The harmonisation of stroke datasets: A case study of four UK datasets

Longitudinal studies of stroke patients play a critical part in developing stroke prognostic models. Stroke longitudinal studies are often limited by small sample sizes, poor recruitment, and high attrition levels. Some of these limitations can be addressed by harmonising and pooling data from existing studies. Thus this thesis evaluated the feasibility of harmonising and pooling secondary stroke datasets to investigate the factors associated with disability after stroke.

Data from the Clinical Information Management System for Stroke study (n=312), Stroke Outcome Study 1(n=448), Stroke Outcome Study 2 (n=585), and the Leeds Sentinel Stroke National Audit (n=350) were used in this research. The research conducted in this thesis consisted of four stages. The first stage used the Data Schema and Harmonisation Platform for Epidemiological Research (DataSHaPER) approach to evaluate the feasibility of harmonising and pooling the four datasets that were used in this case study. The second stage evaluated the utility of using multi-group-confirmatory-factor analysis for testing measurement invariance of the GHQ-28 measure prior to pooling the datasets. The third stage evaluated the utility of using Item Response Theory (IRT) models and regression- based methods for linking disability outcome measures. The last stage synthesised the harmonised datasets using multi-group latent class analysis and multi-level Poisson models to investigate the factors associated with disability post-stroke.

The main barrier encountered in pooling the four datasets was the heterogeneity in outcome measures. Pooling datasets was beneficial but there was a trade-off between increasing the sample size and losing important covariates. The findings from this present study suggested that the GHQ-28 measure was invariant across the SOS1 and SOS2 stroke cohorts, thus an integrative data analysis of the two SOS datasets was conducted. Harmonising measurement scales using IRT models and regression-based methods was effective for predicting group averages and not individual patient predictions. The analyses of harmonised datasets suggested an association of female gender with anxiety and depressive symptoms post-stroke.

This research concludes that harmonising and pooling data from multiple stroke studies was beneficial but there were challenges in measurement comparability. Continued efforts should be made to develop a Data Schema for stroke to facilitate data sharing in stroke rehabilitation research.