Machine learning prediction of human blood optical properties: a comparative study of optimized algorithms with SHAP-based interpretability analysis.

Accurate characterization of the optical properties of blood is essential for biomedical optics, diagnostic imaging, and therapeutic applications. conventional measurement techniques are frequently constrained by complex setups and idealized assumptions, highlighting the need for robust predictive models. This study presents a comparative machine learning framework to predict three blood optical properties: the absorption coefficient (µa), the scattering coefficient (µs), and the refractive index (RI). We developed and systematically optimized support vector regression (SVR), random forest (RF), and extreme gradient boosting (XGBoost) models using the Optuna framework, training them on a comprehensive dataset utilizing wavelength, hematocrit (Hct%), and oxygen saturation (SO2%) as input features. On the test dataset, the Optuna-optimized RF model achieved the highest predictive accuracy, with an R2 of 0.9981 and a root mean squared error (RMSE) of 1.3680. It significantly outperformed both SVR (R2 = 0.9736) and XGBoost (R2 = 0.9865), demonstrating superior generalization capability. SHAP analysis provided model interpretability, confirming that the predictions were based on a physically meaningful relationship. Wavelength was the dominant predictor, while hematocrit and oxygen saturation contributed with light-tissue interaction principles, reinforcing model validity. Additionally, the analysis demonstrated that tree-based models exhibit superior feature utilization with clearer decision boundaries compared to SVR, more effectively capturing the nonlinear, hierarchical relationships inherent in biological systems. Finally, the optimized models were deployed via a user-friendly web interface, providing the biomedical optics community with accessible, interpretable tools for rapid optical property prediction to advance fundamental research and clinical applications.