22691 research outputs found
Sort by
Understanding the Evolution of Diet Quality in U.S. Infants and Toddlers from Households with Low-Income Utilizing Diet Quality Indices
No full textBackground
Diet in early life sets the stage for later eating habits. The 2020-2025 Dietary Guidelines for Americans included infants and toddlers for the first time, making it possible to measure early life diet quality compared to recommendations.
Objectives
To improve our understanding of how diet quality initiates in infancy and tracks into early toddlerhood in children from households with low-income.
Methods
Using 24-hour diet recall and questionnaire data from the Special Supplemental Nutrition Program for Women, Infants, and Children (WIC) Infant and Toddlers Feeding Practices Study-2, we characterized diet quality in infants (year 1) using the Infant Diet Quality Index (IDQI) score and the Diet Quality Index Score (DQIS), and in toddlers (year 2) using the Healthy Eating Index-Toddlers-2020 (HEI-Toddlers-2020). Linear mixed effects models evaluated change in HEI-Toddlers-2020 score over time and the association between infant and toddler diet quality. Subgroups of diet quality trajectory from 12-24 months were investigated with a parallel process growth mixture model of refined grains, sodium, saturated fat, and added sugars. Early life factor association with class membership was investigated.
Results
Mean total HEI-Toddlers-2020 scores ranged from 56.3 to 58.1 across year 2. Scores were lower with age comparing 18- and 24-month scores to 13-month scores ( = -1.83, 95% CI: -2.88, -0.78; and = -1.54, 95% CI: -2.65, -0.45, respectively). Total Vegetables, Refined Grains, Sodium, and Added Sugars decreased. Almost two-thirds of 13-month toddlers consumed no Greens/Beans or Seafood/Plant Protein. There were positive associations between the IDQI ( = 1.74, 95% CI 1.48-1.99) and the DQIS ( = 0.43, 95% CI 0.33-0.52) and the HEI-Toddlers-2020 score. Three subgroups of moderation component trajectory were identified. Class membership was significantly associated with caregiver age, but not duration of human milk or solids introduction. Sugar-sweetened beverages may displace milk for some children.
Conclusions
Diet quality in infancy is positively associated with toddler diet quality but begins to decline by year 2. Declines were driven by increased refined grains, sodium, and added sugars. These findings provide evidence that interventions targeting reduction in moderation foods may have the biggest impact on overall child diet quality
Home Visiting Reach and Engagement of Pregnant Women Who Screen Positive for Substance Use Risk
Full text linkBackground: Evidence-based home visiting (EBHV) is a strategy for supporting expectant families and families with young children to promote healthy family functioning, positive parenting, and child health and development. EBHV has the potential to serve many families with substance use issues. However, reaching and engaging these families presents unique challenges. Understanding how EBHV programs reach and engage families with substance use issues is key to fulfilling its potential to support such families by meeting their needs and improving family outcomes. This multimethod dissertation investigates the extent to which EBHV services in New Jersey reach and engage pregnant women who screen positive for substance use risk.
Methods: Aims 1 and 2 quantitatively examined differences in reach and engagement indicators by substance use risk status using multilevel multivariate logistic regression models fit on data from the statewide Central Intake system. Aim 3 used reflexive thematic analysis to qualitatively explore EBHV engagement experiences among 11 women identified as positive for substance use risk by their home visitors.
Results: Aim 1 results indicate Central Intake was overall more likely to attempt to contact and refer women to EBHV who screened positive for substance use risk prenatally than those who screened negative. There were no overall significant differences in contact success or enrollment by substance use risk. However, interaction analyses highlight how housing stability, race/ethnicity, parenting experience, and social support modify the association of substance use risk with reach indicators. In Aim 2, there was not a statistically significant difference in receipt of a high dose of services by substance use risk status. Aim 3 results highlight the importance of trusted referral sources, tailored provision of functional supports both related and unrelated to substance use recovery, and trusting relationships with home visitors as key to participants’ engagement.
Discussion: This dissertation’s findings suggest that New Jersey EBHV’s efforts to reach and engage women with substance use issues were successful in some areas, while identifying areas for improvement. Grounded in the Home Visiting Precision Paradigm, discussion focuses on program design and implementation strategies to improve reach and engagement and ultimately to strengthen program effectiveness
Further Design and Development of Pendant Photochromic Molecules and Polymers
No full textDiarylethenes (DAEs) are an attractive class of compounds which undergo reversible photochromism, often called molecular switches. Photochromism is the reversible change between two colored states, stimulated by a light source. The switches are of particular interest for photoresponsive materials. To develop such materials, others and our group have tried incorporating these switches into polymers.
Chapter 1 is a literature review of DAE photoswitches and the work done in the Tovar lab. Several different graduate students have investigated pendant photochromic switches, where the DAE is part of fused ring system with a polymerizable “head” and a photoswitching “tail”. Photochromism for many of these systems was mixed, but not fully investigated. This chapter uses computational tools to determine whether other switching motifs on these systems would exhibit photochromism.
Chapter 2 summarizes an outstanding question from our group’s investigation into a thieno[3-4b]thiophene (TT) photoswitch. Previous researchers found that photochromism can be deactivated, due to other available absorption pathways. We envisioned a series of TT-based photochromes with various sterically demanding aromatic caps. Guided by computations, these compounds were synthesized. These model systems provide a better understanding of the behavior of photochromic units within extended oligomeric and polymeric π-conjugated materials.
Chapter 3 describes efforts to incorporate TT switches into responsive materials. First TT switches were functionalized and electropolymerized at various monomer compositions. Preliminary results suggest there is an electrochromic response, though it has yet to be determined whether this deactivates the photochromism of the polymer. In a similar vein the TT switches were functionalized to form metal organic frameworks (MOFs). Since the switching motif is segregated from the polymer head in the pendant TT switch, the photochromism will not perturb the MOF geometry. With the Thoi group at JHU, several MOFs were formed, though photochromism was not observed.
Chapter 4 is an investigation into a different DAE photoswitch, benzo[b]furan (BF). Due to the “L” shaped geometry the BF switch, the DAE switching motif and polymer backbone in closer proximity, allowing for further probing of those interactions. Several switches were envisioned, computed, and were indicative of photochromism. These BF switches and phenyl capped BF switches were then synthesized and were all photochromic. The phenyl capped BF switches serve as models for BF-polymers, indicating that aromatic caps do not deactivate photochromism in these switches, making it a promising switch architecture.
Finally, Chapter 5 summarizes future directions for the photochromic switch project. It highlights previously investigated switches that are worth revisiting and proposes several improved syntheses. Several applications of pendant DAEs are summarized. Particularly the idea of using TT switches as a responsive dye for colorimetric assays. Other novel switches and their syntheses are proposed in this chapter. Finally, this chapter summarizes possible analytical tools which can be used to determine kinetics of photochromism for all the proposed switches
METHODS FOR CAUSAL INFERENCE USING EXPERIMENTAL AND OBSERVATIONAL DATA
Full text linkThe recent development and adoption of electronic healthcare record systems has provided researchers a rich and detailed of data for researchers interested in studying aspects of clinical care. These observational datasets are well suited to supporting exploratory analyses and answering causal questions quickly. At the same time, randomized trials are still the gold standard for assessing causality. However, these interventional datasets are expensive to collect and are often collected with a single question in mind. Historically, methods have been developed to address the analysis of either observational or interventional data, but more recently there have been moves to analyze them jointly, or to use ideas and techniques developed for one context to create new methods in the other.
In this dissertation, I explore methods for causal inference that lie at the intersection of observational and interventional data. First, I provide novel algorithms for data fusion, an endeavor that seeks to combine various observational and interventional datasets to identify causal estimands. I provide improvements over existing algorithms by allowing datasets which have
unobserved variables or selection bias. I also introduce a framework of systematic selection in data
fusion, which allows for the fact that patients might enter an observational or interventional dataset not completely at random. Next, I present a method motivated by a randomized trial of drug
abuse therapies, in which patients often fail to report for testing. I propose a model for this behavior which is computationally tractable, and derive identification the relevant causal estimands using technology originally developed for observational datasets. Finally, I introduce a package for graphical causal inference, Ananke. Named for the Greek goddess of necessity, Ananke provides a suite of tools for interested analysts to more easily conduct end-to-end analyses by providing implementations of various algorithms in causal inference, including
algorithms in this dissertation
MECHANISMS UNDERLYING THE REGIONAL PREDISPOSITION TO AORTIC ANEURSYM IN LOEYS-DIETZ SYNDROME
No full textThoracic aortic aneurysms (TAA) are localized dilations of the aorta that predispose patients to life-threatening tears of the vessel wall. Human genetic studies have uncovered pathogenic variants that cause hereditary forms of TAA, highlighting the importance of the extracellular matrix, mechanosensing pathways, and transforming growth factor-β (TGF-β) signaling in aortic homeostasis. Loeys-Dietz syndrome (LDS) is a connective tissue disorder caused by mutations that impair TGF-β signaling, through partial inactivation of ligands, receptors, and signaling mediators.
Although LDS-causing mutations predispose individuals to aggressive aneurysms throughout the arterial tree, specific arterial regions, such as the aortic root, are especially vulnerable. Sites of severe aortic dilation are characterized by paradoxical secondary upregulation of TGF-β signaling. The work described in this dissertation investigates the dynamic responses to impaired TGF-β signaling, how it is modified by regional factors differentially expressed in specific aortic segments, and how these regional differences may explain the localized vulnerability to LDS-driven aortic pathology.
Chapter 1 provides an overview of the complex pathogenesis, genetic basis, and regional vulnerability of TAA. Chapter 2 delineates the primary and secondary transcriptional response of aortic smooth muscle cells to time-controlled, postnatal TGF-β inhibition. This analysis shows that while TGF-β inhibition causes broad downregulation of transcripts coding for extracellular matrix and focal adhesion components in smooth muscle cells, differential secondary responses, including more pronounced upregulation of these transcripts and TGF-β ligands, can be detected in specific subpopulations located in the aortic root. Chapter 3 examines the contributions of angiotensin II signaling to aortic dilation in an LDS murine model, revealing the differential effects of regional and systemic inactivation. Chapter 4 investigates the heterogeneity of smooth muscle cells in human and mouse aortas, identifying a subset of Gata4-expressing smooth muscle cells in the aortic root. Postnatal Gata4 deletion reduced aortic root dilation in LDS mice, indicating that Gata4 sensitizes the aortic root to impaired TGF-β signaling. Chapter 5 provides commentary on limitations and future directions. This dissertation highlights the importance of elucidating factors influencing regional aneurysm risk, providing insights into adaptive and maladaptive pathways for therapeutic interventions targeting specific arterial segments
Therapeutic Myeloid Cell Targets in the Immune Microenvironment of Bone and Soft Tissue Sarcomas
No full textCancer immunotherapy has become a pillar of treatment, but remains ineffective for many malignancies. Pediatric sarcomas have overall responded poorly to T cell-targeted immunotherapy, due to multifactorial challenges including a heavily immunosuppressed tumor microenvironment. Despite the failure of immunotherapy to-date, these approaches overall are an attractive alternative to cytotoxic chemotherapy, as their long-term side effects tend to be fewer and less severe compared to the risks to growth, development, fertility, and long-term cancer risk of traditional chemotherapeutic agents. Outcomes for pediatric sarcoma patients have plateaued in the last decades and new therapy approaches are desperately needed for these children.
This body of work addresses the potential of therapeutically targeting immunosuppressive innate immune cells in the microenvironment of pediatric sarcomas in combination with cytotoxic chemotherapy to improve treatment outcomes for patients affected by these cancers. We first assess the effects of delivering ADU-S100, a stimulator of interferon genes (STING) pathway agonist, on remodeling the immune microenvironment of osteosarcoma in syngeneic mouse models via analysis of gene expression, chemokine/cytokine production, and immune cell proportions in the tumor microenvironment. We demonstrate that ADU-S100 remodels the immunosuppressive macrophages in the tumor microenvironment to allow longer-term improved infiltration with cytotoxic T lymphocytes. When combined with carboplatin chemotherapy, ADU-S100 is able to induce complete tumor regression and long-term relapse-free survival in survival studies of syngeneic murine osteosarcoma models. This response is lymphocyte-dependent and dependent on STING pathway activation in the immune microenvironment.
We also identify novel innate immune targets in the microenvironment of rhabdomyosarcoma, another canonically immunosuppressed tumor. We use spatial transcriptome sequencing to assess gene expression in fusion-positive and fusion-negative murine rhabdomyosarcoma tumors. By examining gene expression in the tumor and at the tumor border with normal tissue, we identify three innate immune genes – Mif, Ccl8, and Cxcl14 – that are differentially expressed in these regions and are potential therapeutic targets to alter the immunosuppressed microenvironment of these tumors.
This work demonstrates the therapeutic power of targeting immunosuppressive myeloid cells in the microenvironment of pediatric sarcomas and proposes novel targets to manipulate these cells. Our results suggest that combination immunotherapy approaches are essential to the success of immunotherapy in some immune-excluded solid tumors and provides a foundation to design novel combination therapy approaches for pediatric sarcomas
Symmetry in Graph Learning
No full textGraphs are ubiquitous representations to capture complex interactions and structural patterns. Graph neural networks have emerged as popular machine learning tools to learn and reason with graph data. Despite their promise and some success stories, graph neural networks have not managed to absolutely outperform classical methods such as spectral embeddings and graph kernels. Moreover, it is not completely clear when and why graph neural networks work or fail.
In this dissertation, we develop principled graph neural networks by exploiting symmetries. Symmetries appear naturally in graph learning, such as permutation symmetry due to the choice of node labeling, or graph automorphisms arising from self-similarity. Leveraging symmetry can offer elegant descriptions of physical objects and encourage abstract reasoning -- a key feature of intelligence. Therefore, we enforce symmetry in graph neural networks to ensure their performance and rapid generalization to novel situations. Furthermore, symmetry is a fundamental concept in mathematics, spanning across geometry, algebra, and probability. Thus, we utilize rich mathematical insights of symmetry to understand and improve graph neural networks.
Using the notion of permutation symmetry, we study the expressivity of graph neural networks, leading to stronger architectures by incorporating graph spectral invariants or faster algorithms by using random graph embeddings. Beyond permutation symmetry, we analyze the generalization properties of equivariant graph networks when choosing different kinds of natural symmetries induced from the graphs. Leveraging probabilistic symmetry, we evaluate graph neural networks based on novel random graph models arising from joint exchangeability. Empirically, we validate our theoretical insights in numerous graph learning applications across social science, chemistry, and computer vision. We conclude by discussing other notions of symmetries and future research directions that exploit symmetry within and beyond graph learning
EXAMINING FEDERAL FOOD ASSISTANCE PROGRAM PARTICIPATION AS A MEANS OF IMPROVING DIET-RELATED HEALTH EQUITY FOR INDIGENOUS POPULATIONS IN THE UNITED STATES
Full text linkBackground
Indigenous foodways in the United States have been the target of genocidal actions and policies for five centuries, contributing to diet-related health inequities observed today. Though federal food assistance has an important role to play in improving food access and health equity, these programs don’t always meet the needs of Indigenous communities.
Objectives
The objectives of this dissertation are 1) to examine the history of genocide that begot diet-related health injustice; 2) to assess the extent to which accessibility and accessibility of federal food assistance programs have been evaluated among Indigenous Peoples in the U.S.; 3) to describe recent participation changes in tribally-administered WIC programs compared to regional and national changes; and 4) to characterize the system of factors impacting participation in two tribally-administered WIC programs.
Methods
A range of methodologies and approaches were used to address these objectives, including a scoping review, descriptive statistics from USDA WIC data tables, and use of qualitative stakeholder interviews to build causal loop diagrams.
Results
This work highlights a scarcity of available literature on the food assistance participation experiences and trajectories of Indigenous populations in the U.S. Focusing on tribally-administered WIC programs, we find that many experienced remarkably different changes in participation compared to regional and national averages during the years surrounding the COVID-19 pandemic. Causal loop diagrams created to elucidate factors impacting participation in tribally-administered WIC programs suggest that improvements to rural reservation infrastructure and reformation of bureaucratic procedures are necessary to ensure more equitable access to WIC.
Conclusions
If we are to leverage the potential impacts of food assistance on health equity, we must lay the groundwork by allowing Indigenous populations to be seen and counted and by striving to understand and address issues around program accessibility and acceptability
On the Security and Adaptation of Neural Networks: A Study of Adversarial Robustness, Backdoor Attacks, and Transfer Learning
Full text linkNeural networks have become the heart of modern AI systems, playing a crucial role in both commercial and critical applications. However, these models face two significant challenges: security vulnerabilities and adaptation to new test data and tasks. These models are vulnerable to various adversarial attacks, including evasion, data poisoning, and backdoor attacks, which can compromise their performance and reliability. On the other hand, when these models are deployed as prebuilt solutions, pretrained on tasks different from their intended downstream applications, leading to inferior performance due to domain gaps. Motivated by these challenges, this dissertation investigates the areas of adversarial robustness, backdoor attacks, and transfer learning for neural networks.
We first discuss how to reconstruct adversarial perturbations and classify these reconstructed perturbations based on the algorithm that generated them. This pipeline, REDRL, can detect the attack algorithm used to generate a sample from only the sample itself. We then present a new hidden trigger backdoor attack, Sleeper Agent, which leverages gradient matching, data selection, and target model re-training to craft highly effective poisons. Sleeper Agent is the first hidden trigger backdoor attack to be effective against neural networks trained from scratch. We demonstrate its effectiveness on ImageNet and in black-box settings. Next, we use guided diffusion to synthesize base samples from scratch, utilized for creating poisons and backdoors that are significantly more potent than previous state-of-the-art attacks. Our Guided Diffusion Poisoning (GDP) base samples can be combined with any downstream poisoning or backdoor attack to boost its effectiveness.
In addressing the adaptation challenge, we propose a novel approach within the transfer learning paradigm, where highly informative posteriors are learned from the source task, either through supervised or self-supervised methods, and used as priors for the downstream task. Finally, we introduce Battle of the Backbones (BoB), a comprehensive benchmark that evaluates various popular pretrained checkpoints and randomly initialized baselines across a wide range of downstream tasks, including image classification, object detection, segmentation, out-of-distribution generalization, and image retrieval. BoB sheds light on promising directions for the research community to advance computer vision by illuminating strengths and weaknesses of existing approaches
Robot Learning with Efficient Representation and Domain Adaptation
Full text linkRobot learning, integrating techniques from machine learning, computer vision, and robotics, has emerged as a promising field for developing robotic systems capable of performing a wide range of real-world tasks. In building successful learning-based system in robotics as in other domains, data plays a pivotal role. Despite significant achievement in collecting and training on large-scale robotic datasets, the real-world data acquired with physical hardware remains limited, particularly when compared to vast internet-scale image and text datasets. Consequently, a primary challenge for the robot learning community is constructing data-efficient systems that learn from scarce real-world data for deployment on physical hardware. This is particularly demanding for vision-based manipulation problems, where observations are high-dimensional, and tasks are complex.
This thesis addresses this challenge through two main themes: first, employing efficient visual representations to enhance learning efficiency; and second, leveraging abundant, low-cost simulation data for policy learning, followed by domain adaptation using minimal real-world data.
The first part of this thesis presents methods utilizing efficient visual representations to improve learning efficiency. In the "virtual in-hand eye transformer", we propose using virtual in-hand views instead of raw camera views, significantly enhancing performance when learning from a small number of demonstrations. Through self-supervision, in "proportional derivative controllable embedding", we learn embeddings from raw images that can be controlled with a simple proportional derivative-controller; in "keypoint-conditioned neural radiance field", we discover a set of keypoints of the scene that can be efficiently used by model predictive controller.
In the second part of this thesis, we study the problem of visual domain adaptation for sim-to-real transfer. Directly applying policies learned from simulation to real world would result in deteriorated performance because of the differences in appearances and physics. To bridge such domain gap between simulated and real environments, we propose leveraging exploratory experiences in the deployment environment to bridge the domain gap under domain adaptation settings.
To summarize, we propose multiple methods that learn visual policy with limited deployment domain data. Extensive experiments, both simulated and with real physical systems, validate our methods' superiorit