Empirical evaluation of performance degradation of machine learning-based predictive models – A case study in healthcare information systems

While there have been a very large number of academic studies of proposed machine learning-based health predictive models, it is widely recognized that machine learning-based models in all domains typically degrade in performance over time, post training. This known characteristic of machine learning-based models could present significant risks in the healthcare setting to patient quality of care or safety. Nevertheless, there has been little study of the performance degradation of such models on real-world data. In this article, we empirically measure performance degradation of predictive models that predict at time of admission, emergency patient mortality, drawing upon a large dataset of over 1.83 million patient discharge records. We demonstrate important empirical results including both relatively slow performance degradation over two and a half years, but also significant differences in the rate and extent of performance degradation between different machine learning model types and time period of the training set.