Tuberculosis (TB) remains a major global health challenge, causing over a million deaths annually despite being a curable disease. A critical issue is treatment non-adherence, as many patients struggle to complete the required six-month regimen due to a lack of support and access to reliable medical guidance. Improving treatment adherence can significantly increase recovery rates and save lives. This project develops an AI-augmented chatbot powered by GPT-based models to assist Spanish-speaking TB patients. This is done by providing accurate medical guidance, fostering empathy, and enhancing communication between patients and healthcare providers. Integrated into a Human-System Interaction (HSI) interface, the system employs three AI models: a two-step pipeline that classifies messages as informational or emotional to tailor responses appropriately, a few-shot model that generates responses based on examples from prior patient interactions, and a Retrieval-Augmented Generation (RAG) + few-shot model that retrieves relevant medical information from guidelines while maintaining conversational fluency. These models leverage the same underlying technology as ChatGPT, optimizing responses for accuracy, linguistic fluency, and empathy. As part of the research team, I contributed to model development and implementation, ensuring alignment with medical guidelines and human-centered design principles. The chatbot is currently undergoing external evaluation by a multidisciplinary team, including healthcare professionals specializing in TB treatment and AI researchers. Evaluators interact with the chatbot using personas as TB patients, asking medical and support-related questions to assess response quality. They rate the system based on medical accuracy, linguistic fluency, empathy, and other key criteria relevant to patient-provider communication. Insights from this evaluation will guide future refinements, with the goal of improving AI-driven patient support systems in clinical settings.

Dementia, a growing global health concern, affects the nervous system and leads to severe cognitive impairment, with Alzheimer’s disease (AD) being the most common form, currently impacting nearly 7 million Americans. As life expectancy increases, the prevalence of dementia increases in corresponding fashion, driving research efforts like those of the Darvas Lab, where we study AD and other related dementias using adeno-associated viruses (AAVs) to induce neuropathologic changes. The TDP43 protein is involved in neuropathologic changes such as those in Frontotemporal Dementia (FTD) and in Amyotrophic Lateral Sclerosis (ALS), a primary motor neuron disorder. TDP43, primarily localized in the nucleus, plays a crucial role in regulating gene expression and RNA metabolism. TDP43 pathology in neurons involves the presence of TDP43 in the cytoplasm and its accumulation in cytoplasmic inclusions. To better understand the role of TDP43 in neurodegeneration, we use a mouse model where TDP43 proteins are introduced via AAV, a genetically engineered viral vector commonly used in research. This approach allows control over the timing of neuropathologic changes. Our prior AAV constructs included the Synapsin I promoter, which led to a severe ALS-like motor phenotype due to its expression in spinal motor neurons. However, this model could not be used to study the more subtle effects of dementia due to the extreme nature of the physical pathology. Therefore, our goal is to produce a new model to overexpress TDP43 using an AAV that is exclusive to the cortical brain regions relevant to FTD by instead including the CamKIIα promoter, which exclusively drives expression in the forebrain. I assessed behavioral phenotypes in our mouse model by conducting a Y-maze to evaluate effects on short-term memory, and analyzing neurological scoring to evaluate neuromuscular dysfunction. The development of a more dementia-focused TDP43 model will allow us to more specifically investigate its neuropathology.

TDP43 is an RNA/DNA binding protein that forms pathological aggregates in most amyotrophic lateral sclerosis (ALS) and half of frontotemporal lobar degeneration (FTLD) cases. Knockout of TDP43 in animal models leads to neurodegeneration and motor deficits, but overexpression of wildtype TDP43 leads to the same events; therefore, TDP43 protein homeostasis is critical to prevent ALS/FTLD. To achieve this homeostasis, TDP43 autoregulates its own mRNA splicing, resulting in multiple TDP43 isoforms. Although some of these isoforms go through nonsense mediated decay, other isoforms result in unique proteins with differing C-termini. This leads to variable cellular localization. It is unknown if these alternative, protein-coding isoforms are predominantly associated with ALS/FTLD or if aging changes the frequency of these isoforms. To determine how TDP43 overexpression yields different isoforms and interacts with aging and ALS-like symptoms, the Darvas Lab created a novel approach to overexpress human TDP43 (hTDP43) via Adeno-Associated Virus (AAV) delivered through retro-orbital injection, leading to ALS-like motor deficits. We tested this AAV in young and old mice cohorts. Then, to determine if Tardbp alternative splicing is linked to ALS-like symptoms and aging, I designed and validated primers and protocols to measure the nine Tardbp mRNA isoforms in mice via quantitative real-time polymerase chain reaction (qRT-PCR). I have started to determine if hTDP43 overexpression leads to differential splicing compared to mice injected with a sham-control AAV in these old and young mice. Once this is done, we will clone the most interesting differentially spliced isoform in an AAV and inject that AAV and a full-length TDP43 AAV into mice to see if the spliceform causes increased toxicity, manifesting in worsening motor deficits and mortality.

Data is the fuel driving AI innovation. Much of the most valuable data is, however, siloed in research centers, hospitals, banks, etc. The onerous processes researchers must go through to access each silo cause a substantial underutilization of AI in many of the most important domains, including healthcare and genomics. AI researchers cannot train models for personalized medicine if they cannot get their hands on enough relevant patient data. One way to provide broader access for research while also retaining the privacy of the original data is with synthetic data generation (SDG), which uses machine learning to generate a set of synthetic data similar enough to the real data to retain its value for research while also anonymizing it. While in some cases a single data custodian (such as a hospital) alone may have enough data to train a generative model, usually, datasets from multiple custodians need to be combined to reach a cumulative size that enables meaningful AI research. The latter is, for example, often the case for rare diseases, with each clinical site having data for only a small number of patients, which is insufficient to train high-quality synthetic data generators. The goal of my research is to generate synthetic genomics data of patients with Neurofibromatosis type 1, a rare genetic condition that causes changes in skin pigment and tumors on nerve tissue. Thanks to our Privacy-Preserving Machine Learning Lab’s inclusion in the National Artificial Intelligence Research Resource (NAIRR) Pilot and our collaboration with Sage Bionetworks, I have access to the TACC Frontera supercomputer at the University of Texas and multiple sets of NF1 patient data. Results of my work on the NAIRR include an empirical evaluation of cross-silo federated SDG algorithms in terms of quality of the generated NF1 data, computational cost, and level of privacy protection.

As of 2025, the United States has the highest incarceration rate in the world, with its incarcerated population making up 25% of the incarcerated individuals worldwide. Mass incarceration inflicts the most harm on the most vulnerable populations, disproportionately affecting racial and ethnic minorities and creating insurmountable barriers to reintegrating into society. Prison education programs provide opportunities for growth that help prevent recidivism and support rehabilitation efforts, and with the reinstatement of Pell Grants for incarcerated individuals in 2023, there has never been a better time to expand educational opportunities than now. However, little research has been done on prison education programs, with even less research focusing on enhancing and expanding them to address the specific needs of incarcerated individuals, particularly in digital literacy. In a rapidly evolving digital world, it becomes imperative to ensure that incarcerated people, many of whom have had limited experiences with technology due to extended sentences, have the skills to confidently return to a digital society. This project explores how integrating computer science curricula into correctional facilities can increase rehabilitation, reduce recidivism outcomes for incarcerated individuals, and further support other pre-existing educational programs in prisons. To answer this question, we examined legal documents, performed literature reviews, analyzed previous studies on the incarcerated population, and conducted a comprehensive analysis of outcomes from prior prison education programs. Our findings reveal that computer science education for incarcerated people increases self-efficacy rates, post-employment opportunities, and facilitates a smoother transition back into society. Additionally, integrating computer science through enhanced digital infrastructure can address challenges with current educational programs, such as accessibility, course expansion, and classroom segregation. Collectively, this project represents one of the first studies to explore the possibilities for computer education and prisons, offering valuable insights into the potential to improve rehabilitation, reduce recidivism, and address the digital divide.

Ocean eddies contribute significantly to the transfer of heat and energy throughout the world’s oceans, playing a key role in regulating climate. Eddies are observed predominantly through Earth-orbiting satellites that collect data on sea surface height (SSH), a metric that can be used to estimate eddies on a global scale. Historically, satellites could only capture point-wise measurements, resulting in low-resolution SSH maps, which led to underestimations of small-scale eddy strength. Launched in 2022, NASA’s Surface Water and Ocean Topography (SWOT) satellite now provides groundbreaking 2D SSH imagery with higher resolution relative to existing SSH products. However, there are only two years of SWOT data available, unlike other satellites with decades-long records. Here, we considered how the recent SWOT data could be deployed to improve the spatial resolution of SSH products from the preceding 30 years. To achieve this, we trained an image-to-image U-Net neural network to predict the high-resolution SSH from an existing low-resolution product (NeurOST). We used SWOT high-resolution data as a ground truth to train this neural network and minimize the mean squared error of the SSH output with respect to the SWOT data. By evaluating the accuracy of the SSH output maps against an independent withheld satellite, we demonstrated that our method improves the spatial resolution of the SSH product compared to the NeurOST dataset. We next plan to test the accuracy of our method when applied to years before SWOT was launched. Overall, our research highlighted how leveraging deep learning and SWOT can enhance the spatial resolution of a decades-long eddy observation time series, enabling more accurate studies of eddies and their climate impact.

Understanding and predicting changes in primary productivity depend on both upper ocean warming and nutrient supply from the ocean interior. Fronts, where distinct water masses converge, are hotspots for these vertical exchanges, transporting nutrients upward and supporting diverse ecosystems. These fronts create sharp gradients in temperature and salinity, generating strong vertical velocities that upwell nutrients and biomass. However, the exact dynamics of frontogenesis (the formation of fronts) remain poorly understood. Additionally, these processes occur at scales too fine to be resolved in global climate models and are only marginally captured by high-resolution ocean simulations. This underscores the need for observational studies to characterize frontogenesis and test existing theoretical frameworks. In this study, we diagnose frontal dynamics using Petterson’s frontogenesis function, which quantifies the roles of divergence and strain. Using NcCut, a GUI developed by our group, we compiled a unique dataset capturing the full life cycle of numerous ocean fronts in front-following coordinates from a state-of-the-art ocean simulation. Our results indicate that for mesoscale (~100 km) fronts, strain dominates over divergence, aligning with classical theories. In contrast, submesoscale (~10 km) fronts exhibit shorter life cycles and no clear dominant driver of frontogenesis within the Petterson framework. We also identified key limitations in conventional diagnostics and improved our analysis by masking the front from its surrounding environment before diagnosing its drivers. This enhancement provides a more accurate representation of frontogenesis dynamics. In the future, we plan to apply our method to satellite observations to study real-world ocean fronts, validate ocean models, and improve predictions of primary productivity changes. Our findings highlight the importance of refining frontogenesis diagnostics to better capture the small-scale dynamics critical to ocean biogeochemistry and climate predictions.

Following an extensive history of industrial activity in Commencement Bay, Washington, the health of marine ecosystems continues to be affected by persisting pollutants. Commencement Bay has been identified as a Superfund Site, in which the Environmental Protection Agency (EPA) is tasked with cleaning up locations contaminated with hazardous materials. In an effort to gauge just how effective these recovery efforts have been, this study, part of the Puget Sound Foraminifera Project at the Burke Museum, investigates how the density and diversity of benthic foraminiferal assemblages have changed over time. Foraminifera, a diverse and widespread order of shelled marine protists, can be utilized as a reliable measure of marine ecosystem health due to their innate sensitivity to environmental changes. Samples collected by the Washington Department of Ecology (WDOE) from 2014 and 2022 allow for a comparison of diversity indices that are indicative of the success in the bay’s recovery. To quantify this success, calculations of the Shannon Index and the Simpson Index were completed for each sample, supporting our determination of the Foraminiferal Benthic Index (FBI) of the region. The FBI was defined using measures of abundance, diversity, and percentages of tolerant species present in each sample to quantify the extent of adversity. With 2022 density and diversity averages that are statistically similar to those of 2014, we can conclude that clean up efforts have not yet made sufficient measurable improvements in the Foraminiferal Benthic Index over the previous eight years.