Sunday, Aug 4: 4:00 PM - 5:50 PM
5015
Contributed Papers
Oregon Convention Center
Room: CC-E142
Main Sponsor
Biometrics Section
Presentations
In this study, we propose a novel tree-based localized functional principal component analysis method. Eigenfunctions estimated by the proposed method have compact local supports and can be interpreted as local features. We demonstrate the proposed method with an application to the electrocardiogram (ECG) data collected from the Chronic Renal Insufficiency Cohort (CRIC) study. The proposed method identified that delayed and decreased P wave, decreased amplitude of the Q and R wave and abnormal S wave, delayed onset of the T wave, and decreased T wave are associated with atrial fibrillation (AFib). A multivariable predictive model for AFib status using these local features is constructed with a C statistic of 0.771.
Keywords
functional data analysis
functional principal component analysis
electrocardiogram
tree-based method
Monitoring free-living physical activity (PA) can provide valuable insight into daily life activities. However, variations of PA, including the within-subject and between-subject variation, are often large and cause difficulty in analysis. In addition, due to its longitudinal characteristic, it is challenging to summarize and extract interpretable features. In this paper, we propose for such function data an elastic-based clustering algorithm for detecting specific changes in activity patterns. The process of this algorithm includes segmentation, data similarity computation, and pattern clustering. Using this clustering algorithm, we can obtain subject-specific and cluster-specific activity mean functions and perform association analysis to explore the relationship between physical activity and health outcome of interest. This algorithm can detect the phase and amplitude variation, reduce data dimension and offer interpretable findings. The proposed method is demonstrated on mental health studies. The results provide cluster-specific patterns for physical activities and can describe the patterns that are associated with the phy
Keywords
wearable device
functional data analysis
clustering
free-living physical activity
We consider an alternative basis expansion on functional sliced inverse regression that leads to a novel estimator for the functional central subspace. The estimator provides some improvements over conventional functional sliced inverse regression in terms of simplicity of implementation and recovery of less smooth effective directions. We provide some theoretical results, some numerical analyses and an application to the Chronic Renal Insufficiency Cohort study.
Keywords
Sliced Inverse Regression
Sufficient Dimension Reduction
Functional Data
Abstracts
Co-Author(s)
Wensheng Guo, University of Pennsylvania Perelman School of Medicine
Wei Yang, University of Pennsylvania
First Author
Harris Quach, University of Pennsylvania
Presenting Author
Harris Quach, University of Pennsylvania
Human microbiome data exhibit complex spatial structures. Understanding the spatial dependence structures can often enhance inference about microbes' functions. In this work, we propose a novel parsimonious multivariate spatial log Gaussian Cox process (LGCP) model using the concept of the linear model of regionalization, which can explicitly capture within-species and cross-species dependence structures and interactions. The model is inherently latent Gaussian, thus we adopt the integrated nested Laplace approximation-stochastic partial differential equations (INLA-SPDE) method to efficiently speed up the computation using an approximate Bayesian approach. We apply the model to study human oral microbiome biofilm image data from samples of multiple patients obtained using spectral imaging fluorescence in situ hybridization (FISH), where the spatial information of how taxa's cells are located relative to each other and to host cells are preserved.
Keywords
Multivariate Spatial LGCP
INLA-SPDE
Microbiome Image Data
Co-Author(s)
Suman Majumder, University of Missouri, Missouri, the United States
Brent Coull, Harvard T.H. Chan School of Public Health
Jessica Mark Welch, The Forsyth Institute
Patrick La Riviere, University of Chicago
Jacqueline Starr, Brigham and Women’s Hospital
Kyu Ha Lee, Harvard T.H. Chan School of Public Health
First Author
Yan Gong, Harvard T.H. Chan School of Public Health
Presenting Author
Yan Gong, Harvard T.H. Chan School of Public Health
Maps of canonical functional brain networks often guide our interpretation of spatial maps of brain-phenotype associations. However, methods for assessing enrichment of associations within networks of interest have varied in terms of both scientific rigor and underlying assumptions. While some approaches have relied on subjective interpretations, others have made unrealistic assumptions about the spatial structure of imaging data, leading to inflated false positive rates. We seek to address this gap in existing methodology by borrowing insight from Gene Set Enrichment Analysis (GSEA, Subramanian et al. 2005), a method widely used in genomics research for testing enrichment of associations between a set of genes and a phenotype of interest. We propose Network Enrichment Significance Testing (NEST), a flexible framework for testing the specificity of brain-phenotype associations to functional networks. We apply NEST to study associations involving structural and functional brain imaging data from a large-scale neurodevelopmental cohort study.
Keywords
enrichment
permutation testing
neuroimaging
brain networks
spatial data
Co-Author(s)
Simon Vandekar, Vanderbilt University
Aaron Alexander-Bloch, University of Pennsylvania
Armin Raznahan, National Institute of Mental Health
Mingyao Li, University of Pennsylvania, Perelman School of Medical
Raquel Gur, University of Pennsylvania
Ruben Gur, University of Pennsylvania
David Roalf, University of Pennsylvania
Min Tae M. Park, University of Toronto, McGill University
Mallar Chakravarty, McGill University
Erica Baller, University of Pennsylvania
Kristin Linn, University of Pennsylvania
Theodore Satterthwaite, Univ of Pennsylvania
Russell Shinohara, University of Pennsylvania
First Author
Sarah Weinstein, University of Pennsylvania
Presenting Author
Sarah Weinstein, University of Pennsylvania
An important objective in longitudinal analysis is to quantify the dynamic dependence structure between different outcome variables conditioning on a set of time-varying covariates. Existing nonparametric estimation methods do not take the dynamic dependence structure on the covariates into consideration. We propose a series of different approaches to estimate the time-varying conditional correlation functions based on kernel smoothing and structured nonparametric models for the conditional mean, variance and covariance functions, and construct their pointwise confidence intervals using a resampling-subject bootstrap procedure. We investigate the statistical properties of these smoothing estimators through a simulation study and apply these estimation and inference procedures to the Coronary Artery Risk Development in Young Adults (CARDIA) Study. Our findings suggest that the correlation of cardiovascular risk factors for young adults may change with age and other covariates.
Keywords
Functional correlation function
Conditional correlation
Nonparametric estimation
Varying coefficient model
Co-Author(s)
Hongbin Fang, Georgetown University
Xin Tian, NIH/NHLBI-Office of Biostatistics Research
Colin Wu, National Heart, Lung & Blood Institute, Office of Biostatistics Research
First Author
Haiou Li
Presenting Author
Haiou Li
Wearable devices are often used to monitor physical activity behavior to study its influences on health outcomes. These devices are worn over multiple days to record activity patterns resulting in multi-level longitudinal high dimensional or functional data. And excess zeroes may be recorded for non-moving periods or due to missing data. In addition, some recent work has demonstrated that the accuracy of the devices in monitoring physical activity patterns depend on the intensity of the activities and wear time. While work on adjusting for biases due to measurement errors in functional data is a growing field, less work has been done to study missing data patterns, measurement errors and their combined influences on estimation in functional linear regression models. In this work, we propose semicontinuous modeling approaches to adjust for biases due to missing data, zero-inflation, and measurement errors in functional linear regression models. We demonstrate the finite sample properties of our proposed methods through simulations. These methods are applied to a school-based intervention study of physical activity on age and sex adjusted BMI among elementary school aged children.
Keywords
Measurement error
Missing data
Zero-inflated functional covariate
Semicontinuous model
Physical activity data