UCLA

The cornerstone of this dissertation is guayla, a remarkable popular music of the small northeast African nation of Eritrea, its diaspora and northern Ethiopia. Indigenous to Eritrea’s largest ethnic group, the Tigrinya, who make up 50% of the country’s total population of roughly 3-4 million, guayla arose out of traditional music-making and local customs, many of which are still performed today. Guayla has a highly distinctive sound that features soaring vocals, call-and-response singing, and a signature cyclical “five-beat” rhythmic structure that sets everyone dancing at social and political events. Ravinder Rena states that guayla “is [a] powerful and meticulous form of art that brings the emotional integration among the masses. The song is a driving force of an Eritrean society” (Rena 2009). Guayla is particularly present on Eritrean social media and YouTube channels and is a powerful marker of Tigrinya ethnic identity and Eritrean nationhood among the substantial diasporic communities located in North America, Europe, Saudi Arabia, and Australia. Lastly, secular guayla music contrasts with both the prevalence of ancient liturgical music of the Orthodox Tewahedo church which occupies an important place in the cultural history of the country, and the richness of the music of the other eight nationally recognized ethnic groups: Afar, B’dawiyet, Bilen, Kunama, Nara, Rashaida, Saho, and Tigre. The doctoral dissertation “Guayla Nation: Unyielding Tigrinya Music, Dance and Identity in Eritrea” draws from observations, interviews, videos, writings, ephemera, archives, and online testimonials to provide a preliminary ethnographic and historical exploration of Eritrea’s most popular and ubiquitous music legacy.

In the highly-active antiretroviral therapy (HAART) era, sexual minority men (SMM) have achieved high life expectancies. However, they continue to be disproportionately affected by aging-related outcomes (e.g., lung cancer) and low quality of life (e.g., high rates of hospitalization and mortality). Previous literature has shown that overlapping behavioral and psychosocial factors predict HIV infection, low HAART use, and sexual risk behaviors among SMM but whether these factors are linked to adverse health outcomes including hospitalization, lung cancer, and mortality remains unknown. In this dissertation, we used data from the Multicenter AIDS Cohort Study (MACS). The first study included men having sex with men (MSM) who completed at least two assessments between 2004 and 2019 in the Methamphetamine Substudy of the MACS, and we used generalized estimating equation (GEE) models to examine the synergistic effects of psychosocial factors on incident all-cause hospitalization. Our second study involved SMM who underwent at least two assessments in the MACS between 1996 and 2022, and we conducted survival analyses to examine the additive effects of behavioral and psychosocial factors on incident lung cancer. In the third study, we included SMM living with HIV who completed two or more assessments and initiated HAART for at least one year between 1996 and 2022. We examined the predictive capabilities of behavioral and psychosocial factors on all-cause mortality when these factors were added to the Veteran Aging Cohort Study (VACS) Index 2.0, which is a commonly used clinical tool to predict mortality among people living with HIV (PLWH). In our first study, we found synergistic effects of depression, stimulant use, smoking, heroin use, childhood sexual abuse, and intimate partner violence on hospitalization. In our second study, we found a dose-response relationship between the number of behavioral and psychosocial factors (depression, cigarette smoking, heavy alcohol use, and polydrug use) and lung cancer incidence. Our findings from the third study indicated that including depression enhanced the accuracy of the VACS Index 2.0 in predicting mortality. Future interventions should target behavioral and psychosocial factors to reduce healthcare costs and improve the quality of life and overall health among SMM in the US.

Road networks support economic activities, emergency services, and daily operations. However, those located in complex terrains, such as hillside regions, are particularly vulnerable to natural hazards such as earthquakes, landslides, and wildfires, which can severely disrupt connectivity and cause significant social and economic impacts. As climate change accelerates the frequency and intensity of these hazards, there is a growing need for scalable, data-driven approaches to enhance the resilience of these networks. Traditional methods for planning and retrofitting infrastructure fall short in addressing the complexities posed by both terrain and hazard exposure, particularly in large-scale networks.This dissertation aims to develop a comprehensive, data-driven framework with multiple modular components to improve the resilience and capacity of transportation networks in vulnerable regions. It focuses on introducing scalable algorithms for ranking road criticality, integrating hyperlocal data into capacity expansion planning, and optimizing retrofitting strategies to ensure network functionality during and after disaster events. Additionally, the work incorporates an equity dimension in retrofitting strategies to ensure fair resource distribution across different socioeconomic groups. To achieve these objectives, this dissertation presents multiple frameworks that develop and apply advanced machine learning and optimization techniques. First, it introduces a Graph Neural Network (GNN) for ranking road segments based on their criticality. This approach provides a more efficient alternative to conventional methods, significantly reducing the computational time needed to rank road segments while maintaining high accuracy. This capability enables rapid decision-making in both everyday operations and during disruptions. Second, the dissertation focuses on integrating hyperlocal data into capacity expansion planning, utilizing data gathered through mobile mapping systems equipped with LiDAR and vision-based technologies. By incorporating this high-resolution data, the framework allows for more accurate capital allocation, especially in hillside regions where terrain complexity poses a challenge for traditional methods. Third, the dissertation addresses the challenge of optimizing earthquake resilience in large-scale networks. A probabilistic framework is developed that combines a Siamese Graph Convolutional Network with a Genetic Algorithm. This approach evaluates various retrofitting strategies and identifies optimal solutions that balance budget constraints while minimizing welfare losses for low-income populations, ensuring a more equitable distribution of resources. Lastly, a novel framework is developed for managing wildfire risks to road networks using Generative Adversarial Networks (GANs). It generates high-fidelity synthetic weather scenarios that influence wildfire ignition and spread, creating a stochastic catalog of events to simulate fire progression and its impacts on road networks. This information feeds into an optimization model that identifies critical road segments for retrofitting, optimizing investments to maximize resilience. The methodologies were applied to transportation networks, mainly focusing on the hillside regions of Los Angeles, showcasing the effectiveness of each approach. The proposed GNN models improved the speed of road criticality rankings, while maintaining high accuracy and enabled large-scale retrofitting optimization. Hyperlocal data integration proved essential for precise capacity expansion, while the earthquake and wildfire resilience frameworks optimized retrofitting strategies, significantly increasing network resilience even with small investments. The main contribution of this dissertation is the development of scalable, data-driven methods for enhancing the resilience of transportation networks. By integrating advanced machine learning models and hyperlocal data, the proposed approaches provide a robust toolkit for policymakers and infrastructure planners to ensure that transportation networks can better withstand and recover from disruptions. These contributions form a strong foundation for resilient and more equitable infrastructure planning in regions susceptible to natural hazards.

This thesis investigates the use of deep learning and machine learning models for high-frequency stock price prediction and inference. By using models such as Long Short-Term Memory (LSTM) networks, Convolutional LSTMs (CLSTM), and Transformer architectures, this work evaluates the predictive performance of these models in both single-step and multi-step stock price prediction tasks. The models are trained on various datasets, including those with technical indicators, sentiment analysis, and the US Dollar Index, along with Fourier-transformed features for improved feature engineering. The results demonstrate that Transformer-based models, particularly those added with convolutional layers, outperform LSTM-based models in capturing long-term dependencies and making accurate predictions over extended time periods. Additionally, the Fourier-transformed features enhances overall models performance by revealing underlying periodic patterns in stock prices. This research contributes to the growing literature on stock price prediction and inference by offering insights into model architectures and feature engineering techniques that improve the accuracy of financial forecasting.

JWST has begun to reveal the nature of the small-scale structures of our Universe at unprecedented distances, probing an era of galaxy formation with remarkably different properties. This has offered the opportunity for novel tests of our models of cosmology and galaxy formation. In this thesis, I focused on a defining characteristic of ΛCDM structure formation, the streaming velocity, which has a variety of impacts on small-scale structure. Specifically, I developed models of and observational predictions for a new class of small-scale structures called SIGOs derived from this mechanism, with the aim of enabling future detections of these objects that would place constraints on alternative cosmologies. I also showed that--in contrast to previous expectations--the streaming velocity enhances star formation on a per-galaxy basis in the dwarf regime at high redshifts. Taken together, I showed that the streaming velocity has significant influence on high-redshift small-scale structure formation, and may serve as an important signal of our standard cosmology.

Recent research highlights a growing demand for youth mental health services (Barican et al., 2022; Kazdin, 2019; USPSTF, 2022), prompting the need to enhance mental health workforce capacity. Improving workforce capacity entails strengthening critical decision-making activities, including considering client problems, prioritizing them, and selecting the most suitable practices to address them. Clinical supervision, involving dyads of qualified mental health professionals ("supervisors") and direct service providers ("supervisees"), aims to improve these activities (Proctor, 1986; Milne, 2007). Challenges include time constraints, varying competency activity levels, and difficulty in incorporating new scientific findings, compounded by high turnover rates (Bernstein et al., 2015; Brabson et al., 2020; Chorpita et al., 2021; Collatz & Wetterling, 2012; Dorsey et al., 2017; Powell & York, 1992; Simon & Greenberger, 1971). Integrating decision support systems into clinical supervision could address these challenges, promoting use of evidence and ensuring sustained skill retention among supervisory dyads (Bjork & Bjork, 2020). Within the context of a decision-support system integrated within clinical supervision, this dissertation investigated the reliability of quality, effort, and efficiency metrics, and then examined the associations between ordinal repetition of activities and passage of time with those quality and effort metrics. As such, it explored whether time or repetition is associated with improvement, deterioration, or no change in these metrics. The study analyzes existing data from a multi-site randomized implementation trial aimed at promoting the use of evidence-based methods for engaging youth and families in treatment. We audio recorded and transcribed supervision events in which mental health workers discussed cases at-risk for poor treatment engagement. For part one, 26 supervisees and 17 supervisors discussed 30 cases; for part two, 48 supervisees and 16 supervisors, trained and using a decision-support system, discussed 118 cases. Observational coders rated efficiency and the extensiveness of decision-making activities using a subset of the ACE-BOCS coding system (Chorpita et al., 2018). Efficiency was rated holistically for each event on a 5-point scale, from presence of extensive discussions on unnecessary topics (1) to swift and organized decision-making and planning (5). Quality was evaluated using a dichotomous scale, based on whether each activity met sufficient quality criteria, primarily indicating the presence of the activity. Effort was measured by the total number of words spoken for each activity. Two overall effort scores were calculated based on the total words spoken and duration of the entire event. The total number of supervisory events per supervisory dyad was an indicator of repetition of supervisory activities, and the total weeks since training in the decision-support system measured the passage of time. To assess interrater reliability across all coders, we used Fleiss' kappa (κ) for the four dichotomous quality metrics and ICCs (model [2,1], consistency) for the ordinal efficiency metric. To examine possible change in outcomes, we used mixed effects regression models, examining three hierarchical levels: cases nested within supervisees nested within supervisors. Thus, supervisors were the main level of analysis. We assessed the impact of each level on results and simplified the model if it didn't improve it. To manage skewed data with quality and effort measures having excess zeros, we implemented corrections like the Firth logistic regression and employed specialized models such as the Hurdle model, respectively. These strategies helped mitigate bias and stabilize parameter estimates. Interrater reliability estimates showed that coders consistently rated both the decision-making activities and overall efficiency reliably. A strong positive correlation confirmed the initial validity of the effort measure. Findings revealed changes in efficiency, the presence of quality, and the likelihood of putting in effort as dyads moved through each level of supervision for their cases (for example, from the first supervision event type to the second and then to the third type). Increasing repetition of supervision events or time within each supervision stage did not predict whether the dyads improved in these outcomes. This study underscores the sustainability of quality, effort, and efficiency across repeated supervision events within different supervision types and over time. It also identifies areas for further investigation, including the need for more nuanced and robust measures of quality and effort. Future research should address these issues and explore alternative assessment methods to gain a deeper understanding of workforce learning. This understanding will inform strategies aimed at maximizing workforce capacity to meet the growing demand for high-quality youth mental health services.

The manipulation of biological objects is essential in various biomedical applications, such as single-cell analysis, cell-cell interactions, spheroid fabrication, drug delivery, and tissue engineering. Traditional methods of cell manipulation—including optical, magnetic, dielectrophoresis (DEP), and hydrodynamic techniques—each offer distinct advantages and limitations. In comparison, acoustics has been demonstrated to have superior biocompatibility and wide range of operable sizes of target, making it an attractive option to be widely utilized in many applications. However, traditional acoustic manipulation methods like BAWs and SAWs, which rely on standing waves, are constrained by limitations such as spatial resolution, periodic patterning, and restricted operational area. To overcome these limitations, the Compliant Membrane Acoustic Patterning (CMAP) mechanism was developed. Unlike BAWs and SAWs, CMAP does not rely on standing acoustic waves. Instead, it utilizes large acoustic impedance mismatches to generate a near-field acoustic potential gradient, enabling the precise patterning of microparticles and cells at sub-wavelength resolution with complex, non-periodic acoustic potential profiles over a large operational area. While CMAP offers high-resolution, reconfigurable patterning in complex, non-periodic shapes, which is challenging for traditional acoustic methods like BAWs and SAWs to achieve, several limitations remain. The fabrication process for CMAP devices is still complex and hard to control, with the operational area and patterning resolution restricted by the constraints of current fabrication methods. Moreover, CMAP is only currently capable of trapping particles and cells in groups, falling short of achieving single-cell level trapping and manipulation. Additionally, once microparticles and cells are trapped, these platforms lack the ability to selectively release specific particles or cells for downstream collection and further analysis. In this dissertation, two innovations are explored to address the limitations described. The first is a rapid prototyping method for fabricating Compliant Membrane Acoustic Platform (CMAP) devices using laser-assisted manufacturing. This direct-write approach not only offers a large area manufacturing capability but also eliminates the need for photolithography, significantly reducing the cost and time associated with the fabrication process. This new manufacturing method allows the transfer of a thin membrane onto the acoustic device reliably. The devices created using this approach achieve sub-wavelength, complex, and non-periodic patterning of microparticles and biological objects, with a spatial resolution of 60 μm across a large active manipulation area of 10 × 10 mm2—which is nearly 10 times larger than the priorly demonstrated CMAP device (CMAP 3mm x 3mm). The second innovation involves the development of a novel single-cell CMAP device that integrates a photothermal mechanism, enabling precise single-cell trapping and selective release. This advancement allows for the isolation and collection of individual cells, which is crucial for applications in single-cell analysis. The device utilizes spherical air cavities embedded in a PDMS substrate to create localized acoustic potential wells for trapping cells, enabling the individual trapping of both synthetic microparticles and biological cells across a 1 cm2 with more than 75,000 traps, accommodating a broad size range from 8 to 30 µm. The integration of a near-infrared laser facilitates selective release, achieving a release rate of approximately 40 cells per minute without compromising cell viability or proliferation.

Resource disaggregation has emerged as a promising solution to enhance both resource utilization and management efficiency in datacenters. Existing disaggregation solutions have largely centered on generic, low-level system optimizations such as minimizing remote access latency at the operating system and hardware levels. However, these solutions often yield suboptimal performance due to the lack of alignment between application semantics and the underlying system layers.

This dissertation presents a novel approach that enhances the performance of disaggregated systems by incorporating application semantics, including memory access patterns, data object ownership, and computational intensity, into the system design. Our methodology is demonstrated through three techniques -- Mako, MemLiner, and DRust. Each technique applies program semantics at different levels of the system stack, ranging from programming languages and compilers to runtime environments and operating systems. Specifically, Mako and MemLiner utilize program semantics to develop a new runtime that is optimized for disaggregated memory architectures. Meanwhile, DRust employs data object ownership semantics in applications to build a programming framework tailored for compute disaggregation.

Collectively, these proposed techniques aim to enhance the performance, efficiency, and consistency of disaggregated systems, making them more viable for practical implementation in today's datacenters. This body of work lays a foundational framework for the co-design and co-optimization of techniques across system layers, aimed at advancing future disaggregated datacenters.

Rituximab, an anti-CD20 monoclonal antibody, has revolutionized the treatment for lymphoma, particularly B- cell non-Hodgkin lymphoma; the therapeutic efficacy, however, is limited by its temporary activity, non-ideal biodistribution, and heterogeneity of the cancer cells. In this dissertation research, two auxiliary systems were designed to improve the therapeutic efficacy of rituximab. The first system involves a novel combination treatment for non-Hodgkin lymphoma using rituximab and a whole cell-based therapeutic cancer vaccine, which effectively elicited a specific immune response, established immune memory, offered both immediate tumor killing and sustained protection from relapse of cancer. The second system involves a design of novel protein delivery systems that target lymph nodes, which help eliminate circulating cancer cells within the lymph nodes, block metastatic pathways, and reduce cancer metastasis and relapse. The strategies described in this dissertation research can be extended to other monoclonal antibody therapeutics and cancer therapies beyond non-Hodgkin lymphoma therapeutics, providing a platform towards more effective cancer therapy.

Purpose Beam orientation and biological dose optimization are interdependent features of Intensity-Modulated Proton Therapy (IMPT). Current automated beam orientation optimization (BOO) methods are robust and able to provide beam convergence but have not accounted for accurate biological modeling that narrows the therapeutic window. Biological models such as relative biological effectiveness (RBE) and oxygen enhancement ratio (OER) and machine parameters such as dose-averaged dose rate (DADR) are highly complex and lead to computationally challenging frameworks that may be solved by novel optimization methods. Methods The robust BOO framework for IMPT was formulated with physical dose fidelity to provide accurate dose to the tumor and limit dose to organs at risk (OARs), a heterogeneity-weighted L2,1/2-norm group sparsity term to reduce the number of active beams to 2-4, and a sensitivity regularization term. The dose fidelity term was updated to consider variable RBE values, lower oxygenation status in tumor regions, and the normal tissue sparing effects caused by high dose rate. These biologically-informed BOO frameworks were solved with RBE and dose rate linearization along with splitting schemes. The plans were generally tested on challenging head-and-neck (H&N) cases and compared against previous plans in terms of dosimetry and robustness. Results Compared to IMPT BOO plans solved with constant RBE=1.1, McNamara RBE-based dose was able to improve OAR [Dmean, Dmax, worst Dmean, worst Dmax] by an average of [36.1%, 26.4%, 25.0%, 19.2%] with modest CTV coverage and robustness improvement. Additionally, hypoxia-based RBE dose fidelity was able to increase tumor [HI, Dmax, worst HI, worst Dmax] by [31.3%, 48.6%, 12.5%, 7.3%] with only [8.0%, 13.1%] increase in OAR [Dmean, Dmax], increasing the therapeutic index. Next, compared to spread-out Bragg peak IMPT BOO plans, dose rate-optimized plans with Bragg peak and shoot-through beams combined were able to increase volume of ROIs receiving >40 Gy/s by approximately 41.1%, while improving CTV homogeneity by 5.6% and improving OAR dose in several structures. Conclusions Novel optimization methods were developed for biologically-guided IMPT. The objective function integrates RBE-weighted dose, hypoxia-informed dose, and dose rate optimization into a unified framework with BOO as the baseline objective. Compared with the physical dose BOO or manual selection, our method generates plans with superior tumor and normal tissue dosimetry and robustness.