GLADIATER: An algorithm to support anomaly detection from a longitudinal cardiovascular and diabetes dataset

Jason Thomas Chew, Yakub Sebastian, Valliappan Raman, Xun Ting Tiong, Alan Yean Yip Fong, Patrick Hang Hui Then

Research output: Contribution to journalMeeting Abstract

Abstract

Objectives: To develop a new graph-based anomaly detection system for discovering unexpected Type 2 Diabetes Mellitus (T2DM) progression patterns from a longitudinal cohort study.

Methods: LIFECARE (Life Course Study in Cardiovascular Disease Epidemiology) study is a multinational longitudinal cohort study that aims at identifying cardiovascular risk factors among the South East Asian populations. LIFECARE has completed two phases. Part of this study was carried out in Sarawak, where the baseline cohort was initiated in 2010 and recruited 2,541 subjects. The second phase took place from 2013 to 2016 and successfully followed up 1,941 subjects.

A new software, GLADIATER, was developed to model the unexpected blood glucose level progression patterns between the two phases of LIFECARE study. Investigations into such cases are expected to provide fresh insights into the pathogenesis of T2DM. The algorithm models each study subject as a node in a graph. A link connects two or more nodes with homogeneous clinical/non-clinical profiles. Subsequently, groups of highly-connected homogeneous subjects are semi-automatically identified from the graph by employing a graph clustering algorithm. Using a Bayesian inference method, individuals whose blood glucose progression trajectory deviate from the expected trajectory of other individuals in the same homogenous group are flagged as being anomalous and subjected to further clinical investigations.

Results: We tested the system on 1,711 LIFECARE study subjects who satisfied pre-specified inclusion criteria. Variables studied included social demographics profiles, anthropomorphic measurements, laboratory biochemical results, physical activity, and dietary intake patterns. Eighty-six groups of homogeneous study subjects were found based on a combination of sociodemographic data (age, gender, income level, civil status, education level), dietary intake patterns (types of food and intake frequency), physical activities (occupational and non-occupational), and current medications.

An interesting cohort of 17 subjects with highly similar profiles were identified. Through the comparison of baseline and second phase data of the LIFECARE study, the majority of the subjects in this cohort were found with lowered or stagnant blood glucose levels. GLADIATER flagged two subjects who, despite sharing strongly similar characteristics with other subjects, experienced an opposite blood glucose level trajectory. Follow-up investigation revealed that one subject was morbidly obese, whereas the other had experienced a major family crisis. Obesity and stress-induced depressions might have thus contributed to the pathogenesis of T2DM in subjects who were otherwise might have experienced non-elevated glucose levels.

Conclusions:
Initial results obtained with GLADIATER appeared promising in isolating individuals with unexpected T2DM progression patterns. Further studies of such individuals could possibly suggest novel T2DM risk factors.
Original languageEnglish
Article numberGW29-e0830
Pages (from-to)C244-C245
Number of pages2
JournalJournal of the American College of Cardiology
Volume72
Issue number16 (Supplement)
DOIs
Publication statusPublished - Oct 2018
Externally publishedYes

Fingerprint Dive into the research topics of 'GLADIATER: An algorithm to support anomaly detection from a longitudinal cardiovascular and diabetes dataset'. Together they form a unique fingerprint.

Cite this