FaceFuel: A Multi-Stage Heterogeneous Fusion Pipeline for Non-Invasive Nutritional Deficiency Screening from Facial Imagery

We present FaceFuel, a five-stage deep learning pipeline that estimates nutritional and lifestyle deficiency probabilities from a single selfie image without blood tests or clinical contact. The pipeline combines MediaPipe face alignment, YOLOv8m lesion detection (mAP@0.5 = 0.790) across 11 skin feature classes, region-aware DINOv2 ViT-S/14 feature extraction from eight anatomical facial zones, a per-feature severity MLP with Monte Carlo Dropout uncertainty quantification, and a calibrated Bayesian inference engine. The system achieves mean F1 = 0.6765 at 17.1 FPS (58.3 ms/image) on an NVIDIA RTX 4070 Super. Ablation experiments demonstrate that region-aware DINOv2 features improve F1 by 0.149 over whole-face features, and the full pipeline outperforms a rule-based baseline by 0.131 F1.