Limitations of linear statistical methods for detecting associations between dental status and systemic patient’s health

Relevance. The DMFT index remains the principal epidemiological measure of caries experience. However, as a summary measure, it does not capture the spatial distribution of lesions within the dentition, including their location, symmetry, or complex patterns across individual teeth and tooth groups. Although numerous meta-analyses based on large cohorts have demonstrated links between oral and systemic health, few studies have used detailed dental variables derived from individual teeth and tooth groups to identify such associations. Conventional statistical methods based on pairwise linear associations may therefore be insufficient to detect nonlinear and contextdependent patterns. Objective: To systematically evaluate the ability of conventional statistical methods to detect clinically meaningful associations between detailed dental variables and systemic health indicators, in comparison with large language models (LLMs). Materials and methods. An original digital medical questionnaire was developed, comprising 133 fields across 9 modules, and implemented on the Amvera Cloud (Yandex) platform. A total of 127 consecutive patients were examined at Vident Company LLC (St. Petersburg). Based on the dental chart (32 teeth × 4 statuses) and medical history data, 121 derived variables were generated across 11 categories. More than 2,420 pairwise tests were performed using Bonferroni correction and Benjamini-Hochberg false discovery rate control (q = 0.05). De-identified data were then uploaded to a large language model to explore nonlinear patterns. Results . Neither the Mann-Whitney U test (1,815 tests) nor point-biserial correlation identified any significant associations. Logistic regression based on 12 dental variables yielded an AUC of 0.43-0.61. Adding dental variables to age did not improve predictive performance (ΔAUC = -0.15 to +0.02). Of 25 prespecified clinical hypotheses, only 2 were confirmed (8%), a proportion comparable to the expected false-positive rate. In contrast, the large language model identified four nonlinear patterns in the same dataset: a threshold effect of smoking (ρpart = 0.228, p = 0.010, despite the absence of a significant linear association, p = 0.653), a context-dependent association between asymmetry in missing teeth, cardiovascular disease, and anticoagulant use, a nonlinear comorbidity threshold at an M/F ratio > 1.0, and clinically interpretable clusters. Conclusion . Conventional linear statistical methods appear insufficient for detecting clinically meaningful associations between specific dental patterns and systemic health indicators. These findings support the development of specialized neural network models capable of processing the dental chart as a spatial structure and identifying multifactorial nonlinear patterns.