Type 2 diabetes mellitus (T2DM) is a chronic metabolic disorder that substantially affects oral health, increasing the incidence of periodontal infections and contributing to dental treatment failures through mechanisms such as chronic hyperglycemia. While laser‑induced breakdown spectroscopy (LIBS) coupled with machine learning (ML) has previously identified diabetes‑related changes in hair, nails, and urine, mineralized tooth tissues have not been extensively studied as a long‑term metabolic archive. Teeth are chemically stable tissues that can preserve systemic elemental alterations over extended periods, providing a potential substrate for non‑destructive diagnostic screening.
The study collected dental tissues from 30 individuals, split evenly between 15 healthy controls and 15 people with T2DM. Four distinct dental tissues were sampled: enamel, coronal dentine, radicular dentine, and cementum. Laser‑induced breakdown spectroscopy was used to acquire spectral data, yielding a total of 3600 LIBS spectra across the sampled tissues. The abstract does not report additional procedural details such as instrument model, laser parameters, spatial sampling strategy, or preprocessing workflows; those specifics were not provided in the source abstract.
Quantitative comparison of LIBS elemental signals between diabetic and healthy teeth identified several elements with differential abundance. Diabetic teeth showed higher levels of iron (Fe) and tin (Sn). Conversely, zinc (Zn), silicon (Si), and potassium (K) were observed at lower levels in the diabetic group. The authors identify these elements as diabetes‑related elemental biomarkers preserved in dental tissues. The abstract does not give absolute concentrations, statistical test details, or effect sizes beyond directional differences.
Multiple supervised classification algorithms were evaluated to distinguish diabetic from non‑diabetic tooth spectral profiles. Algorithms tested included logistic regression (LR), support vector machine (SVM), and artificial neural network (ANN). Two feature‑selection or dimensionality‑reduction strategies were compared: principal component analysis (PCA) and a correlation‑based forward/backward selection (CFS‑BFS). The models were trained and evaluated using the LIBS spectral features reduced or selected by these methods. Specifics on model architectures, hyperparameter values, cross‑validation schemes, and training/test splits are not reported in the abstract.
Performance varied across algorithm and feature‑selection combinations. The PCA‑ANN combination achieved the highest reported mean metrics with an accuracy of 96%, sensitivity of 94%, and specificity of 96%. PCA‑SVM produced a mean accuracy of 95% with a notably narrow confidence interval reported as 94.98–95.02%, which the authors highlight as indicating high stability of accuracy for that model. The study notes that no single model was consistently the most stable across all evaluated performance metrics, indicating tradeoffs between peak performance and reproducibility depending on the chosen classifier and metric.
Principal component analysis was an effective dimensionality‑reduction approach in this dataset. Using six principal components captured 94% of the variance in the LIBS spectra and PCA‑based pipelines consistently outperformed the CFS‑BFS feature‑selection approach when integrated with the tested ML algorithms. The abstract does not report the loading patterns or the spectral regions driving the principal components.
The results demonstrate that dental tissues retain diabetes‑associated elemental alterations that are detectable by LIBS combined with machine learning, enabling a potential non‑destructive screening method. In a dental setting this approach could provide pre‑treatment risk identification for undiagnosed diabetes, with the potential to reduce dental treatment failures and improve outcomes. However, the authors emphasize that larger, multicenter validation studies are required before this method can be adopted in therapeutic practice. The abstract omits several methodological details needed for replication and clinical translation, including full instrumentation parameters, detailed model training methodology, and external validation results.
This work supports the feasibility of using LIBS‑derived elemental fingerprints from mineralized dental tissues, combined with machine learning classifiers, to classify type 2 diabetes status non‑destructively. Key diabetes‑related elemental markers identified were increased Fe and Sn and decreased Zn, Si, and K in diabetic teeth. PCA with six components was an effective feature‑reduction strategy, and the PCA‑ANN and PCA‑SVM pipelines produced the strongest classification performance and stability metrics reported in the abstract. The authors call for large‑scale multicenter validation to confirm these findings and to enable translation into routine dental screening practice. Further reporting of experimental parameters and external validation results will be necessary for implementation and regulatory consideration.