Characterisation, identification, clustering, and classification of disease

Abstract

The importance of quantifying the distribution and determinants of multimorbidity has prompted novel data-driven classifications of disease. Applications have included improved statistical power and refined prognoses for a range of respiratory, infectious, autoimmune, and neurological diseases, with studies using molecular information, age of disease incidence, and sequences of disease onset (“disease trajectories”) to classify disease clusters. Here we consider whether easily measured risk factors such as height and BMI can effectively characterise diseases in UK Biobank data, combining established statistical methods in new but rigorous ways to provide clinically relevant comparisons and clusters of disease. Over 400 common diseases were selected for analysis using clinical and epidemiological criteria, and conventional proportional hazards models were used to estimate associations with 12 established risk factors. Several diseases had strongly sex-dependent associations of disease risk with BMI. Importantly, a large proportion of diseases affecting both sexes could be identified by their risk factors, and equivalent diseases tended to cluster adjacently. These included 10 diseases presently classified as “Symptoms, signs, and abnormal clinical and laboratory findings, not elsewhere classified”. Many clusters are associated with a shared, known pathogenesis, others suggest likely but presently unconfirmed causes. The specificity of associations and shared pathogenesis of many clustered diseases provide a new perspective on the interactions between biological pathways, risk factors, and patterns of disease such as multimorbidity.

Statistical tests for heterogeneity of clusters and composite endpoints

“…An alternative clustering-based approach, is to assume that diseases are in one or more clusters with equal associations, and test if the log-likelihood for the model [32] is minimised by one, or more clusters. This objective test can incorporate a prior, and examples suggest it is more lenient than a Q 2 test, with the disease clustering data of Webster et al [1] minimising the log-likelihood when I 2 50%. The merits of this approach for applications such as meta-analyses, will need exploring in greater detail elsewhere.…”

“…The null hypothesis (in a fixed effects model), is that all diseases in a composite endpoint have the same associations with one or more parameters, such as a drug, or a collection of potential risk factors. These might be a subset of associations, with potential confounders adjusted for, and subsequently removed by marginalisation [1]. Consider m composite endpoints (or clusters of diseases), labelled by g. Under the null hypothesis of the same associations for diseases in a composite endpoint, labelled i = 1 to i = n g ,…”

“…For a flat prior, the sum is from i = 1 to i = n g . The subscripts g allow the discussion to include more than one cluster of diseases, as was considered in Webster et al [1], or a random effects model. Here unless stated otherwise, we consider a single composite endpoint, and the subscript g could be omitted.…”

“…It is common for epidemiological studies [1] and clinical trials [2][3][4] to group similar diseases together into a cluster of diseases, providing more total cases, and more statistical power to detect associations. This procedure has been essential since the pioneering epidemiological studies of John Graunt in the 1600s [5], providing sufficient cases to allow meaningful statistical study, while attempting to ensure that clustered diseases have a similar etiology.…”

“…The ICD system clusters increasingly detailed disease descriptions into larger clusters with similar disease etiology, often allowing larger clusters to be used in epidemiological studies. In practice, diseases are selected by a clinician to help ensure that only diseases with a clearly defined and common etiology are clustered together [1]. With datadriven clustering studies increasingly being used to identify potential composite endpoints [1,[7][8][9][10][11], it seems appropriate to review the assumptions needed to study a cluster of diseases, and the existing arguments for and against doing so in clinical trials.…”