Work

Development and Machine Learning-Based Optimisation of a Microfluidic Bone-on Chip Model

Year: 2026
Type: dissertation
Abstract: Traditional cell culture and animal models cannot reproduce the dynamic mechanical and biochemical environment of human bone, limiting studies of remodelling, osteogenesis and diseases such as osteoporosis. This PhD develops a microfluidic bone-on-chip (BoC) platform that combines an optimised hydrogel to create a 3D culture environment, combined with mechanical stimulation to promote osteoblast proliferation and differentiation. The aim of this research was to refine the BoC environment by systematically analysing cell responses to flow shear, cyclic loading, and incubation conditions. A prototype was designed, fabricated and validated for long-term use by leak and flow tests. MC3T3 preosteoblast cells were seeded and their responses were measured under varying conditions. Optimisation targeted a hydrogel coating with suitable composition and stiffness and adjusted oxygen and media (including a cell protectant supplement) to maximise viability, proliferation and long-term development under mechanical stimulation. The optimal setup used a 1 mm hydrogel of 10 % GelMA and 1 % HAMA under normoxia (19 % O2), with 10 % (w/v) cell protectant. Multi-stage experiments evaluated combinations of centrifuge-based cyclic loading and flow shear, with and without cell protectant. Stimulated and supplemented cultures maintained higher levels of cell viability and proliferation than static controls. Flow stimulation had the strongest effect, increasing proliferation and differentiation by 60 to 100 % at 4 to 5 weeks, followed by interaction effects of cell protectant (30 to 45 %) and centrifugation (25 to 35 %) at later stages (5 to 7 weeks). Machine learning (ML) models were then trained on experimental data for data-driven optimisation of in-chip bone tissue culture. Bayesian-optimised Random Forest models performed best (R2 = 0.93). Deep learning and neural networks were trained on extended data, of which Randomised deep learning was the best (R2 = 0.77 with the best output). PCA and Explainable AI methods (LIME and SHAP) clarified model behaviour and data. This work presents a novel BoC platform that integrates multimode dual mechanical stimulation with data-driven modelling, advancing strategies for bone tissue regeneration.
Language: English
Open Access status: green
OpenAlex ID: W7224984030
Cites: 272
Cited by:
Related to: 9