TRIDENT: Safety-Gated, On-Device Triage for Paediatric Teledermatology on Pigmented Skin

Sub-Saharan Africa (SSA) has fewer than one dermatologist per million people, and care increasingly flows through WhatsApp, which recompresses every photo before any model sees it. Existing teledermatology classifiers are forced-choice: every image gets a diagnosis regardless of readability or confidence, conflating a blurry photo (needs a retake) with an ambiguous case (needs specialist referral). We present TRIDENT, a 4M-parameter multi-task EfficientNet-B0 deployed on-device (8.1MB FP16 ONNX, 73ms median, Galaxy A34) that routes each image to retake, treat, or refer via heads for image quality, diagnostic uncertainty, and a safety gate for secondary bacterial infection (impetiginization), where steroids on an infected lesion are actively harmful. On PASSION (four SSA countries, Fitzpatrick III–VI), TRIDENT is on par with a published ResNet-50 baseline (subject balanced accuracy 0.703 vs. 0.70, within seed variation) while resolving 57% of encounters offline. We also report two deployment lessons: (1) a dedicated quality signal and simple confidence-based uncertainty are nearly interchangeable for routing cost despite being near-orthogonal (r=0.08), because uncertainty spikes specifically where degradation flips the diagnosis; and (2) retraining that quality signal on the corruptions a retake actually fixes (blur, exposure) makes it track them well (r up to 0.92, from 0.17 under JPEG-only supervision), yet even then it cannot tell which degraded images will be misdiagnosed — uncertainty can — a caution that fixing what a quality head is trained on does not fix what it is structurally unable to see.