Non-destructive chick embryo mortality prediction at pre-incubation and early incubation using hyperspectral imaging and explainable artificial intelligence

1. This study evaluated the potential of visible-near infrared (Vis-NIR) hyperspectral imaging (HSI) combined with machine learning and explainable artificial intelligence (AI) to non-destructively predict chick embryo mortality before incubation and at 4 d of incubation.2. The partial least squares discriminant analysis (PLS-DA), random forest (RF) and categorical boosting (CatBoost) calibration models were developed and independent validation and test sets evaluated the performance of the calibration models. In addition to raw figures, synthetic data was utilised for classification model development. Various spectral pre-processing and feature selection methods were evaluated to enhance predictive robustness. The best model was interpreted using Shapley additive explanations (SHAP) for AI.3. At full wavelength (501-921 nm), the PLS-DA model demonstrated the best performance for chick embryo mortality classification, achieving an accuracy of 91.3% for calibration, 88% for validation and 86.7% for the test set for pre-incubation. At d 4 of incubation (ED4), the model showed 97.3% accuracy for calibration, 96% for validation and 97.3% for the test set, highlighting its robustness across different data sets.4. The PLS-DA models, using a reduced set of important spectral features, demonstrated strong predictive performance, offering computational efficiency, robustness and enhanced interpretability.5. The SHAP explainable AI revealed that wavelengths associated with embryo hydration status, blood formation and metabolic differences between live and dead embryos are critical for classifying chick embryo mortality during early incubation.