Development and implementation of machine learning models for dynamic risk prediction models in health care applications (2023)

Traditional risk prediction generates a risk estimate at a defined timepoint in a patient’s disease trajectory, for example the risk of death within 30 days following a surgical procedure. In contrast, dynamic risk prediction enables prediction of risk at any time point. This allows to continuously monitor a patient’s risk profile and forms the basis for intervention if the predicted risk increases. In this course, we will explore methodological and technical solutions, as well as corresponding challenges, for developing and implementing such solutions in health care. The course includes the following topics: 1) Data management: This part of the course considers the challenges of preparing heterogenous longitudinal health data for prediction. We will cover the various steps involved in this process, including data formatting, feature engineering, and splitting strategies for model validation. This will include discussion about how to handle irregularly sampled health data, data leakage, class imbalance, temporal robustness, normalization, and other potential biases. 2) Modelling: In this part of the course, participants will be led through the process of building such models. We will introduce both basic and more advanced dynamic machine learning prediction algorithms, such as gradient tree boosting, random forest, and LSTM and discuss issues related to performance metrics and hyperparameter optimization, for example Bayesian optimization. 3) Implementation: In the last part of the course, we will consider the challenges associated with the implementation of predictive tools in the clinic. This includes technical aspects about hosting, user interface, and access to live data, including an introduction to the FHIR standard. Regulatory and organizational issues will also be discussed. During the project the participants will get hands on experience covering realistic scenarios related to the subjects discussed. This will include data management of representative data sets, training models and hands on introduction to the FHIR set-up