Integrated omics and machine-learning strategies for non-invasive endometriosis diagnosis: A critical narrative review

Endometriosis is biologically heterogeneous, and current diagnostic pathways remain imperfect, particularly for superficial disease and for patients with persistent symptoms despite non-diagnostic imaging. This critical narrative review evaluates whether integrated molecular profiling and machine-learning methods can support clinically useful non-invasive testing. We distinguish formal multi-omics integration—joint analysis of at least two molecular layers—from single-omics models, multi-analyte panels, and multi-site profiling. The most mature evidence is a multicenter blood-based model combining microRNAs, proteins, a steroid hormone, and clinical variables; an independent retrospective cohort showed an area under the curve of 0.944, sensitivity of 80.0%, and specificity of 97.5%, although prospective intended-use validation and independent replication remain necessary. Plasma metabolite–autoantibody integration and a serum-metabolite/salivary-microRNA pilot also reported high internal performance, but feature selection, small effective sample sizes, spectrum bias, and absence of locked external validation limit inference. In a newly published serum proteome–metabolome study, the combined model did not outperform a metabolite-only panel for endometriosis versus healthy controls. Stool microbiome–metabolome, menstrual-cell methylome, cell-free-DNA, urinary microRNA, and salivary microRNA studies broaden biological plausibility but remain discovery-stage or single-omics evidence. No current assay has demonstrated sufficient analytical validity, external clinical validity, clinical utility, and interlaboratory reproducibility to replace established pathways. Near-term development should prioritize adjunct triage in symptomatic patients, prespecified compact panels, clinically relevant controls, transparent machine-learning pipelines, multicenter prospective validation, and comparison with expert imaging and usual care. The field has moved beyond proof of concept, but reported discrimination must not be equated with clinical readiness.