Mackintosh, M;
Gonzalez-Izquierdo, A;
Rossor, M;
Whitaker, K;
Direk, K;
Denaxas, S;
(2019)
Network Structures and Dynamics Of Early Dementia Events Recorded in Primary Care Electronic Health Records.
In: Kostkova, P and Wood, C and Bosman, A and Grasso, F and Edelstein, M, (eds.)
DPH2019: Proceedings of the International Conference on Digital Public Health.
(pp. p. 127).
Association for Computing Machinery (ACM): New York, NY, USA.
Preview |
Text
Rossor_po08-mackintosh.pdf - Accepted Version Download (291kB) | Preview |
Abstract
Background: Identifying early signs of heterogeneous conditions like dementia is challenging. We used electronic health records (EHR) and data-driven methods in order to represent prodromal dementia as a dynamic temporal network of healthcare events and move beyond reductionist representations of complex syndromes. / Methods: We used UK primary care EHR data from Clinical Practice Research Datalink (CPRD) through the CALIBER resource and identified patients with a dementia diagnosis. We constructed a weighted, undirected network. We calculated centrality measures of the network and compared three community detection algorithms, Louvain, InfoMAP and Walktrap. Distinctive temporal communities of events in two year windows were derived, and we explored community membership, interactions and dynamics using evaluation measures including Jaccard distance, modularity, Rand Index and Normalised Mutual Information. / Results: We analysed data from 89,102 patients, where nodes (n=816) were connected to edges (n=392,345) based on the frequency with which two features were recorded in the same time window. Across the whole prodrome, repeated cardiovascular medications accounted for 22% of the network edges, followed by repeated central nervous system medications, accounting for 5% of edges. Cardiovascular conditions had the highest eigen centrality (influence on the entire network structure) and QRISK2 and malignant neoplasms had the greatest betweenness centrality (bridge between events). Louvin had the highest modularity and clustered the temporal network into six communities: The largest community was enriched for respiratory diseases (late-prodrome) and circulatory conditions (mid). Community 2 was enriched for musculoskeletal conditions (late), and Community 3 for administrative events (early) and nervous system conditions (late). In the early prodrome, there were more transitions between communities, however from 10 years to diagnosis, most events occurred within communities. / Conclusion: By understanding the interdependencies of conditions and associated medications across a disease network, we can deepen our understanding of prodromal dementia and create an accurate phenotype of the earliest stages of cognitive decline.
Archive Staff Only
 |
View Item |
