Development and validation of the Cambridge Multimorbidity Score

Problem

Multiple long-term health conditions are increasingly common and significantly affect health and well-being. Multimorbidity places considerable pressures on primary and secondary care, yet services and policies have failed to respond to these demands. There is a need to develop new measures of multimorbidity that could be used in future studies of multimorbidity and for planning health services and resource allocation.

Approach

Full clinical data were extracted from three samples of the UK’s Clinical Practice Research Database (CPRD): a development sample of 300,000 and two validation samples of 150,000 patients each. Codes were derived for 37 commonly-coded conditions and related to three outcomes: primary care consultations, unplanned hospital admissions and mortality. Outcomes were measured at one and five years. Additional analyses estimated the added benefit of including prescribing information as a predictor, constructed a general-outcome multimorbidity score by averaging the standardised weights of the separate outcome scores, and compared our models with the Charlson co-morbidity index.

Findings

Models including all 37 conditions were good predictors of GP consultations, emergency hospital admissions and mortality at one year (C-indices 0.732, 0.742 and 0.912 respectively, adjusted for age and gender). Reducing the models to the 20 conditions which had the greatest combined prevalence/weight made little difference to the predictive value of the models (C-indices 0.727, 0.738 and 0.910 respectively). Adding data on prescribing patterns made little difference to the predictive power of the models. Prediction of outcomes at five years for the 20-condition model remained good for consultations and mortality admissions (C-indices 0.735, 0.889) but performed less well for unplanned admissions (C-index 0.708). A 20-condition general-outcome score performed similarly to the outcome-specific models (C-indices 0.723, 0.735 and 0.913). The models performed substantially better than models based on conditions in the Charlson index.

Consequences

We have developed several robust, simple-to-use, multimorbidity scores, both tailored and not tailored to specific health outcomes. These scores have the potential to be of considerable value to clinicians and policymakers alike, providing a common-sense, transparent, easy-to-implement and effective means of optimising the delivery of healthcare to an ageing and increasingly multimorbid population.