Model card & safety

Intended use

DermaMarker is clinical decision support for general practitioners and nurses in settings without a dermatologist on call. It prioritises referrals and drafts documentation. It is not a diagnostic device — every output is framed as a preliminary assessment for a licensed clinician to confirm or overrule.

Model & training data

Known limitations

• Skin-tone representation. HAM10000 is drawn overwhelmingly from lighter Fitzpatrick skin types (I–III). Accuracy on darker skin tones — common across the Gulf, South Asia, and Africa — is not established. The confidence gate below is a partial mitigation; fine-tuning on regional data (e.g. dermoscopy sets that include Fitzpatrick IV–VI) is the roadmap fix.
• Dermatoscopy vs phone photos. Training images are dermatoscopic. Smartphone photos degrade accuracy; the UI instructs users on framing and flags low-confidence results instead of hiding the uncertainty.
• Seven classes only. Conditions outside the HAM10000 taxonomy (e.g. cutaneous infections, drug eruptions) will be forced into the nearest class. Out-of-distribution detection is on the roadmap.
• Class imbalance. 67% of training images are benign nevi. Training uses weighted sampling to counter this, but rare classes (dermatofibroma, vascular lesions) have thin support.

Safety guardrails

• Asymmetric triage thresholds. Melanoma at just 30% probability or basal cell carcinoma at 40% already escalates to urgent — the system is deliberately tuned to over-refer rather than under-refer. A missed melanoma costs a life; an unnecessary referral costs a consultation.
• Confidence gating. Any top prediction below 50% confidence is automatically bumped to monitor with a re-scan recommendation, and the uncertainty is stated in the report rather than suppressed.
• Degradation, not failure. Each pipeline step (classify → triage → report → store) fails independently and visibly. If the AI narrative layer is unavailable, a deterministic rule-based report is generated instead — the clinician always gets the raw probabilities.
• Human in the loop by design. The product output is a referral recommendation addressed to a clinician, never a diagnosis addressed to a patient.

Privacy & data handling

• No patient identifiers are collected — scans store the image, the model output, and timestamps only.
• Images are stored in Supabase object storage; deployments can be region-pinned to satisfy data-residency requirements (e.g. UAE health data regulations).
• The referral note is generated on demand and printed or copied by the clinician — it is not transmitted anywhere by the system.