Insulin resistance is a critical yet underdiagnosed driver of metabolic disease, preceding type 2 diabetes by years. The Homeostasis Model Assessment for Insulin Resistance (HOMA‑IR) quantifies fasting glucose‑insulin dynamics, with a score >2.9 indicating resistance.

While total body fat percentage gives a baseline, biomarkers such as the Android‑to‑Gynoid (A/G) ratio and Visceral‑to‑Subcutaneous (V/S) ratio reveal deeper risk. Dual‑Energy X‑Ray Absorptiometry (DXA) remains the gold standard for body composition but is costly, requires specialized equipment, and involves low‑dose radiation, limiting everyday screening.

Photo Scan, an investigational deep‑learning framework, estimates BF%, A/G ratio and V/S ratio from standard 2D smartphone photos. It was pre‑trained on over 35,323 UK Biobank records and fine‑tuned with a diverse cohort of 677 adults. Independent validation on the Metabolic Mosaic cohort (N=132) showed mean absolute errors of ~2.1 % for BF%, 0.09 for A/G ratio and 0.09 for V/S ratio, matching DXA accuracy.

For insulin resistance classification, a gradient‑boosting classifier using baseline demographics achieved AUROC 0.692. Adding Photo Scan features raised AUROC to 0.760 and NRI to 0.593, approaching the performance of DXA (AUROC 0.773, NRI 0.748). Smartwatch BIA added no improvement, underscoring the value of A/G and V/S ratios.