Defining phenotypes of disease severity for long-term cardiovascular, renal, metabolic, and mental health conditions in primary care electronic health records: A mixed-methods study using the nominal group technique

Abstract

Objective

Inclusion of severity measures for long-term conditions (LTC) could improve prediction models for multiple long-term conditions (MLTC) but some severity measures have limited availability in electronic health records (EHR). We aimed to develop consensus on feasible severity phenotypes for nine cardio-renal-metabolic and mental health conditions.

Methods

This was a mixed-methods study using novel methodology. From existing literature, we identified potential severity phenotypes and explored feasibility of their use in EHR through analysis of data from 31 randomly selected general practices in the Clinical Practice Research Datalink (CPRD) Aurum database, a large UK-based primary care EHR database. We recruited clinical academic experts to participate in a survey and nominal group technique workshop. Participants used a Likert scale to rate clinical importance and feasibility for each severity phenotype independently (informed by the exploratory analysis). For the optimal severity phenotype (highest combined score) for each condition, adjusted hazard ratios (aHR) of five-year mortality were calculated using Cox regression on the full CPRD database.

Results

Fifteen existing severity indexes for nine conditions informed the survey. Eighteen clinical academics participated in the survey, twelve also participated in the workshops. Combined mean scores for clinical importance and feasibility were highest for estimated glomerular filtration rate (eGFR) for chronic kidney disease (CKD) (9.42/10) and for microvascular complications of diabetes (9.08/10). Mortality was higher for each reduction in eGFR stage; Stage 3b aHR 1.42, 95 %CI 1.41–1.44 versus Stage 3a CKD and for each additional microvascular complication of diabetes; one complication aHR 1.44, 95 %CI 1.32–1.57 versus none. Some phenotypes (e.g., aneurysm diameter) were not well recorded within the database and could not feasibly be applied.

Conclusion

We developed a methodology for identifying severity phenotypes in EHRs. Severity phenotypes were identified for diabetes (type 1 and 2), ischaemic heart disease, CKD and peripheral vascular disease. Data quality in EHR should be improved for under-recorded severity measures.