Evaluating the harmonisation potential of diverse cohort datasets

Variable selection

A set of 124 variables optimised for neurodegeneration was identified by consensus within the ADDI Data Harmonisation group. Variables were selected to reflect the frequency of being requested in Dementias Platform UK (DPUK) data access proposals [1], to cover a range of data modalities, and to include modifiable and non-modifiable risk factors.

Standardisation

Datasets were curated to a common structure and labelling conventions using C-Surv as the data model [14]. C-Surv is a simple four level acyclic taxonomy intended to capture the breadth of data typically collected in research cohorts. The tiered structure supports grouped and individual variable discovery. C-Surv comprises 18 data themes (level 1) leading to > 146 data ‘domains’ (level 2), > 500 data ‘families’ (level 3) and then to a growing number of data ‘objects’ (level 4) i.e. the measured variable. C-Surv has been adopted by DPUK [1], Dementias Platform Australia [3], and the ADDI workbench [4]. Other models, developed for other purposes were available, such as the Observational Health Data Sciences and Informatics (OHDSI) OMOP Common Data Model for administrative health data [15], and CDISC Clinical Data Interchange Standards Consortium (CDISC) for trials data [16], but these have structural and semantic complexity that is alien to the cohort study design.

Schema development

The harmonisation schema was optimised to be inclusive of datasets by using relatively simple harmonisation rules and widely used value-labelling conventions. Three strategies for harmonisation, as described in the Maelstrom harmonisation guidelines [17] were used.

Simple calibration, using direct mapping between the source variable and the harmonised variable, was adopted for widely used standard metrics such as weight or height. Direct mapping, including cut-off points was used for validated clinical scales. The Gregorian calendar was used for dates and conventional units were used for age (years), durations (hours), concentrations (mg/ml), volumes (mm3), etc.

Algorithmic transformation was used for non-clinical questionnaire responses including lifestyle. The algorithm was selected to be inclusive by using a relatively simple transformation and was developed iteratively as it was applied to each dataset. Gender was transformed as male, female; smoking as ‘current, past, and never’, and ethnicity as white, black, Asian, mixed, other. Cohabitation was coded as single, married/cohabiting, divorced/separated, widowed, whilst education was considered as educational experience and transformed into junior or less, secondary, degree or equivalent, postgraduate or equivalent. For type of accommodation a straightforward transformation was house/bungalow, apartment, sheltered/residential, other.

Non-clinical cognitive performance scores were standardised into z-scores by default, with an option for refining this rule on a scale-by-scale basis according to the variable distribution. More sophisticated methods such as latent variable modelling or multiple imputation were not used.

Schema evaluation

The utility of the harmonisation rules was tested using four DPUK collaborating cohorts. These were selected on the basis of having diverse primary scientific objectives, providing longitudinal multimodal data, and being frequently requested by DPUK users. The cohorts were the Airwave Health Monitoring Study (Airwave); an occupational cohort [18], the English Longitudinal Study of Ageing (ELSA); a social science focussed study [19], Generation Scotland; a genetics cohort [20], and Memento; a neurodegeneration cohort [21]. The coverage of each cohort and overlap of variables across cohorts was assessed, along with the utility of the harmonisation rules. All analyses were conducted within the DPUK Data Portal [22].

留言 (0)

沒有登入
gif