Explainable AI for predicting crystallinity in doped hydroxyapatite: Implications for biomedical applications

Abstract

Hydroxyapatite (HAp) is a widely used bioactive material in medical applications such as bone tissue engineering, where the crystallinity plays a critical role in determining mechanical strength, bioactivity, and bioelectric behavior. The crystallinity is strongly affected by complex interactions between dopant elements and processing conditions. In this study, a public database was used to develop an explainable artificial intelligence (AI) framework to predict the crystallinity of multi-doped HAp based on compositional parameters. A public dataset comprising 37 samples with various dopants (e.g., Sr, Zn, Ag, F, and transition metals), processing temperatures, and duration was used. Three machine learning (ML) models, including XGBoost, random forest, and decision tree, were developed using all features. An explainable AI method, Shapley additive explanations (SHAP), was used to identify the most relevant features for each model. The features relevant to all models were selected (temperature, processing duration, Zn, HAp, Er, and Al). Model performance was assessed before and after feature reduction. XGBoost achieved the best performance with all features (R² = 0.917, root mean squared error as RMSE = 6.03), while feature reduction slightly decreased its performance (R² = 0.881, RMSE = 6.18). In contrast, RF and DT showed notable improvements after feature selection, with RF increasing from R² = 0.814 to 0.872 and DT from R² = 0.574 to 0.688, indicating reduced overfitting and improved generalization. Final model interpretation using SHAP revealed that processing parameters (time and temperature) dominate crystallinity prediction, followed by specific dopants such as Zn and Er. It was demonstrated that combining SHAP-based consensus feature selection with ML provides both accurate prediction and meaningful interpretation, offering valuable insights into the factors governing HAp crystallinity.