Longitudinal visual tracking of nail lesions using a deep learning-based EMR-integrated platform

This study proposes and evaluates an EMR-integrated, artificial intelligence (AI)-based platform for the image-level classification and longitudinal visual tracking of nail lesions in real-world clinical workflows. Because the morphology and extent of nail lesions often change over time, continuous visual monitoring is crucial to assess disease progression and treatment responses. However, conventional electronic medical record (EMR) systems provide limited functionality for systematically managing and tracking these temporal changes. To address this, we developed a two-stage AI pipeline seamlessly integrated into a clinical platform. First, an RF-DETR-based detection model automatically identifies and crops individual nails from full-hand images. Second, a MedSigLIP-based classification model categorizes the cropped images into four target classes: normal nail, melanonychia, nail psoriasis, and onychomycosis. The classifier was trained via partial fine-tuning and rigorously compared against standard baseline models (ResNet50, ViT, CLIP, and SigLIP). Model performance was evaluated using a bootstrapping method with 1,000 resamples to derive 95% confidence intervals (CIs). Finally, the technical feasibility and clinical usability of the platform were assessed through a structured user experience (UX) survey by eight board-certified dermatologists. The RF-DETR model successfully isolated individual nails, facilitating automated downstream analysis. The MedSigLIP classification model consistently outperformed the standard baseline architectures, demonstrating robust reliability with an F1 score of 0.928 and an Area Under the Curve (AUC) of 0.988. Furthermore, Grad-CAM + + was integrated to provide interpretable visual explanations for the AI predictions. In the UX survey, the dermatologists reported high overall satisfaction, particularly valuing the platform’s utility for longitudinal visual tracking (mean score 3.75 out of 4.00). The proposed system demonstrates the technical feasibility of longitudinal visual tracking for lesion progression, serving as a platform for automated, per-nail analysis. By successfully bridging high-performing AI models with practical clinical workflows, the platform enables the storage of long-term visit records and facilitates their time-series comparisons. Future work will focus on establishing standardized imaging protocols, expanding disease categories, and conducting multi-center clinical validation to further enhance its real-world applicability.