Statement of Purpose

ICS Ph.D. portfolio, category 1.

Research

My research focuses on a central challenge in medical AI: whether diagnostic models generalize to the populations and clinical settings in which they are ultimately used. I build and evaluate machine learning systems for medical screening, with particular attention to populations underrepresented or excluded during model development. I am interested not only in identifying where performance deteriorates and for whom, but also in understanding why it happens and developing methods to address it.

This work is particularly meaningful in Hawaiʻi, where Native Hawaiian and Pacific Islander women experience disparities in breast cancer stage at diagnosis and outcomes, yet mammographic AI models have rarely been validated in these populations. Most existing models were developed using cohorts from the continental United States and Europe. Their performance in Hawaiʻi cannot simply be assumed, especially when differences in patient populations, imaging equipment, and clinical practices may affect how models behave.

My current research follows two related directions. The first investigates heterogeneity across populations and institutions in mammographic AI. I aim to distinguish the effects of population characteristics, including age, race and ethnicity, and breast density, from institutional factors such as scanner vendor, acquisition protocol, and screening environment. These sources of variation are frequently entangled, making it difficult to establish why a model performs differently across settings. Beyond measuring performance loss, I investigate whether personalized federated learning can recover it while allowing institutions to retain control of their patient data. I am especially interested in understanding where adaptation occurs: whether it changes learned image representations, the mapping from those representations to predicted risk, or simply the calibration of model outputs. Distinguishing these mechanisms could lead to more targeted, efficient, and clinically practical approaches to model adaptation.

The second direction addresses women with breast implants, a population frequently excluded from mammographic AI development. Implant-displaced mammography introduces additional views and imaging characteristics that conventional four-view architectures are not designed to accommodate. These differences create both technical challenges and opportunities to develop models specifically suited to this population. My interest extends beyond improving classification performance to determining when a model has sufficient information to make a meaningful prediction and when it should appropriately abstain.

These directions share a common research framework: identify where existing models fail, investigate the population or acquisition characteristics responsible, and develop adaptations or new models that address those limitations. Both draw on the Hawaiʻi and Pacific Islands Mammography Registry and explore the potential of language-grounded representation learning for populations with limited annotated data. Together, they inform my developing dissertation research and a broader program centered on reliable, population-aware medical imaging AI.

My preliminary work demonstrates the importance of this approach. In research presented at the MICCAI DeCaF workshop, I evaluated federated fine-tuning of GMIC across a three-site federation that included a diverse Pacific Island cohort. Standard global strategies, including FedAvg and FedProx, primarily benefited the largest site while leaving smaller sites at or below their pretrained performance. In contrast, Ditto, a personalized federated learning method, improved performance at the worst-performing site. These findings illustrate why aggregate improvements cannot be assumed to benefit every participating population and motivate my interest in understanding the mechanisms through which personalization succeeds.

This emphasis on generalization also connects my earlier research. In my QCHAT-10 autism screening work, I evaluated cross-population transfer using cohorts from New Zealand, Saudi Arabia, and Poland. In Parkinson’s disease research with the Hawaiʻi Digital Health Lab, our group investigated how device characteristics and handedness influence remote AI assessments. In an umbrella review of AI for skin cancer screening, I examined the limited demographic diversity of validation datasets and the lack of implementation science frameworks in the reviewed literature. Across these projects, I have pursued a consistent question: how can we establish that a model’s reported performance translates into reliable performance for the people and settings where it matters?

Professional Goals

My long-term goal is to become a tenure-track faculty member leading a research group focused on trustworthy, equitable, and clinically useful medical AI. I envision a career that integrates methodological research, real-world deployment, and the training of researchers who can work across computer science and clinical disciplines.

I plan to begin with postdoctoral research that deepens my expertise in machine learning methodology, particularly federated learning, multi-view medical imaging, and representation learning. My doctoral work has given me experience applying and critically evaluating existing methods under challenging clinical and population shifts. I want to build on that foundation by developing new approaches to personalization, generalization, and model reliability.

I also intend to spend part of my career in industry, working on the development and deployment of machine learning systems at scale. Before entering doctoral study, my professional experience centered on operational systems and the responsibility of ensuring that they worked reliably for the people who depended on them. That perspective continues to shape how I approach research. Industry experience would allow me to deepen my understanding of the engineering, monitoring, regulatory, and organizational challenges that determine whether promising models become effective real-world tools. I see this experience as complementary to academic research and particularly valuable for building a research program that bridges methodological advances and clinical implementation.

Teaching is also an important part of the academic career I hope to build. Although I have not yet had the opportunity to teach in a university classroom, I discovered how much I enjoy teaching while serving as an Army officer, where I taught introductory IT and command-line skills to soldiers with varied technical backgrounds. I continue to find that same satisfaction in coaching my son’s youth soccer team. I look forward to teaching computer science at the university level, helping students develop confidence with challenging technical material, and mentoring those entering the field from nontraditional backgrounds.

Ultimately, I intend to bring these experiences together in academia, establishing a research group that develops medical AI with rigorous validation, practical deployment considerations, and meaningful clinical partnerships built into the research process. I want to contribute to a field in which underrepresented populations are considered throughout model development and evaluation, rather than addressed only after systems have already been built. Through research, teaching, and mentorship, I hope to help advance medical AI that is both methodologically rigorous and genuinely useful to the people it is intended to serve.