Abstract
Dental implant failure and subsequent reimplantation are sequential clinical events influenced by complex, patient-specific factors. Traditional survival models often fail to capture the conditional dependencies in multi-stage outcomes. This study introduces a multi-task survival framework that integrates conditional masking, attention-based feature interactions, monotonicity regularization, and dependency alignment to jointly predict implant failure and subsequent reimplantation. The model was trained on real-world data from 1,627 patients with 57 clinical features. Model performance was assessed using five-fold center-blocked cross-validation across multiple discretization intervals. The proposed approach achieved superior predictive accuracy. For implant failure, the concordance index was 0.8133, the area under the curve was 0.8321, and the Brier score was 0.1354. For reimplantation, these metrics were 0.9722, 0.9885, and 0.0352, respectively, using inverse probability of censoring weighting to ensure valid time-dependent discrimination and calibration under right censoring. The framework consistently outperformed established survival models, including classical statistical methods such as Cox proportional hazards and deep learning approaches such as DeepSurv. Risk scores effectively discriminated and tracked progression across patient groups. Feature importance analyses revealed both shared and stage-specific predictors, and calibration curves confirmed the reliability of the estimates. This framework models sequential implant outcomes and provides individualized risk estimates that may inform personalized treatment planning, monitoring, and clinical decision-making.