Evaluation of data processing pipelines on real-world electronic health records data for the purpose of measuring patient similarity

BACKGROUND: The ever-growing size, breadth, and availability of patient data allows for a wide variety of clinical features to serve as inputs for phenotype discovery using cluster analysis. Data of mixed types in particular are not straightforward to combine into a single feature vector, and techniques used to address this can be biased towards certain data types in ways that are not immediately obvious or intended. In this context, the process of constructing clinically meaningful patient representations from complex datasets has not been systematically evaluated. AIMS: Our aim was to a) outline and b) implement an analytical framework to evaluate distinct methods of constructing patient representations from routine electronic health record data for the purpose of measuring patient similarity. We applied the analysis on a patient cohort diagnosed with chronic obstructive pulmonary disease. METHODS: Using data from the CALIBER data resource, we extracted clinically relevant features for a cohort of patients diagnosed with chronic obstructive pulmonary disease. We used four different data processing pipelines to construct lower dimensional patient representations from which we calculated patient similarity scores. We described the resulting representations, ranked the influence of each individual feature on patient similarity and evaluated the effect of different pipelines on clustering outcomes. Experts evaluated the resulting representations by rating the clinical relevance of similar patient suggestions with regard to a reference patient. RESULTS: Each of the four pipelines resulted in similarity scores primarily driven by a unique set of features. It was demonstrated that data transformations according to each pipeline prior to clustering can result in a variation of clustering results of over 40%. The most appropriate pipeline was selected on the basis of feature ranking and clinical expertise. There was moderate agreement between clinicians as measured by Cohen's kappa coefficient. CONCLUSIONS: Data transformation has downstream and unforeseen consequences in cluster analysis. Rather than viewing this process as a black box, we have shown ways to quantitatively and qualitatively evaluate and select the appropriate preprocessing pipeline