Explainable machine learning analysis of factors for predicting hypodontia of permanent teeth among CLP patients: a retrospective study.
Authors
Affiliations (2)
Affiliations (2)
- Oral Department, Children's Hospital of Chongqing Medical University, National Clinical Research Center for Child Health and Disorders, Ministry of Education Key Laboratory of Child Development and Disorders, Intelligent Application of Big Data in Pediatrics Engineering Research Center of Chongqing Education Commission of China, Chongqing, P.R. China.
- Oral Department, Children's Hospital of Chongqing Medical University, National Clinical Research Center for Child Health and Disorders, Ministry of Education Key Laboratory of Child Development and Disorders, Intelligent Application of Big Data in Pediatrics Engineering Research Center of Chongqing Education Commission of China, Chongqing, P.R. China. [email protected].
Abstract
This study aims to develop and validate an interpretable machine-learning model to quantify the predictive factors associated with hypodontia of permanent teeth in patients with non-syndromic cleft lip and palate (CLP). A total of 213 patients with cleft lip and palate were enrolled in this retrospective cohort study, and stratified into two groups: those with hypodontia of permanent teeth (n = 130) and without hypodontia (n = 83). Analytical variables were extracted directly from the patient's original medical records and high-resolution computed-tomography (CT) imaging data. Following Least Absolute Shrinkage and Selection Operator (LASSO) regression, both univariate and multivariate logistic regression analyses were conducted to identify independent risk factors. Subsequently, Logistic Regression (LR), Lasso regression and Random Forest (RF) were employed for comparative modelling. SHapley Additive exPlanations (SHAP) visualizations provided both global and local interpretations of the top-performing model. Based on these rankings, a diagnostic nomogram was further constructed and the accuracy of the diagnostic model was evaluated using ROC curves. After LASSO screening and integrating multivariable logistic regression analyses with clinical relevance, the final parsimonious model including the following variables: birth weight, number of supernumerary teeth, alveolar cleft width, incomplete cleft palate (ICP), complete cleft palate (CCP). Age sensitivity analysis excludes the age factor from the final model. RF demonstrated superior performance compared to other algorithms in precision-recall analysis. The top three predictive factors for increased hypodontia risk include a wider alveolar cleft width, ICP and a smaller number of supernumerary teeth. Receiver operating characteristic (ROC) curves analysis indicated that the model exhibited good predictive and diagnostic performance (the area under the receiver operating characteristic curve [AUC] = 0.766, confidence interval [CI] (0.743-0.872), Accuracy = 0.755). Finally, the diagnostic nomogram constructed based on SHAP values demonstrated a notably accuracy and predictive performance (AUC = 0.820, CI:0.761-0.879, Accuracy = 0.779). In this retrospective single-center cohort, five variables were identified to be associated with hypodontia of permanent teeth in CLP patients. The potential as candidate predictors that warrants further validation in prospective, multi-center studies. Registry: ClinicalTrials.gov, TRN: MR-50-24-035828, Registration date: September 5th, 2024.