Predictive modeling of patulin accumulation in apple lesions infected by Penicillium expansum using machine learning

Abstract

Penicillium expansum causes blue mold in fruit, leading to the accumulation of patulin, a harmful toxin. Current detection methods for patulin are costly and complex, highlighting the need for a simplified approach. This study investigates the impact of P. expansum on ‘Golden Delicious’ and ‘Fuji’ apples, focusing on significant changes in lesion characteristics and fruit properties. Correlation analysis revealed strong positive associations between patulin content and lesion diameter, lesion depth, and fruit weight loss, while negative correlations were observed with apple firmness and flesh firmness. Of the four machine learning models used - Back Propagation Neural Network (BPNN), Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Gradient Boosting Regression Tree (GBRT) - all showed good predictive performance without overfitting the data. BPNN outperformed the others in accurately predicting patulin accumulation (Golden Delicious: RMSE=0.72, MAE=0.217, R²=0.982; Fuji: RMSE=0.289, MAE=0.08, R²=0.994). Key features influencing patulin accumulation, identified using SHapley Additive exPlanation (SHAP) measures, included lesion diameter, lesion depth, and fruit weight loss. This research highlights the potential of machine learning to predict patulin accumulation based on these critical factors, providing a simplified detection method for postharvest fruit quality management.