Explainable Machine Learning Framework for Predicting Cold-Chain Failure Risk in Vaccine Supply Chains

Cold-chain failures during vaccine storage and transport can reduce vaccine quality and cause serious public health risks. This study proposes a hybrid explainable artificial intelligence and machine learning (XAI-ML) framework to predict cold-chain failure risk in vaccine supply chains. The framework focuses on three main problems: imbalanced risk data, limited model interpretability, and practical use in real supply chain operations. The study used a dataset of 26,374 records and 18 features covering environmental, logistical, and operational factors. Several balancing methods were tested, including Random Oversampling, SMOTE, Tomek Links, Instance Hardness Threshold, and their hybrid combinations. Among all models, the Gradient Boosting Classifier with SMOTE and Tomek Links performed best. It achieved 96.31% accuracy, 96.50% precision, 96.31% recall, 96.36% F1-score, 91.73 Kappa, 91.78 MCC, and 98.93% ROC-AUC. The same preprocessing steps were used for all models, including missing value treatment, scaling, label encoding, label binarization, and a 70/30 train-test split. No manual re-labeling or extra data cleaning was applied. To explain the model results, SHAP, LIME, and Permutation Importance Analysis were used. The main risk factors were out-of-bound temperature hours, location, and ultra-low temperature freezer hours. Against a state-of-the-art XGBoost baseline re-implemented on the same dataset, the proposed framework achieved higher accuracy (96.31% versus 96.22%) and stronger class-balance metrics, and it also exceeded the 94% accuracy of a prior cold-chain benchmark that is used only as an indicative external reference. The model can support early action, such as rerouting shipments, adjusting storage temperature, and warning logistics staff before vaccine damage occurs.