Poster Topical Area: Nutritional Epidemiology

Location: Hall D

Poster Board Number: 767

P20-042 - Distance metrics optimized for clustering temporal dietary patterning among U.S. adults

Monday, Jun 11
8:00 AM – 3:00 PM

Objective: Few attempts to determine dietary patterns have incorporated concepts of time, specifically time of energy intake, frequency of intake, and proportion of energy consumed throughout a day, despite the intrinsic nature of time to patterning. Modified dynamic time warping (MDTW) has been previously developed as an appropriate distance metric for patterning these aspects of dietary patterns to determine temporal dietary patterns (TDP) using the 1999-2004 National Health and Nutrition Examination Survey (NHANES) dataset. The objective of this study was to further explore DTW distance metrics ("unconstrained", "constrained" and MDTW) with modern spectral clustering methods to optimize the classification of TDP related to dietary quality using this dataset. The hypothesis was that MDTW would classify the TDP with the strongest relationships to dietary quality among U.S. adults 20-65y of NHANES 1999-2004.


Methods:
Proportional energy intake by time of day and frequency metrics were optimized from complete day-one 24-hour dietary recalls to create MDTW, conventional "unconstrained" DTW with only a standard local constraint, and "constrained" DTW with both a standard local and global banding constraint. All three distance metrics were clustered using spectral clustering. The association between each TDP distance metric clustering and mean dietary quality, as indicated by the 2005 Healthy Eating Index (HEI-2005), were determined using multiple linear regression controlled for potential confounders. Strength of association for each model was compared using adjusted R-squared.

Results: Four clusters representing distinct TDP for each distance metric by spectral clustering were generated among participants. MDTW exhibited TDP clusters with strong associations to HEI and the widest significant differences (p<0.0001) in HEI-2005 (at 35.7 points, 2.1 standard error and 51.9 points, 0.5 standard error) among clusters compared with the TDP clusters generated from unconstrained and constrained DTW.

Implication: MDTW paired with spectral clustering is an ideal method for integrating multiple aspects of time with dietary data to determine TDP.




Funding Source:

Purdue University

CoAuthors: Youngha Hwang, PhD – School of Electrical and Computer Engineering, Purdue University; Saul Gelfand, PhD – School of Electrical and Computer Engineering, Purdue University; Yanling Zhao, MS – Department of Nutrition Science, Purdue University; Anindya Bhadra, PhD – Department of Statistics, Purdue University; Edward Delp, PhD – School of Electrical and Computer Engineering, Purdue University

Heather A. Eicher-Miller

Assistant Professor
Department of Nutrition Science, Purdue University
West Lafayette, Indiana