Integrating multi-dimensional prosthodontic features into an interpretable XGBoost framework for cross-sectional diagnostic classification of peri-implant diseases

Objective The accurate prediction of peri-implant disease severity is crucial for timely intervention and successful long-term outcomes. This study aimed to develop and validate an interpretable machine learning model that evaluates whether multi-dimensional prosthodontic configurations alone, without relying on direct inflammatory indicators, can assist in classifying the cross-sectional severity of peri-implant diseases. Methods A retrospective analysis was conducted on 300 implants from 285 patients. Data encompassed patient demographics, implant characteristics, and prosthodontic features, including Margin Fit (MF), Screw Access Space (SAS), prosthesis type, prosthesis material, and implant position. To prevent label leakage and tautological reasoning, direct diagnostic criteria (Probing Depth and Bleeding Index) were strictly excluded from the predictor matrix. The dataset was partitioned at the patient level into an internal training set ( n = 200 patients; 210 implants) and an internal hold-out validation set ( n = 85 patients; 90 implants). Four machine learning algorithms (Ordinal Logistic Regression, SVM, Random Forest, and XGBoost) were developed and validated. Model performance was evaluated using accuracy, F 1-score, and AUC. Model interpretability was assessed using SHapley Additive exPlanations (SHAP), and clinical utility was determined via Decision Curve Analysis (DCA). A nomogram was also constructed. Results The XGBoost model achieved the highest discriminative performance on the internal validation set, with an overall accuracy of 78.9% and a macro-averaged AUC of 0.821 (95% CI: 0.760–0.882). SHAP analysis identified Margin Fit, Screw Access Space, and prosthesis material as the most influential features associated with disease stage classification. Decision Curve Analysis (DCA) demonstrated a positive clinical net benefit across decision threshold probabilities from 15% to 65%. The multivariable ordinal logistic regression model yielded a diagnostic AUC of 0.774, confirming the additional discriminative gain provided by non-linear ML ensembles. Conclusion A non-linear machine learning framework incorporating prosthodontic features—free of label-leaking clinical indices—can objectively aid in classifying peri-implant disease stages in a cross-sectional setting. This tool should be framed as an automated diagnostic aid rather than a longitudinal prognostic system, pending multi-center external validation.