Predictive Model for Bone Availability in Patients With Unilateral Cleft Lip and Palate: A Machine Learning Approach

Objective To develop a predictive model for estimating cortical bone thickness at any maxillary location in patients with unilateral cleft lip and palate (UCLP). Design Retrospective cross-sectional cohort study with machine learning. Setting University hospital, Department of Oral and Maxillofacial Radiology. Patients, Participants Fifty patients with non-syndromic UCLP and 50 age- and gender-matched controls (total N = 100). Mean age: UCLP 14.8 ± 5.1 years; controls 16.5 ± 4.6 years. Age range: 8 to 27 years. Interventions All participants underwent CBCT (0.2 mm voxel, 14 × 16 cm FOV). Cortical bone thickness was measured at 4, 6, 8, 10, and 12 mm from the alveolar crest across six interradicular regions, yielding 4500 data points. Main Outcome Measure(s) Cortical bone thickness; model performance metrics ( R 2 , RMSE); feature importance; accuracy by region and depth. Results Random Forest regression explained 75% of variance (patient-level cross-validation: R 2 = 0.75, RMSE = 0.102 mm). Anatomical region was the strongest predictor (0.38), followed by depth (0.27), age (0.18), cleft status (0.12), and side (0.05). Best accuracy in posterior regions ( R 2 = 0.84) and at 12 mm depth ( R 2 = 0.81); lowest in anterior regions ( R 2 = 0.72) and at 4 mm depth ( R 2 = 0.69). The model identified high-risk zones with 89% sensitivity (AUC = 0.94). External validation has not yet been performed; the author welcomes collaboration with other centers for independent testing. Conclusions This model provides accurate, patient-specific predictions of cortical bone thickness in UCLP patients, assisting in pretreatment risk assessment, surgical planning, and patient counseling.