Answers are grounded in this paper and cite their sources.
Creatinine, estimated glomerular filtration rate and histology detect kidney disease late and say little about the molecular heterogeneity driving progression. Artificial intelligence can extract molecular features from biofluid omics; digitised biopsies and retinal images supply proxy readouts, not molecular measurements. In this review we grouped the literature by histopathology, multi-omics, foundation models and noninvasive systems and checked each result against its cohort, endpoint and validation setting. Reported performance is often high, but most results come from retrospective single-cohort evaluations; external and prospective validation are uncommon, and we found no study among those reviewed showing improved patient outcomes. High discrimination therefore establishes technical performance only; clinical utility and patient benefit remain untested. The evidence to date supports feasibility, not benefit, and the weakness lies in evaluation design rather than model architecture. The evidence so far establishes feasibility, not benefit. Until prospective studies on shared external cohorts are available, these tools are best used to augment existing diagnostic practice, with patient outcomes as the measure of success.
No takes yet. Share an insight, caveat, or question.
Zeng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: