Temperature deviations are critical in a pharmaceutical supply chain (SC) due to quality deterioration concerns and resulting health risks. The current solutions ensuring temperature maintenance are either labor-intensive or prone to triggering alarms that require no corrective measures, which, in turn, increase the alarm investigation costs. Machine learning (ML) methods have fared well both in the areas characterized by the execution of repetitive tasks and in the identification of false alarms; however, they have not been applied in the context of temperature monitoring in a pharmaceutical SC. In this paper, we used the real-world data of a large international logistics service provider for the period of 2013–2018 and compared the optimized performance of 10 ML classification methods in the task of false temperature alarm identification. Such additional features as temperature in the location of possible physical handling and average temperature deviation were either externally collected or estimated to enrich the models. In general, gradient boosting achieved the best performance in our evaluations, with an accuracy of 95.9% in comparison with the value of 16.6% demonstrated by the current legacy rule-based system. The feature ranking and sensitivity tests pointed to the strength of the features indicating an absolute temperature deviation and the location of cargo along the SC. The tests simulating model applications on new dissimilar observations showed various performance losses across classifiers, with the best stability retained for a new customer scenario and largest performance decreases for a new temperature range scenario.

由于对质量下降的关注以及由此带来的健康风险，温度偏差在药品供应链（SC）中至关重要。当前确保温度维护的解决方案要么是劳动密集型的，要么容易触发不需要更正措施的警报，这反过来又增加了警报调查的成本。机器学习（ML）方法在执行重复性任务和识别虚假警报方面都表现良好。但是，它们尚未应用于药物SC中的温度监控。在本文中，我们使用了一家大型国际物流服务提供商2013-2018年的真实数据，并比较了10种ML分类方法在错误温度警报识别任务中的优化性能。外部收集或估计了可能的物理处理位置中的温度和平均温度偏差等其他特征，以丰富模型。通常，梯度提升在我们的评估中获得了最佳性能，其准确度为95.9％，而当前的基于旧规则的系统显示的数值为16.6％。特征等级和灵敏度测试指出了特征的强度，表明了绝对温度偏差和沿SC的货物位置。在新的不同观测值上模拟模型应用程序的测试显示了各分类器的各种性能损失，其中对于新客户场景保留了最佳稳定性，而对于新温度范围场景则保留了最大的性能下降。