22689 research outputs found
Sort by
OFFLINE REINFORCEMENT LEARNING FOR MDPS AND CONFOUNDED POMDPS: POLICY GRADIENT ALGORITHMS AND PESSIMISTIC MODEL-BASED APPROACHES
No full textThis dissertation aims to develop new offline reinforcement learning (RL) methods with theoretical guarantees. The first contribution is a model-based offline RL algorithm under the framework of Markov decision processes (MDPs). Existing approaches either emphasize theoretical aspects without practical implementations or restrict the policy space, limiting the potential of model-based methods. To address this, we propose MoMA, a model-based mirror ascent algorithm with general function approximations under partial offline data coverage. Unlike previous methods, MoMA utilizes an unrestricted policy class. We provide theoretical bounds on the performance gap between the output and optimal policies, supported by numerical experiments demonstrating the algorithm's efficacy.
The second contribution is a computationally efficient offline algorithm for partially observable Markov decision processes (POMDPs). Most existing works on offline RL were developed for the fully observable environment with Markovian transition dynamics, which may not known a priori, and little has been done for POMDPs. To develop a computationally practical algorithm for POMDPs, we propose a first policy gradient method in the offline setting for POMDPs. Our contributions include: 1) a novel identification for directly estimating the gradient of any (history-dependent) policy using the offline data, circumventing the issue of partial observability, 2) a min-max estimation procedure to compute the policy gradient non-parametrically, with an investigation on the statistical error, and 3) a proof of last-iterate global convergence of the proposed algorithm.
Lastly, we introduce a model-based offline RL method for confounded POMDPs under general function approximations, which is provably efficient under the assumption of partial coverage imposed on the offline dataset. Specifically, we first establish a novel model-based identification result to learn the impact of actions on rewards and transitions in confounded POMDPs. Using this, we design a nonparametric two-stage estimation procedure to construct estimators for policy values which permits general function approximations. Finally, we learn the optimal policy by performing a conservative policy optimization within the confidence regions which are constructed based on the proposed estimation procedure. Under some mild conditions, we establish a finite-sample upper bound on the suboptimality of the learned policy
THE ROLE OF TDP-43 IN NUCLEAR RNA PROCESSING AND EXOSOME-TARGETED DEGRADATION IN NEURODEGENERATION
No full textThis study investigates the role of TDP-43 in nuclear RNA processing and exosome-targeted degradation in the context of neurodegeneration. We demonstrate that depletion of TDP-43 in human neuronal cells leads to significant accumulation of LINE1 RNA, which results from impaired regulation of RNA stability rather than altered transcription. Mechanistically, we provide the first evidence that TDP-43 directly interacts with ZCCHC8, a core component of the NEXT complex, and that this association is maintained independently of RNA. Using RNA immunoprecipitation and loss-of-function assays, we show that ZCCHC8 physically associates with LINE1 transcripts and acts as a negative regulator of their abundance. Importantly, our analyses of ALS/FTD patient tissues and iPSN models reveal that ZCCHC8 expression is consistently reduced in disease states, suggesting a clinically relevant impairment of the nuclear RNA surveillance pathway. Furthermore, knockdown of ZCCHC8 enhances cryptic exon inclusion beyond levels induced by TDP-43 depletion alone, revealing cooperative regulation of aberrant transcripts. These results indicate that TDP-43 and the NEXT complex cooperatively repress aberrant transcripts, thereby preserving neuronal RNA homeostasis. These findings establish a novel functional relationship between TDP-43 and the NEXT complex in maintaining neuronal RNA homeostasis and implicate disrupted nuclear RNA surveillance in ALS/FTD pathogenesis
Latent HIV-1 Reservoir Decay and Antibody Response after Long-term Antiretroviral Therapy
No full textThe major barrier to curing HIV-1 is a stable, latent reservoir in resting memory CD4+ T cells. These cells carry replication-competent virus but do not express viral proteins and therefore cannot be cleared from the immune system. Antiretroviral therapy (ART) stops disease progression by preventing new infection events. However, viral rebound occurs upon ART cessation because ART does not eliminate the latent reservoir. The latent reservoir decays slowly with a half-life of 44 months during the first seven years of ART. We now have PWH (people with HIV) on ART for more than 20 years, and it is unclear whether this decay has continued and whether there have been qualitative changes in the reservoir as suggested by recent integration site studies. Our lab originally shown that in vivo, HIV mainly integrated within highly transcribed genes. Recently, it has been suggested that after 20 years of ART, there is a gradual selection such that the remaining proviruses are mainly in regions of the genome that are less transcriptionally active, allowing for a deeper state of latency. This implied that after two decades of ART, PWH could stop treatment without viral rebound. To assess this, we found 42 PWH on long-term ART and used the quantitative viral outgrowth assay to test whether any inducible, replication competent proviruses were still present. We found that reservoir decay did not continue after seven years of ART and that these study participants had higher than expected frequencies of inducible, replication competent proviruses which began to increase with a doubling time of 23 years. Using the intact proviral DNA assay on the same samples, we found that the frequency of intact proviruses also did not continue to decrease after seven years of ART. Through sequencing, we determined that most of the outgrowth viruses had the same env gene, suggesting that this lack of decay reflected the proliferation of cells that were infected before ART was started. Prior studies showed that autologous neutralizing antibodies (aNAbs) can block outgrowth of a substantial but variable fraction of reservoir viruses. Thus, to characterize aNAb dynamics over time, we performed neutralization assays for outgrowth viruses from a large cohort of PWH on ART, including those with longitudinal aNAb timepoints spanning decades of ART. We found that the majority of outgrowth viruses were resistant to neutralization by aANbs. Over time on ART, the aANb response waned, while for other reservoir viruses, we observed long-term stability of aNAb neutralization. These responses depended on each viral isolate and each PWH. As ART completely stops virus replication and evolution in adherent PWH, our data suggests that initiation of ART shortly after infection will prevent accumulation of aNAb-resistant viruses in the reservoir that can contribute to viral rebound
ADVANCING EQUITY IN PUBLIC HEALTH DASHBOARDS: DEVELOPMENT, VALIDATION, AND APPLICATION OF THE DASHBOARD INSTRUMENT TO REVIEW EQUITY (DIRE) TOOL
No full textBackground
The COVID-19 pandemic placed unprecedented strain on the global public health system, demanding rapid decision making with limited data on disparities affecting vulnerable populations. While dashboards informed public health decisions, they failed to address equity gaps— underscoring limitations to meet the needs of an emergency.
This dissertation presents the development of an innovative approach— the Dashboard Instrument to Review Equity (DIRE) framework and checklist— an equity-centered and data-driven decision support tool designed to systematically identify and address health equity gaps and better support public health decision making.
Methods
In Study 1, the DIRE framework was developed based on the emerging themes from scoping reviews. In Study 2, a mixed-methods study was conducted to refine the framework and develop the DIRE checklist. Interviews and surveys were conducted with decision makers and data practitioners. In Study 3, the reliability of the checklist was assessed by two reviewers who applied the tool to a sample of U.S. state-based dashboards and calculated the interrater reliability.
Results
In Study 1, the DIRE framework integrated five key themes from literature and scoping reviews. In Study 2, 16 interviews and 43 surveys were conducted. The interview thematic analysis demonstrated high interrater agreement (94.3%). Survey and interview findings guided the development of the DIRE checklist, which was organized into 6 categories. In Study 3, the checklist was applied to 26 dashboards. Interrater reliability was calculated at 72.7%, indicating substantial agreement. Dashboard scores reached a maximum score of 45.5% out of 100%. Results confirmed the reliability of the tool.
Conclusions
The DIRE framework and checklist tool innovatively equips public health departments with a practical, reliable, and validated tool to effectively bridge gaps in health equity, decision support, and emergency preparedness. The dashboard assessment reinforces the current need to improve the capacity of dashboards to guide decision makers. While DIRE is equity-driven, it is not equity-exclusive, as it also promotes multisectoral collaboration, strengthens communication, and enhances data-driven decision making. DIRE is a holistic and essential tool that provides critical guidance, best practices, and a clear roadmap to strengthen public health dashboards
Navigating Uncertainty: Understanding Housing Insecurity Among Low-Income Families
Full text linkThis dissertation examines the persistent U.S. affordable housing crisis and its implications for low-income families through three related studies. Using qualitative narrative data from in-depth interviews with Housing Choice Voucher recipients in the Seattle metropolitan area, it explores how housing insecurity is experienced, navigated, and responded to. Together, these chapters provide a comprehensive analysis of the dynamics of housing insecurity, emphasizing the interplay between structural limitations, social relationships, and individual decision-making. Chapter 2 examines the common drivers of housing insecurity, revealing both a high frequency of forced and reactive moves and a high prevalence of such moves across the sample. Based on these findings, this study provides recommendations to improve common survey measures, advocating for broader frameworks to capture housing insecurity fully. Chapter 3 focuses on doubling up arrangements, this sample's most common response to housing crises. I present three categorizations of doubling up arrangements—stable and supportive, stable yet strained, and fragile—demonstrating that while under the best conditions, doubling up can offer stable housing, childcare, and emotional support, even hosts that are secure or willing to provide consistent support encounter challenges that strain these arrangements. Less stable shared housing arrangements, such as when hosts depend on housing assistance, tend to be short-lived and precarious. Chapter 4 more broadly examines the decision-making process of families seeking shelter following a housing crisis. Families facing tight timelines and scarce resources must make challenging compromises on their well-being. In addition, selecting from a suboptimal set of housing strategies creates additional consequences for families and their children. Together, these chapters reveal that housing insecurity is often an ongoing process in which one episode of insecurity contributes to the next and the solutions families turn to can leave them vulnerable to additional precarity. The insights from this dissertation contribute to our understanding of how poverty and inequality are reproduced among low-income families
Controllability and Configuration Space Enhancement of Powered Multi-DOF Upper Limb Prostheses
Full text linkThe control paradigm for commercial electrically powered prosthesis has remained relatively unchanged since the advent of this technology in the late 1960s. Electromyographic (EMG) signals are measured from a user's residual limb as a means to estimate motor intention. Today machine learning-based `pattern-recognition' algorithms that classify EMG signals into a predefined set of motion classes are considered the state of the art in research and have seen increasing use in clinical settings. Unfortunately, these methods are not well suited to multi-DOF systems, and there is a great body of research demonstrating the deficits of modern upper limb prostheses. Even when presented with multiple controllable DOF, patients using interfaces that require sequential joint modulation tend to utilize just one of the available DOF, leading to inefficient, unnatural motion, and eventually long-term overuse injuries. It is important to note that the issue with prosthesis functionality is not due to dexterity issues with the mechatronics. The field of robotics shows us that purely mechanical systems are capable of autonomous dexterous object manipulation. This indicates that issues with prosthesis dexterity stem from the user interface and the ability to reliably interpret and execute user intent. This dissertation specifically explores the translation of user will to prosthesis action. First we maximize the efficacy of existing techniques for decoding motion intent from EMG by developing a tool for rapid and effective electrode placement during prosthesis fitting. Next we develop and test a hybrid gaze+ EMG semi-autonomous prosthesis interface that is capable of interpreting task-level intent, allowing for functional control of higher DOF systems. We demonstrate that through use of such an interface, users can effectively utilize full 3DOF prosthetic wrists, demonstrating a significant (p < 0.01) decrease in compensatory motion over PR systems. Thirdly, we develop a bio-mimetic method for implementing these interpreted actions by developing a human-in-the-loop trajectory planner for cooperative hand placement with a prosthetic elbow. This work spans methods to estimate motion intent, interfaces to enable exertion of that will onto high DOF systems, and algorithms to generate machine motion that are effective and efficient for the user
BRIDGING THEORY AND PRACTICE: DEVELOPING STATISTICS-INFORMED OPTIMIZATION MODELS TO ADVANCE HEALTHCARE DECISION SUPPORT
Full text linkHealthcare systems face unprecedented challenges, including resource limitations, disparities in access to care, and the increasing complexity of patient needs. This thesis addresses these issues by integrating optimization, machine learning, and statistical methodologies to develop data-driven frameworks for decision support in healthcare. Through three interrelated studies, this work aims to improve resource allocation, patient safety, and public health outcomes.
The first study investigates disparities in COVID-19 mortality across U.S. counties, utilizing longitudinal analyses to understand the dynamic associations between healthcare access, socioeconomic factors, vaccination coverage, and mortality outcomes. The findings highlight the persistent impact of social vulnerability and healthcare access on COVID-19 outcomes and provide actionable insights for mitigating future public health crises.
The second study advances fall risk assessment in hospital settings by optimizing the Johns Hopkins Fall Risk Assessment Tool (JHFRAT) and augmenting it with electronic health record (EHR) data. By incorporating machine learning and optimization techniques, the enhanced models significantly improve sensitivity and specificity while maintaining clinical interpretability. These advancements enable precise identification of at-risk patients, inform targeted interventions, and enhance operational efficiency.
The third study introduces a probabilistic framework for inverse optimization to address challenges in noisy and incomplete data. Leveraging maximum likelihood estimation, this framework enables robust parameter recovery and provides confidence intervals for decision-makers. By bridging theoretical rigor and practical applicability, the study ensures robust and interpretable solutions to critical healthcare challenges.
This thesis demonstrates how the synergy of optimization, machine learning, and statistical methods can address pressing challenges in healthcare. The contributions made here not only advance methodological innovation but also ensure their applicability in real-world clinical and public health settings. Future research will focus on expanding the applicability of these frameworks, incorporating additional data streams, and developing real-time decision support systems to enhance responsiveness and predictive accuracy
ANGLER’S WAY A COLLECTION OF NOVEL CHAPTERS
No full textAngler’s Way is a collection of four novel chapters that follows the lasting friendship of two men, Adam Richards and Colin McCormick, bonded by their experiences in Vietnam and the four decades that followed. Set in post-Vietnam America, the story follows the return of Adam to his hometown of Davisville, Pennsylvania and Colin’s efforts to build a stable family in Prescott, New Jersey. As they near the end of their lives, both men are compelled to reflect upon the choices that shaped their lives and their consequences for themselves and those they love. This work of literary historical fiction explores the themes of personal growth through grief, guilt, resilience and forgiveness
INVESTIGATING POST-TRANSLATIONAL MODIFICATIONS OF BCL-12 TO REGULATE BCL-2 FAMILY PROTEIN ACTIVITY
No full textApoptosis, a regulated form of programmed cell death, is critically governed by the BCL-2 protein family. This study focuses on BCL-12 (BCL2L12), a lesser-characterized member of this family, and its phosphorylation-mediated regulation of apoptotic activity. We demonstrate that BCL-12 interacts with the anti-apoptotic protein BCL-xL and that phosphorylation at critical serine residues likely modulates this interaction. Phosphatase assays and treatment with pharmacological inhibitors okadaic acid reveal that PP2A phosphatases may regulate BCL-12 dephosphorylation. We explored whether DNA damage or nutrient deprivation stimulates BCL-12 phosphorylation. The functional assays highlight BCL-12’s role as a dynamic regulator of apoptosis, with phosphorylation acting as a molecular switch to modulate BCL-xL activity. These findings provide novel insights into BCL-12’s regulatory mechanisms and its potential as a therapeutic target in cancers resistant to anticancer therapy. Limitations include the need for further validation of phosphorylation sites and stimulus-specific pathways. Overall, this study underscores the importance of post-translational modifications of BCL-12 in apoptotic signaling and proposes combinatorial strategies for targeting BCL-12
FIGURING THINGS OUT A COLLECTION OF ARTICLES AND ESSAYS ON SCIENCE
Full text linkWhile this thesis has no definite theme, I want it to reflect my progression as a writer. Much like how the rest of us move through the world, there is a lot of stumbling, yelling, and figuring things out on our own before we find a community that helps lift us up. The pieces in this thesis range from personal essays to scientific explainers to comedic stories with a science twist. The point of this thesis is to give a small window into the work of scientists and conservationists as they strive to better our understanding of the world and to better the environment for the animals that live here (plus a bit of comedic relief at the end)