Context-aware computing of actionable information using electronic health records data

Abstract

The vast amount of routinely collected information in Electronic Health Records (EHRs) is increasingly used for other purposes than direct care - and in particular, for research. Although research would not aim to improve the care for any specific patient, it can produce generalisable knowledge that can be translated into action within specific healthcare contexts. The main aim of this PhD thesis was computing actionable information from EHR data for specified healthcare contexts, including health research itself, population health management, and health information technology (HIT) engineering. We focused on two main research areas: predictive modelling using EHR data and patient portals. First, we explored how to use more effectively longitudinal EHR data to investigate multimorbidity. In a 10-year retrospective cohort study in the UK primary care, we tested different longitudinal comorbidity metrics and their value in predicting mortality. We found that explicitly accounting for longitudinal changes in comorbidities, as measured with the Charlson Comorbidity Index (CCI), better captures comorbidity burden on mortality, with more rapid changes in CCI posing a greater mortality risk. We suggest that the survival model proposed in the study should be considered by health researchers when investigating multimorbidity in EHR data. Second, we followed international guidelines to externally validate available models for predicting onset of CKD in the UK primary care (i.e. seven in total). We tested their performance on a five-year time horizon. All models had good discrimination on a five-year time horizon, however the majority over-predicted the CKD risk. QKidney, the only model originally developed in the UK, outperformed the other models and was shown to support a high risk approach to CKD prevention. This finding is actionable at a population health management level: on the basis of our results, policy makers should consider to update clinical practice guidelines by including QKidney among the CKD screening criteria. Finally, we focused on providing actionable information for HIT engineers. Particularly, we carried out a controlled study assessing whether patient interpretation and decision-making is influenced by the way the laboratory test results are presented to them in patient portals. We did not find any statistically significant differences between the three presentations that we tested, but we did find that misinterpretation of risk was high across all three presentations. Furthermore, we developed a method to calculate dynamic, patient-tailored alerts. Our method underwent proof-of-concept testing using one type of laboratory test value (i.e. potassium) and a group of GPs. Although representing a substantial methodological advancement and promising results were obtained, further evaluation of this method is required before HIT engineers can implement it in EHR systems. In this thesis, we used routinely collected EHR data to investigate decision-making in different contexts and involving different stakeholders. These included patients, clinicians and policy makers, as well as HIT engineers and researchers. Ultimately, we produced actionable information across health research and population health management, with methodological advances in predictive modelling using EHR data and findings from evaluation studies that are relevant to policy makers.

Date of Award	1 Aug 2017
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Iain Buchan (Supervisor), John Ainsworth (Supervisor), Niels Peek (Supervisor) & Sabine Van Der Veer (Supervisor)