Year: 2026
Type: preprint
Study design: Observational Study
Abstract: Background: Artificial intelligence (AI) is increasingly utilized to support the early assessment of skin lesions. While numerous studies report high diagnostic accuracy under controlled experimental conditions, considerably fewer integrate technical validation, independent clinical stress-testing, and large-scale real-world evidence (RWE) within a single transparent framework. Objective: To evaluate the capabilities, clinical utility, and operational limitations of a mobile AI-assisted skin analysis system by synthesizing three complementary streams of evidence. Methods: This study integrates: (1) an analysis of longitudinal technical performance metrics and big-data usage patterns from over 6.57 million real-world smartphone assessments; (2) an independent clinical evaluation ((n = 160) analyzed cases) conducted at a specialized tertiary referral dermatology center comparing AI predictions and triage routing against histopathological, cytological, or expert clinical reference standards; and (3) a comparative synthesis with contemporary peer-reviewed literature, including external validation on skin of color. Results: Longitudinal technical validation demonstrated consistent, balanced improvements across sensitivity (97.4%), specificity (93.1%), and overall accuracy (94.2%) by 2026. In real-world consumer usage, common inflammatory conditions (e.g., acne/rosacea, (n = 708,356)) accounted for the largest share of assessments, while preventable image-acquisition errors (poor lighting: 42.0%; poor focus: 28.0%; motion blur: 18.0%) represented the primary modifiable determinants of failure. In the independent tertiary referral cohort - heavily enriched with malignant and diagnostically challenging lesions - exact Top-1 and Top-3 diagnostic agreements were 54.4% and 70.6%, respectively. However, combining Top-3 agreement, correct broader disease-group assignment, and appropriate risk-based routing yielded a clinically useful triage outcome in 80.0% of cases. Conclusions: AI-assisted skin analysis provides clinically meaningful support for early risk stratification and patient referral when deployed as a clinical decision-support tool rather than an autonomous diagnostic system. Comprehensive evaluation of medical AI requires pairing conventional accuracy metrics with triage-oriented clinical endpoints and transparent RWE reporting.
Source: Research Square
Authors: Viktor Shpudeiko, Kirill Sokolov
Language: English
Subfield: Oncology
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7234394826
