Year: 2019
Type: article
Abstract: Electronic health records (EHR) offer higher diagnostic resolution and a breadth of clinically relevant features recorded over the lifecourse. Applying clustering methods in this large, representative, longitudinal data can potentially identify novel, clinically-actionable Alzheimer's Disease (AD) subtypes. The effect of the different clustering methods has not been systematically investigated. Here, we perform four clustering experiments using different methods, comparing the results internal and clinical validity. We identified AD patients using a previously-validated rule-based phenotyping algorithm based on diagnosis codes in longitudinal primary care EHR (CPRD) available in the CALIBER resource (Denaxas 2012). We identified relevant clinical features to include from literature and extracted them for each patient. Data were clustered using four methods: k-means, kernel k-means, affinity propagation and latent class analysis (LCA). Clusters were manually labelled by examining clinical feature prevalence. Internal validity was assessed using the mean silhouette score (measuring cluster separation). Clinical validity was assessed using healthcare utilization and progression markers such as rate of decline in MMSE score. We identified 10,065 AD patients (66% female, median age: 81).We identified and extracted four symptomatic variables, 22 comorbidities (such as depression, cancer), and four lifestyle risk factors. The number of optimal clusters varied from 2 (kernel k-means) to 6 (LCA). The highest silhouette score of 0.19 was achieved for five clusters using the k-means algorithm. A common cluster was found in K-means, Affinity propagation and LCA. It had the earliest diagnosis age, and the highest prevalence of depression, anxiety, and current smokers and drinkers. It represented 13.7% of patients that are consistently grouped together. This group had significantly faster progression determined by examining MMSE scores. As there is no ground truth available when using clustering methods the fact that this cluster has appeared in 3/4 methodologies indicates that this is a distinct group of interest rather than an artefact of the clustering methods. Further research should be conducted into how this group of variables relates to faster progression and also possible interventions. This group has clinical significance due to the complications of treating AD patients with additional mental health diagnoses.
Source: Alzheimer s & Dementia
Language: English
Subfield: Psychiatry and Mental health
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
Funder: Medical Research Council
Award: G6507636021
OpenAlex ID: W2980854764