MEL-IA AI achieves 86% accuracy across five skin lesion classes in hospital-integrated deployment

The MEL-IA platform combines a mobile image-capture app, a multimodal classifier, and secure middleware that writes results directly into the hospital electronic record. In live testing it processed 980 studies with sub-second latency while maintaining longitudinal lesion tracking. Unlike binary malignant-benign tools, the five-class output supplies differential probabilities intended only as decision support for clinicians.

Training incorporated dermatoscopic images and metadata (age, sex, site) to reach 88% melanoma sensitivity and 92% basal-cell-carcinoma sensitivity. These figures derive from retrospective validation; no prospective diagnostic RCT has yet been reported. Integration into existing IT infrastructure reduces workflow friction but also creates new attack surfaces for data exfiltration or model poisoning if security controls lag.

Over-reliance on any single AI output risks missed melanomas when image quality falls or when rare lesion subtypes appear. Comparable systems have shown performance drops of 10-15 percentage points on external validation sets. The authors correctly flag the need for prospective studies with dermatologists before wider rollout.

Next steps include smartphone acquisition trials and expansion to additional lesion categories. Until those data exist, hospitals adopting MEL-IA should retain mandatory clinician override and audit all discordant cases.