Feature importance in multi-dimensional tissue-engineering datasets: Random forest assisted optimization of experimental variables for collagen scaffolds

Ice-templated collagen-based tissue-engineering scaffolds are ideal for controlled tissue regeneration since they mimic the micro-environment experienced in vivo. The structure and properties of scaffolds are fine-tuned during fabrication by controlling a number of experimental parameters. However, this parameter space is large and complex, rendering the interpretation of results and selection of optimal parameters to be challenging in practice. This paper investigates the impact of a cross section of this parameter space (drying conditions and solute environment) on the scaffold microstructure. Qualitative assessment revealed the previously unreported impact of drying temperature and pressure on pore wall roughness, and confirmed the influence of collagen concentration, solvent type, and solute addition on pore morphology. For quantitative comparison, we demonstrate the novel application of random forest regression to analyze multi-dimensional biomaterials datasets, and predict microstructural attributes for a scaffold. Using these regression models, we assessed the relative importance of the input experimental parameters on quantitative pore measurements. Collagen concentration and pH were found to be the largest factors in determining pore size and connectivity. Furthermore, circular dichroism peak intensities were also revealed to be a good predictor for structural variations, which is a parameter that has not previously been investigated for its effect on a scaffold microstructure. Thus, this paper demonstrates the potential for predictive models such as random forest regressors to discover novel relationships in biomaterials datasets. These relationships between parameters (such as circular dichroism spectra and pore connectivity) can therefore also be used to identify and design further avenues of investigation within biomaterials.