Machine Learning-Guided Selection of 3D Printing Parameters for Hybrid Biomaterial Scaffolds: A Leakage-Controlled, Multi-Objective Analysis of Open Bioprinting and Fused Deposition Data

Hybrid tissue-engineering scaffolds combine a printed thermoplastic frame for load bearing with a cell-laden hydrogel for biological function, but their printing parameters are usually chosen by trial and error. This study evaluates whether open data and machine learning can guide that choice. Three openly released datasets were used: 617 extrusion-bioprinting cell-viability records and 339 filament-printability records compiled from 75 published studies, and 500 fused deposition modelling runs with tensile and surface measurements. Each dataset was modelled separately with ridge or logistic regression, random forest, XGBoost, Gaussian process and multilayer perceptron models under repeated nested cross-validation. Splits for the bioprinting data were grouped by source study to prevent leakage. Models were interpreted with SHAP values and partial dependence, and NSGA-II and exhaustive surrogate search with bootstrap uncertainty and an applicability-domain constraint were used to propose parameter windows. Tensile strength and modulus were predicted accurately (XGBoost R² = 0.92 and 0.96). Printability transferred moderately to unseen studies (AUC = 0.82), whereas continuous viability did not (best R² = 0.01). Random K-fold splitting inflated viability R² from -0.27 to 0.49 and printability AUC from 0.63 to 0.96. PLA+ frame settings with predicted lower-bound strength above 30 MPa were found at up to 80 % nominal void fraction. No alginate-based bioink window inside the data domain met both the viability and printability targets with confidence. Literature-pooled bioprinting data therefore support frame design but not quantitative bioink design across laboratories, and all recommendations require experimental validation.