AAAI-22: Time series in healthcare

Time series datasets such as electronic health records (EHR) and registries represent valuable (but imperfect) sources of information spanning a patient’s entire lifetime of care. While learning from temporal data is an established field and has been covered in a number of prior tutorials, the healthcare domain raises unique problems and challenges that require new methodologies and ways of thinking.

Perhaps the most common application of time series is forecasting. While we will discuss state-of-the-art approaches for disease forecasting, we will also focus on other important problems in time series, such as time-to-event or survival analysis, personalized monitoring, and treatment effects over time. These topics will be introduced in the context of healthcare, but they have broad applicability to other domains beyond medicine.

In addition, we will explore several characteristics that are necessary to make AI and machine learning models as useful as possible in the clinical setting. We will discuss automated machine learning and we will address the challenges of understanding and explaining machine learning models as well as uncertainty estimation, both of which are critical in high-stakes scenarios such as healthcare.

We will aim for minimal required prerequisite knowledge. However, we will assume basic knowledge of standard machine learning methods (e.g. MLPs, RNNs). While our tutorial will include some detailed explanations of machine learning techniques, significant focus will be placed on the problem areas, their unique challenges, and ways of thinking to overcome these.