Big Data Analysis of Electronic Health Records: Clinically interpretable representations of older adult inpatient trajectories using time-series numerical data and Hidden Markov Models

Abstract The implementation of Electronic Health Records (EHR) in UK hospitals provides new opportunities for clinical ‘big data’ analysis. The representation of observations routinely recorded in clinical practice is the first step to use these data in several research tasks. Anonymised data were extracted from 11 158 first emergency admission episodes (AE) in older adults. Irregular records from 23 laboratory blood tests and vital signs were normalized and regularised into daily bins and represented as numerical multivariate time-series (MVTS). Unsupervised Hidden Markov Models (HMM) were trained to represent each day of each AE as one of 17 state spaces. The visual clinical interpretation of these states showed remarkable differences between patients who died at the end of the AE and those who were discharged. All states had marked features that allowed their clinical interpretation and differentiation between those associated with the patients’ disease burden, their physiological response to this burden or the stage of admission. The most evident relationships with hold-out clinical information were also confirmed by Chi-square tests, with two states strongly associated with inpatient mortality (IM) and 12 states (71%) associated with at least one admission diagnosis. The potential of these data representations on prediction of hospital outcomes was also explored using Logistic Regression (LR) and Random Forest (RF) models, with higher prediction performance observed when models were trained with MVTS data compared to HMM state spaces. However, the outputs of generative and discriminative analyses were complementary. For example, highest ranking features of the best performing RF model for IM (ROC-AUC 0.85) resembled the laboratory blood test and vital sign variables characterising the ‘Early Inflammatory Response-like’ state, itself strongly associated with IM. These results provide evidence of the capability of generative models to extract biological signals from routinely collected clinical data and their potential to represent interpretable patients’ trajectories for future research in hypothesis generation or prediction modelling.