Supply Chain Modeling for Temperature-Sensitive Pharmaceutical Goods

Abstract

The pharmaceutical industry relies on cold chain supply infrastructure to preserve the integrity of temperature-sensitive products. Specifically, passive controlled temperature solutions present an increased challenge as the inherent uncertainty of ambient temperatures and process lead time variations in the delivery dramatically increase the risk of inactivating the product. Despite the sponsor company's efforts to test their package solutions in laboratory temperature-control chambers, a lack of visibility exists on the likelihood of success of the packaging solution in real-life conditions. Hence, developing predictive forecasting capabilities for their deliveries across the United States can provide significant financial and operational benefits. This research studies and compares two families of methods of predicting temperature ranges of the goods: statistical methods, Autoregression and ARIMA, and machine learning methods, K-Nearest Neighbor, Support Vector Machines, Random Forest, Quantile Regression, and Long Short-Term Memory Neural Networks. Additionally, one-step and multi-step ahead forecasting techniques were analyzed in all models to determine the best forecasting approach. In addition, the forecasting models were tested on two types of packaging solutions, one for the summer profiles and the second for the winter profile. The results confirm that one-step ahead models outperform multi-step ahead forecasting for long-term horizons when compared by RMSE and MAE. Both statistical and machine learning models accurately predicted training and test set values with relatively lower RMSE. Nonetheless, it was found that testing the models in new external temperature conditions presented contradictory results for predicting the internal temperature, mainly due to the limited data set utilized to train and validate the models. Quantile Regression, on the other hand, successfully predicted the internal temperature of the payload’s given new ambient conditions. Therefore, we concluded that a forecasting model can be implemented as part of a predictive risk assessment analysis, considering the impact of variability in both temperature and process lead times for the sponsor company’s passive- controlled temperature solutions. These models can be extended for future applications with different configurations of insulator materials, amount of gel packs, and package dimensions.