Electronic health record (EHR) systems provide a rich source of timestamped patient data which can be leveraged for several secondary purposes. Using temporal EHR data to develop models for clinical outcome prediction is a rapidly expanding and promising area of research. However, while state-of-the-art methodologies have been developed for modeling EHR data, there are underlying data issues that reduce model utility in clinical settings. Among these issues are data shifts, or differences between the setting for model training and model implementation that can negatively affect prediction performance. Of the methods developed for EHR data that do not require data from target environments are those that require expert knowledge of the causal pathways for each outcome of interest. However, many conditions, including Long COVID, are caused by poorly understood physiological processes. We thus focus on alternative methods which use data from heterogeneous sources to extract invariant predictors without explicit knowledge of causal pathways. In the process of adapting these methods for prediction of Long COVID onset, we also developed reproducible phenotyping methods to consistently identify Long COVID patients in the absence of a formal diagnosis.