Work

Three Data‐Driven Lip Soft‐Tissue Phenotypes and Their Dento‐Skeletal Associations in Orthodontic Patients: An Exploratory Unsupervised Learning Study

Year: 2026

Type: article

Abstract: ABSTRACT Background Conventional facial diagnosis emphasizes skeletal and dental measures while lip features are assessed piecemeal. We aimed to identify data‐driven lip soft‐tissue phenotypes using unsupervised machine learning and examine their associations with skeletal and dental characteristics. Methods In a retrospective cross‐sectional study, 10 lip measurements were extracted from standardized lateral cephalometric radiographs. Variables were z ‐score standardized and analysed using k‐means clustering. Of 26 NbClust validation indices, 14 (53.8%) supported a three‐cluster solution. Cluster stability was assessed by bootstrap resampling (B = 100) and 10‐fold cross‐validation. Post hoc feature importance was evaluated using PCA, Random Forest, SVM and XGBoost. A decision tree was trained and evaluated by 10‐fold cross‐validation repeated five times. Associations with craniofacial parameters were examined via within‐class ANOVA, Cramér's V and hierarchical regression. Concordant pairings were defined as Cluster 1–Class I, Cluster 2–Class II and Cluster 3–Class III; all others were classified as discordant. Results Three phenotypes emerged—Thin‐Straight ( n = 638), Protrusive‐Convex ( n = 644) and Thick‐Concave ( n = 389)—showing broad correspondence with skeletal Classes I, II and III, respectively (Cramér's V = 0.386; overall concordance 52.8%). Lip protrusion and thickness were the primary discriminators (4/4 method consensus for UL_EP_mm, Cm_Sn_Ls, SSD; 3/4 for ULT). Within each skeletal class, all lip variables differed significantly across phenotypes (all p < 0.001; η 2 = 0.10–0.54 for lip variables), and lip phenotype independently predicted incisor inclination beyond skeletal covariates (Δ R 2 adj = 0.058–0.078, p < 0.001). An interpretable decision tree achieved 88.0% cross‐validated accuracy (macro‐F1 = 0.876) for phenotype assignment. Conclusions Unsupervised learning identified three reproducible lip soft‐tissue phenotypes that complement conventional dentoskeletal description by capturing within‐class soft‐tissue heterogeneity. The framework is exploratory and requires prospective external validation and outcome‐based studies before clinical triage application.

Source: Orthodontics and Craniofacial Research

DOI: 10.1111/ocr.70138