Breath-Based Detection of Oral Diseases Using Sensors and Machine Learning.

OBJECTIVES: Diagnostic limitations contribute to variability in clinical decision-making, highlighting the need for objective diagnostic tools. Biomarker-based approaches enable detection of subclinical pathological changes and may improve diagnostic accuracy. Exhaled breath contains volatile organic compounds (VOCs) that reflect host and microbial metabolic activity and therefore represent a promising non-invasive diagnostic medium. This study evaluated the feasibility of a nanosensor array outputs to classify caries, gingivitis, periodontitis, and peri-implantitis using breath VOC profiles. METHODS: A priori power analysis (η²=0.02) yielded 346 participants across five oral phenotypes. Exhaled VOCs were analyzed via GC-MS, with a subset assessed using a 40-sensor nanoarray. VOCs were age-adjusted (residual regression), filtered (≥50% presence; 47 compounds), log-transformed, and tested (Kruskal-Wallis, Bonferroni αadj=0.0167). Nanoarray data underwent Principal Component Analysis (PCA) (15 factors; αadj=0.01), Canonical Discriminant Analysis (CDA), and supervised Machine Learning (ML) (80:20 split). RESULTS: GC-MS identified one significant VOC (RT≈2.81 min) differentiating periodontitis from caries (p=0.006) and gingivitis (p=0.043), while 13 VOCs distinguished implant groups (p<0.05). Nanoarray CDA showed two significant functions (Wilks' Λ=0.1697, p<0.0001; 87.64% variance). A 10-sensor ML model achieved 69.6% training and 58.8% validation accuracy. Validation sensitivities were polarized (100% periodontitis/healthy implants; 25-50% others), with high specificities (0.714-1.000) and ROC-AUC (0.88-0.97). CONCLUSIONS: Although validation performance across all disease categories remained variable, and discriminatory capacity was limited for early dysbiotic and non-inflammatory conditions, the predictive models demonstrated strong classification accuracy for inflammatory phenotypes, particularly periodontitis and implant-related states. These findings highlight both the promise of volatilomic diagnostics and the current limitations imposed by sample size, biological variability, and metabolic overlap, particularly in early disease stages. CLINICAL SIGNIFICANCE: Nevertheless, the strong performance observed in inflammatory conditions underscores the potential clinical utility of this platform as a non-invasive adjunct for disease detection and monitoring. Continued advancements in sensor technologies and machine learning methodologies are likely to enhance model robustness and generalizability, facilitating translation into clinical practice.