupus erythematosus is crucial for improving patient outcomes. While artificial intelligence, particularly deep learning, is being increasingly applied in the medical field, its role in diagnosing lupus has not been explored, and there is limited systematic evidence supporting its diagnostic value.

Objective:

This study aimed to evaluate the accuracy of machine learning models, including deep learning, in detecting lupus-related manifestations using histopathology, radiological, and ultrasound images.

Methods:

PubMed, Embase, Cochrane Library, IEEE Xplore, and Web of Science were comprehensively searched for studies published since 2014. Studies were assessed for bias risk using the QUADAS-AI tool to ensure quality evaluation. 26 studies providing data for constructing contingency tables were included to assess the diagnostic performance of machine learning models for lupus.

Results:

From 6,611 studies screened, 26 were included. Pooled results demonstrated strong diagnostic accuracy of machine learning models (sensitivity 90% [89–92%], specificity 94% [92–95%], AUC 0.97). Subgroup analyses spanned five categories: algorithm (ML vs. deep learning), task types (detection vs. SLE classification), disease types (lupus nephritis, neuropsychiatric lupus), feature types (MRI, Raman spectroscopy), and DL-versus-clinician performance comparisons.

Conclusions:

Pooled estimates indicate high diagnostic accuracy across modalities; however, very high between-study heterogeneity and the predominance of retrospective designs warrant caution. Given the limited external validation and sparse head-to-head comparisons with clinicians, current evidence, though promising, is preliminary. Prospective, multi-center validation is required before routine clinical adoption. Clinical Trial: PROSPERO CRD42024545109.