Characterisation of inpatient trajectories as clinically interpretable representations using Hidden Markov Models from anonymised electronic hospital records in older adults

Aim: The aim of this study is to assess the interpretability of a representation of patients’ trajectories, collected from electronic health records (EHR) in the form of multivariate time series, using both time aware machine learning techniques, such as Hidden Markov Model (HMM) or neural networks and fixed frame machine learning techniques. . The quality of this representation is evaluated comparing its value as input variables for the prediction of adverse outcomes in clinical admission episodes (inpatient mortality, 30-day post-discharge readmission and 30-day post-discharge mortality) and clinical diagnosis at admission and discharge. Previous work: An unsupervised HMM model has been trained using expectation maximisation using as input data numeric multivariate time series from laboratory tests results and vital signs from EHR patients’ data. No outcome data was used in the model, or the training of the model or the assessment of the model. As an extra precaution a random subset of the data was held back as fully “unseen” data so far. This generative model has been used to represent each day of admission as a unique state instead of 23 numeric variables. The visual interpretation of this characterization of patients’ trajectories has shown relationships between states and outcomes of admission (e.g., inpatient mortality or length of admission episode) and diagnosis at admission and discharge. Some of these relationships have been confirmed using chi-squared t-tests but a more detailed assessment of the quality and interpretability of this representation will be explored in this research.