COMORBIDITY CLUSTERS IN CLINICAL CONDITIONS: AN ANALYSIS OF ELECTRONIC HEALTH RECORD DATA
HE-THESIS-2019.pdf (830.9Kb) (embargoed until: 2021-08-01)
MetadataShow full item record
Background and Objective: Comorbidity is defined as other conditions present alongside a major condition at the same time. Knowledge of existing comorbidities in study participants may help to guide their disease assessment and management. The objectives of this research were to understand comorbidity patterns and to assess the performance of a range of clustering methods applied to study participant comorbidity profiles in order to stratify study participant’s by disease severity. Methods: We selected study participants who had diabetic retinopathy, glaucoma, or chronic kidney disease from an Electronic MEdical Records & GEnomics dataset (our data source). Then we then created a “gold standard” categorization of study participants into disease severity groups using International Classification of Diseases, Ninth Revision, Clinical Modification codes (i.e., mild, moderate or severe disease). After that, we applied K-means, hierarchical and spectral clustering methods to see how well each performed to classify study participants into the correct severity group, considering two different data subsets. The first data subset considered all “EDC” diagnostic categories and the second data subset only considered selected EDCs that were considered relevant to the conditions. Results: Our results show that there are no significant differences in the number of comorbidities among different severity levels for diabetic retinopathy (p = 0.8261) and glaucoma (p = 0.5748). However, there was a statistical difference among severity levels for chronic kidney disease (p = 0.0008). Also, we found that for diabetic retinopathy study participants, when using K-means and spectral clustering methods and taking all EDCs disease categories into consideration, it is possible to stratify study participants into three groups based on diagnostic category clustering which corresponds to their severity. Conclusions: We found a statistical difference in the number of comorbidities present among patients categorized into different severity group for one condition (chronic kidney disease). But there are no significant differences in the number of comorbid conditions among different severity levels for diabetic retinopathy (p = 0.8261) and glaucoma (p = 0.5748). However, there is statistical difference among severity levels for chronic kidney disease (p = 0.0008497). When applying clustering approaches to all EDCs of study participants, we found that, two clustering approaches (K-means and spectral clustering) could be used to classify study participants with diabetic retinopathy into the correct severity group. Clustering approaches were not successful for other scenarios we explored.There were some limitations to this work due to a reliance on administrative data to categorize study participants into severity groups. Findings from this work, however, are promising start to exploring machine learning approaches to identify the severity of disease.