How should we define severity phenotypes for long-term health conditions in analyses of primary care electronic health records?

Problem

Primary care electronic health records (EHR) are a rich real-world data source to study research priorities in patients with multiple long-term conditions (MLTC) who are often excluded from clinical trials. Studies of MLTC in EHR rarely account for disease severity since it is often not routinely recorded in patients’ records. However, those with severe disease are phenotypically very different to those with mild disease in terms quality of life, treatment regime and susceptibility to and progression of other conditions. We aimed to explore views and develop consensus on reliable proxy indicators of severity within EHR for analyses of key long-term health conditions.

Approach

We used a multiphase, building sequential mixed-methods study design to identify severity phenotypes within EHR for nine key long-term conditions including diabetes, heart disease, and depression. Informed by existing literature, our previous research and clinical experience, we identified potential severity phenotypes based on measures including disease complications, symptoms, medications, test results, and interventions. We then conducted exploratory data analysis in a primary care EHR database containing over 12 million patients' records to determine the feasibility of using these options in primary care data. Purposive sampling was used to recruit participants with both clinical training and expertise in analysing EHR. Participants completed a survey, and contributed to a structured nominal group technique discussion session, which was facilitated to elicit participants’ views (informed by the exploratory data analysis). Each participant used a 5-point Likert scale to rate clinical importance and feasibility of each of the proposed severity phenotypes independently.

Findings

11 male and 7 female clinical academics predominantly from general practice and public health backgrounds were recruited. Mean scores for clinical importance were highest for severity phenotypes based on disease complications (e.g. retinopathy in diabetes), and lowest for phenotypes based on symptom codes (e.g breathlessness in heart failure) and medication prescriptions (e.g. use of antidepressants in depression). However, use of prescription data and use of Quality and Outcomes Framework (QOF) incentivised measures (e.g. foot ulcer classification in diabetes) were ranked most feasible in EHR. Several clinical measurements such as ankle-brachial pressure index for peripheral vascular disease were too poorly accessible in the EHR database to be feasible measures of severity classification.

Consequences

Many severity indicators that are important to patients and clinicians are not well captured in primary care EHR records databases. Better incentivisation and standardised recording methods (such as via QOF or templates) for primary care staff may improve data quality in EHR. However, several important proxy measures of severity are feasible in electronic health records, and should be used to improve the granularity of future analyses in studies involving patients with multiple long-term health conditions.