Analytical validation of Smilo.ai: evaluating an AI-driven platform for automated oral health screening and global...

Oral diseases affect more than 3.5 billion people worldwide and remain among the most neglected global health challenges. Access to preventive oral healthcare is constrained by workforce shortages, transportation barriers, and patient apprehension toward dental procedures. Artificial intelligence (AI)-enabled diagnostic systems offer scalable solutions for early detection and triage through non-dental personnel. This study presents the Phase 1 analytical validation of Smilo.ai, an AI-driven platform designed for automated detection of common oral health conditions using standardized digital oral photographs.

A cross-sectional validation study was conducted among 45 adult participants in a community-based oral health screening program in India between February 1 and 10, 2025. Each participant contributed one composite oral image containing four standardized views (frontal, left lateral, right lateral, and occlusal), yielding an analytical dataset of 45 images (n = 45). Smilo.ai's diagnostic outputs were compared with dentist evaluations of identical photographs across seven diagnostic categories: dental decay, gingivitis, plaque, calculus, tooth wear, discoloration, and crowding. Performance metrics included sensitivity, specificity, precision, accuracy, and Cohen's κ, computed using confusion matrix-based analysis in accordance with STARD 2015 and CONSORT-AI guidelines. Inter-rater reliability between the two reference dentists was retrospectively quantified, yielding substantial agreement (κ = 0.61).

Smilo.ai demonstrated an overall sensitivity of 73.6%, precision of 46.0%, accuracy of 47.3%, and a macro-level Cohen's κ of 0.010 relative to dentist assessments. High sensitivity was achieved for dental decay, tooth wear, and discoloration (100% each); however, the corresponding near-zero specificity values indicate a pattern of positive prediction bias in the current model configuration. Calculus exhibited the strongest balanced performance (sensitivity 61.0%, specificity 65.0%, precision 73.9%, κ = 0.24). Gingivitis demonstrated moderate sensitivity (56.0%), while plaque and crowding showed partial detection with variable agreement.

This Phase 1 validation establishes a performance baseline for Smilo.ai and identifies optimization priorities, including positive prediction bias correction and threshold recalibration necessary before screening workflow integration. AI-assisted screening models hold significant promise for improving preventive oral healthcare access among underserved populations. The upcoming Phase 2 validation will evaluate clinical generalizability through synchronous clinical examinations to confirm real-world applicability.