Prediction and risk factor analysis of dental implant failure using Cox regression and machine learning

Survival analysis is commonly used to identify factors affecting dental implant longevity. While conventional studies typically employ the Cox proportional hazards model, the potential benefits of modern machine-learning algorithms remain unclear. In this study, we aimed to compare Cox regression with three machine-learning approaches-Random Survival Forest (RSF), DeepSurv, and TabNet-to evaluate their predictive performance and the risk factors they identify.

Clinical records of 543 implants placed in 256 patients were reviewed, with 32 failures observed over a mean follow-up of 8.0 ± 2.5  years. Cox models incorporating a patient-level frailty term were subjected to an exhaustive subset search. RSF, DeepSurv, and TabNet were tuned using 50 Optuna Bayesian trials, with the best settings refitted on a 90% training split and evaluated on the remaining 10%. Variable importance was assessed using hazard ratios for Cox regression and Shapley Additive Explanations for the machine-learning models.

The optimal Cox model achieved a concordance index (C-index) of 0.81 but demonstrated a limited area under the curve (AUC) performance. Among the machine-learning methods, RSF performed best (C-index, 0.72; AUC, 0.79; Integrated Brier Score, 0.044), outperforming DeepSurv and TabNet across all metrics. Both Cox regression and RSF identified anterior placement and periodontitis-related tooth loss as the primary risk factors for implant failure.

RSF showed good, but not superior, predictive accuracy compared with the best-performing Cox model. The choice of analytical method for implant survival studies should consider model complexity in relation to the dataset's sample size and event rate.

Dental implants are a well-established treatment for replacing missing teeth; however, reports of associated complications are increasing. Identifying risk factors and estimating the likelihood of failure can assist clinicians in treatment planning and follow-up care. In this study, we analyzed 543 implants placed in 256 patients with disabilities, monitored over an average of 8 years, during which 32 implant failures occurred. We first developed traditional statistical models using Cox regression, a well-established method for survival analysis. Then, we evaluated three modern machine-learning models capable of handling time-to-event outcomes: RSF, DeepSurv, and TabNet. Among the machine-learning approaches, the tree-based RSF demonstrated the highest predictive performance; however, it did not outperform the best Cox model after adjusting for several patient- and implant-related characteristics. Both RSF and Cox models consistently identified anterior implant placement and replacement of periodontitis-related tooth loss as the strongest predictors of failure. Our findings suggest that newer machine-learning tools do not inherently outperform conventional models, particularly in datasets with few events. The selection of analytical methods should be guided by dataset size and event frequency rather than the novelty of the algorithm.