Beyond Prevalence: An Exploratory Framework for Patient Phenotyping and Clinical Burden Analysis of Incidental...

To move beyond simple prevalence reporting by applying an exploratory, transferable multi-method analytical framework to incidental findings on dental cone-beam computed tomography (CBCT). The framework combines multiple correspondence analysis (MCA), data-driven patient clustering, association rule mining, co-occurrence network analysis, and clinical burden quantification, and is intended to generate hypotheses for prospective validation.

In this retrospective, cross-sectional, single-center study, 100 consecutive CBCT scans acquired for implant treatment planning were interpreted by a single board-certified oral and maxillofacial radiologist using a structured reporting template. Findings were classified into three investigator-defined clinical-action categories (I: intervention; II: monitoring/correlation; III: no action) and by anatomical domain. MCA (prevalence ≥5%) was followed by agglomerative clustering in MCA space; association rules were mined with the Apriori algorithm; a phi-coefficient co-occurrence network was built with Louvain community detection; and a weighted Clinical Burden Index (CBI) was computed per patient. Finding counts were modeled by Poisson regression, with sensitivity analyses for the prevalence threshold and CBI weights. This study is exploratory and hypothesis-generating.

Across 100 patients, 488 incidental findings (86 unique types) were identified (mean 4.9 ± 2.0; range 1-10), with Category I findings in 86% of patients. MCA yielded two dimensions explaining 43.9% of inertia. Clustering identified two exploratory phenotypes, "High-Burden Multimorbidity" (n = 23, mean 6.1 findings) and "Standard-Burden" (n = 77, mean 4.5; Kruskal-Wallis p = 0.001). Five association rules were found (top lift = 1.90, stable across thresholds), and a sparse, modular co-occurrence network bridged dental and nondental domains. Poisson regression identified Category I (incidence rate ratio [IRR] = 1.21, 95% confidence interval [CI]: 1.14-1.28) and Category III (IRR = 1.25, 95% CI: 1.16-1.34) counts as independent predictors of burden. Mean CBI was 10.3 ± 4.8; 18% of patients were high-burden, with tier assignment robust to weighting (Spearman ρ ≥ 0.90). An estimated 86% of patients would potentially require at least one specialist referral based on incidental findings alone.

Applied to a retrospective single-center dataset, this exploratory analytical framework reveals patient phenotypes, clinically relevant co-occurrence patterns, and a potentially substantial specialist-referral burden that univariate prevalence reporting does not capture. These preliminary observations require prospective, multi-center external validation, and the framework may become a valuable tool for structured, data-driven CBCT interpretation.