Towards the ideal vascular implant: Use of machine learning and statistical approaches to optimise manufacturing parameters

Abstract

Introduction: Electrospinning is a manufacturing technique that creates a net of nano and microfibres able to mimic the natural extracellular matrix (ECM) of biological tissue. Electrospun scaffolds' morphology and mechanical behaviour can be tailored by modifying the environmental, solution and process parameters. This study aims to produce biomimetic vascular implants optimising the manufacturing set up through two machine learning techniques and statistical approaches.Methods: Polyvinyl alcohol (PVA) based scaffolds were produced by modifying the concentration of the polymer, flow rate, voltage, type of collector, diameter of the needle, distance between needle and collector and revolutions of the mandrel. The scaffolds were morphologically and mechanically characterised using scanning electron microscopy and mechanical testing respectively to inform the morphological model (simultaneously predicting diameter of the fibres and inter-fibre separation) and mechanical model (predicting strain at rupture and ultimate tensile strength).Results: Prediction and traditional techniques led to an optimum set up of: 12% PVA, 1 ml/h flow rate, 20 kV, 8 cm between the needle, 18 G gauge needle, rotational mandrel of 15 cm and 2000 rpm. Optimised PVA scaffolds replicated the mechanical properties and morphology of the vascular tissue with an ultimate tensile strength of 6.17 ± 0.18 MPa, a strain at break of 97.39 ± 5.06, fibre diameters of 126 ± 6.11 nm and inter-fibre separation of 1488 ± 91.99 nm.Discussion: This work revealed for the first time that machine learning Chi-squared Automatic Interaction Detection (CHAID) models are a novel and visual route to elect the optimum manufacturing set up to develop biomimetic vascular implants. Novel two-output Artificial Neural Networks (ANN) and multivariate analysis of variance and covariance (MANOVA, MANCOVA) models presented comparable prediction results (R2=0.91); however, two-output ANN predicted models demonstrated to be the most powerful tool for non-parametric conditions, showing cross-validation mean squared errors (MSE) of 0.0001943. Multi Linear Regression models (MLR) exhibited the lowest accuracy in their predictions (R2=0.6). Machine learning, statistical approaches and traditional characterisation methods were studied to successfully achieve vascular substitutes with analogous biomechanical behaviour and physical structure to the native vascular tissue.