Optimization of primary screw stability in Trabecular bone using neural network-based models

• Published in: Computer methods and programs in biomedicine Vol. 264; p. 108720
• Main Authors: Zhou, Yijun, Helgason, Benedikt, Ferguson, Stephen J., Persson, Cecilia
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.06.2025
• Subjects: Bone Screws; Cancellous Bone - diagnostic imaging; Cancellous Bone - surgery; Design optimization; Humans; Linear Models; Machine Learning; Neural network; Neural Networks, Computer; Orthopaedic screw; Pull-out; Surrogate model; Tomography, X-Ray Computed; Neural network; Orthopaedic screw; Surrogate model; Pull-out; Design optimization
• ISSN: 0169-2607; 1872-7565; 1872-7565
• DOI: 10.1016/j.cmpb.2025.108720

•Goal: optimize patient-specific screw designs to enhance primary stability.•Neural networks predict pull-put stiffness and strength with 2–6 % error•Optimized designs showed approximately 15 % improvement in pull-out stiffness and strength. Screw implant stability in bone is crucial to the success of many orthopaedic procedures, yet the relationship between screw design parameters and specific bone characteristics remains underexplored. This study aims to optimize screw designs to enhance primary stability by leveraging subject-specific bone data and advanced surrogate modelling techniques. In this study, 2880 screw pull-out simulations were conducted to assess primary screw stability by analysing pull-out stiffness and strength. The resulting dataset was used to develop surrogate models using multiple linear regression, random forest, and neural networks (NN). An optimization process was then applied to find optimal screw designs for 80 distinct trabecular bone specimens, in terms of inner diameter, pitch, and thread angle. The models, trained with various input parameters, including bone morphological parameters and computed tomography images, promisingly predicted the results of the simulations. The prediction errors varied by model type, with multiple linear regression yielding approximately 12 % error, while non-linear machine learning models achieved lower errors, ranging between 2–6 %. The series of subsequent optimization tasks provided optimized screw designs showing statistically significant improvements in pull-out stiffness and strength compared to the average screw designs (approximately 16 and 14 %, respectively). This even though our study focused only on screw design parameters that generally have a smaller impact on stability compared to factors such as screw outer diameter and insertion depth. Multiple linear regression models were found to be insufficient for generating optimized screw configurations, and more complex surrogate models, such as NN, are needed. It could be concluded that different trabecular bone morphologies can benefit from distinct optimal screw designs. The insights gained from this study could have implications for the development of patient-specific orthopaedic treatments.