TriModal: Face, Tongue, and Eye as Complementary Visual Channels for Non-Invasive Nutritional and Lifestyle Deficiency Screening

This paper introduces TriModal FaceFuel, a tri-modal computer vision system that analyzes a selfie photograph and an optional tongue photograph to produce calibrated probability estimates over 16 nutritional and lifestyle deficiency categories. Building on previously published face and tongue pipelines, a new eye analysis module is developed using YOLO11m trained on 2,497 clinically labeled images across three eye feature classes: conjunctival pallor (iron/B12 deficiency), scleral icterus (liver stress), and xanthelasma (cholesterol imbalance). The eye module achieves mAP@0.5 = 0.913 - the highest of any single modality in the system. A LAB color gate prevents false positive scleral icterus detections by verifying genuine yellow pigmentation (B > 145 in LAB space). A three-way weighted product-of-experts fusion combines face (α=0.40), tongue (β=0.35), and eye (γ=0.25) posteriors into a unified 16-dimensional output. The face detector is upgraded from YOLOv8m to YOLO11m, improving mAP from 0.790 to 0.872 (+10.4%). The complete tri-modal system runs in under 235 ms on consumer GPU hardware and introduces cholesterol imbalance as a new eye-exclusive deficiency dimension detectable from a standard selfie photograph.