This study investigated the correlation of age in male and female specimens with physico-mechanical properties of trabecular bone including compressive strength, bone volume fraction, structural model index, trabecular thickness factor, level of inter-connectivity and pore morphology. An artificial neural network was designed to analyse 35 available samples in order to account for complex inter-dependencies of the key parameters in multi-dimensional space. Trained by using Levenberg-Marquardt back propagation algorithm, the network achieved regression factor of 0·96 by optimisation and showed that age correlates strongly with the physical properties of the bone affected by severe osteoarthritis. In addition, the compressive strength was found to be the most important factor for predicting the bone aging. Within the limitations of the input data set, the model developed provides a reliable predictive tool to tissue engineering applications.
IntroductionBone is a natural biocomposite comprising hierarchical cortical and trabecular (cancellous) structures. The skeletal structurefunction relationship of trabecular tissue in vivo is complex 1 since the mechano-biological properties are intrinsically dependent on the physical and geometrical parameters of bone, 2 for example, three-dimensional (3D) trabecular architecture, 3 pore size, compressive strength, 4 yield strain and modulus, 5,6 strain energy density and bone remodelling. The task of optimising or predicting these properties especially for trabecular bone becomes even more onerous due to the fact that mechanical properties of bone differ according to their anatomical location.7 Tissue engineers are thus often faced with problems of selecting the most successful strategy for both the design and fabrication of synthetic scaffold for the treatment of patients suffering from degenerative orthopaedic diseases triggered by osteoarthritis, osteoporosis, trauma, injury and metastatic cancer occurring in specific age group and gender. 8,9 Computational techniques such as finite-element analysis (FEA) and other mathematical procedures have been used for clinical data in hip fractures with limited success.10-12 Phenomenological (data-driven) models, based on the experimental or clinical data, are known to have poor accuracy and restricted by the size of available data sets.13 Resulting statistical models are generally very sensitive to the inaccurate data and outliers, as well as the level of ambiguity allowed by the fuzzy logic and soft algorithms. This necessitates generation of large volume and high-quality data, and in tissue engineering, these requirements are highly impractical and often unrealistic. To resolve this dilemma and make statistical datadriven models applicable, it is desirable to design a model based on the principles of machine learning, which enables the ability of such model to learn from the additional data and to dynamically adapt to new inputs. [14][15][16][17]