Deformation prediction in innovative implant design with machine learning approaches

Abstract

This study includes experimental research on a classic implant and a sleeved implant, both designed for experimental purposes and produced as samples for testing. The effects of the sleeved implant, an innovative design in dental implantology, on the jawbone were examined through experimental tests, and the data obtained were analyzed using machine learning (ML). Unlike classic implants, the sleeved implant is predicted to provide a more homogeneous distribution of chewing forces on the bone by reducing stress concentrations around the implant. This innovative design aims to strengthen the integration between the jawbone and the implant, increase the implant’s long-term stability, and reduce bone deformation.

In the experimental part of the study, sleeved and classic implants were produced as samples mounted on cow bones. Then, these implants were tested by subjecting them to various pressures and chewing forces using a double-acting pneumatic mechanism. The deformations of the implants on the bone were recorded with digital load cells and measured with relevant measuring devices. 4263 usable experimental data were collected for the classic implant, and 8832 for the sleeved implant. The data obtained were modeled using a finite element analysis system (ANSYS), and instantaneous deformation data were collected during the modeling process. These instantaneous deformation data were included as an additional feature in the ML dataset and used in the analysis processes.

In the study, the Kernel Support Vector Machine (Kernel SVM), Kernel Logistic Regression (Kernel LR), and extreme gradient boosting (XGBoost) classification methods were employed to assess the impact of the implant on the jawbone. In the ML models applied to the experimental data, the deformation levels created by the sleeved and classic implants were classified as slightly, medium, and serious loss classes. Among the evaluated models, the XGBoost model demonstrated outstanding classification performance for both implant types. For the XGBoost model, the training and test accuracies were 100 % for both classic and sleeved implants. The model was also subjected to the robustness test, which confirmed its stability and consistent performance across varying data conditions. Moreover, the Kernel SVM model provided 95 % training accuracy and 95 % test accuracy for the traditional implant and 92 % training accuracy and 95 % test accuracy for the sleeved implant. Finally, the Kernel LR model provided 92 % training accuracy and 93 % test accuracy for the classic implant. For the sleeved implant, a training accuracy of 88 % and a test accuracy of 90 % were achieved. These results show that XGBoost was the best-performing model overall, followed by Kernel SVM, while the Kernel LR model performed slightly lower on the sleeved implant than the classic implant. Additionally, the study evaluated precision, recall, F1-score, and Area Under the Receiver Operating Characteristic Curve (AUROC) metrics. The ROC analyses revealed that XGBoost exhibited the highest classification performance among all models, followed by Kernel SVM, while Kernel LR showed relatively low, yet acceptable, discrimination capability.

According to experimental results, it has been determined that the sleeved implant produces similar effects and results to the classic implant. The findings of the obtained ML models confirm that the sleeved implant creates deformation at a similar level to the classic implant and can be considered a reliable alternative in implantology.