The Use of a Bayesian Hierarchy to Develop and Validate a Co-Morbidity Score to Predict Mortality for Linked Primary and Secondary Care Data from the NHS in England

BACKGROUND We have assessed whether the linkage between routine primary and secondary care records provided an opportunity to develop an improved population based co-morbidity score with the combined information on co-morbidities from both health care settings. METHODS We extracted all people older than 20 years at the start of 2005 within the linkage between the Hospital Episodes Statistics, Clinical Practice Research Datalink, and Office for National Statistics death register in England. A random 50% sample was used to identify relevant diagnostic codes using a Bayesian hierarchy to share information between similar Read and ICD 10 code groupings. Internal validation of the score was performed in the remaining 50% and discrimination was assessed using Harrell's C statistic. Comparisons were made over time, age, and consultation rate with the Charlson and Elixhauser indexes. RESULTS 657,264 people were followed up from the 1st January 2005. 98 groupings of codes were derived from the Bayesian hierarchy, and 37 had an adjusted weighting of greater than zero in the Cox proportional hazards model. 11 of these groupings had a different weighting dependent on whether they were coded from hospital or primary care. The C statistic reduced from 0.88 (95% confidence interval 0.88-0.88) in the first year of follow up, to 0.85 (0.85-0.85) including all 5 years. When we stratified the linked score by consultation rate the association with mortality remained consistent, but there was a significant interaction with age, with improved discrimination and fit in those under 50 years old (C = 0.85, 0.83-0.87) compared to the Charlson (C = 0.79, 0.77-0.82) or Elixhauser index (C = 0.81, 0.79-0.83). CONCLUSIONS The use of linked population based primary and secondary care data developed a co-morbidity score that had improved discrimination, particularly in younger age groups, and had a greater effect when adjusting for co-morbidity than existing scores.