Predicting the risk of mission failure in shipments of pharmaceuticals

We applied a set of machine learning techniques to predict the risk of shipment-failure in shipments of temperature-sensitive pharmaceutical goods.

Background

This client is one of the world’s leading providers of solutions for tracking sensitive pharmaceutical goods during shipment and throughout the whole supply chain. Each shipment is tracked minute-by-minute by a temperature logger. If the goods have been exposed to prolonged temperatures above certain thresholds by the final destination, the shipment will be rejected.

Solution

The models were based on time series and auxiliary data from hundreds of thousands of historical missions. The approach developed can be used to optimize the choice of packaging solution, and transport route, and act as the basis for a warning system that notifies the operators of increased risk along the way. The final solution was deployed in a web app for showcasing and internal use.

Tools/Tech
We used bayesian regressions to estimate the distributions in shipment max and mean kinetic temperatures, tree ensemble models to build classifiers for probability point estimates, and Deep neural networks for ahead of time predictions of the time series. Preprocessing, modeling, and validation flows were developed in Python with data fetched from a relational database. Deep learning architectures were developed in PyTorch and Tensorflow.