Data efficiency
How can models learn meaningful biomedical representations when expert labels and patient data are limited?
Explore the ECG studyResearch that holds up beyond the lab.
I’m drawn to questions where better machine learning can support better decisions for people. Biomedical data make that responsibility concrete: labels are limited, populations shift, errors carry different costs, and every useful result must stand up to careful scrutiny.
Useful research begins with real consequences.
I care about methods that do more than improve a benchmark. The deeper task is to understand when a model works, where it fails, and whether its behavior is clear enough to support responsible use. That makes data quality, evaluation, uncertainty, and reproducibility part of the research itself.
The questions guiding my work.
Generalization
How can evaluation reveal whether a model will remain reliable across patients, devices, sites, and populations?
Explore cold-start evaluationInterpretability
Can inspectable structure be built into learning itself, so explanations reflect how the model reaches a decision?
Explore the heart-risk studyUncertainty
How can calibrated uncertainty help a model recognize when human review matters most?
Inspect model calibrationRobust evaluation
Which leakage checks, stress tests, and subgroup analyses should be standard before a result is considered convincing?
Read the evaluation noteReusable representations
Which representations can carry useful knowledge across signals, clinical variables, and eventually other biological modalities?
Explore multimodal learningMolecular and multimodal learning
Recent drug–target affinity work extends these principles to molecular graphs, protein language-model representations, and cold-start evaluation. I’m treating this as a developing research direction that still needs complete benchmarks, careful ablations, and broader biological validation.