Enhancing Dental Implant Risk Prediction with an Interpretable Multi-Instance Learning Model

Variations in clinical and biological factors often lead to differing risks of dental implant failure among patients, even when undergoing similar procedures. Traditional predictive models often oversimplify outcomes into binary classifications, lacking the interpretability needed for accurate risk assessment and effective patient stratification. To address this challenge, this paper presents a novel multi-instance learning (MIL) framework that incorporates a Hosmer-Lemeshow-based loss function for implant failure risk assessment based on patient-specific clinical features. The framework also identifies the optimal threshold for key features, such as bone density, to enable robust patient stratification. The effectiveness label for each patient was constructed first and sampled patients into 600 and 1,000 groups. Results demonstrate that the proposed framework captures inter-patient variability with improved statistical calibration and enhanced risk stratification, offering practical insights to guide clinical decision-making. This study highlights the scalability and versatility of the proposed framework, bridging computational methodologies and practical applications in personalized implantology. Furthermore, the approach provides a transparent and effective tool for risk assessment, with potential applications in broader clinical stratification problems.