Three threads, one closed loop
We develop computational methods for cancer care that observe how patients change during treatment, translate those signals into timely treatment decisions, and make those advances practical and accessible in real-world care.
Response monitoring & biomarker discovery
Cancer treatment produces biological and physiological changes long before they become visible as conventional clinical endpoints. We develop and validate biomarkers that capture this evolving patient state during therapy, with particular emphasis on measurements that are longitudinal, low-burden, and practical to acquire repeatedly.
Our work spans digital biomarkers derived from wearable physiological sensing, imaging biomarkers that characterize tumor evolution from routinely acquired treatment imaging, and molecular biomarkers, including circulating tumor DNA and other blood-based measures. Across modalities, the goal is the same: identify signals that tell us, early and reliably, how an individual patient is responding.
Dynamic treatment adaptation
A biomarker becomes clinically useful when it can inform a decision. We develop computational methods that translate evolving patient signals into timely, actionable treatment interventions.
This requires deciding not only how treatment should change, but whether and when adaptation is warranted under uncertainty. Our work spans optimal stopping and sequential decision-making, biologically informed adaptive radiotherapy, bilevel and multicriteria optimization, and the integration of machine-learning-based outcome and toxicity models directly into treatment planning. The goal is to move from prediction to prescription: using what we learn about the patient to determine the next best treatment decision.
Accessible & patient-centered delivery
Precision should not require complexity that puts it out of reach. We design monitoring and decision-support systems with scalability, patient burden, and clinical resources in mind, so that advances in personalization can extend beyond highly specialized settings.
This principle shapes both sides of our work: patient-facing digital health technologies that enable inexpensive, longitudinal monitoring outside the clinic, and health-system optimization tools that help allocate scarce treatment resources more effectively. Projects such as CardiMom and NexusRT aim to make sophisticated monitoring and treatment delivery easier to use, easier to access, and more responsive to the needs of individual patients.
From observation to action
Across these programs, the underlying question is the same: what can we learn about a patient's evolving state, and how should that information change the next decision? We approach that problem using mathematical optimization, statistical learning, machine learning, and decision-making under uncertainty.
Complete publication list on Google Scholar.