My laboratory focuses on developing advanced artificial intelligence methods for early disease detection using multimodal clinical data. Our overarching goal is to design computational systems that assist clinicians in identifying clinically relevant information that is difficult or impossible to perceive through conventional human interpretation, thereby enabling earlier and more precise diagnosis. Based on the different types of clinical data we work with, our research is organized into several complementary areas.

Medical imaging data. A major direction of our work focuses on imaging modalities including MRI, CT, ultrasound, PET, and digital pathology. We believe that radiological images may contain latent cancer-related signatures that are invisible to the human eye, similar to how bees can perceive ultraviolet light or migratory birds sense magnetic fields while humans cannot. We aim to build AI systems capable of extracting these sub-visual micro-lesion patterns, surpassing the limits of human visual perception to achieve early cancer detection at scales below human visibility.

Clinical text data. Our work on radiology reports, physician notes, discharge summaries, and other unstructured narratives aims to uncover hidden relationships among diseases, symptoms, phenotypes, medications, and genetic factors. We combine public knowledge bases with private patient-level records to identify subtle diagnostic clues that are often overlooked. For example, intracranial hypotension–induced headache may be associated with cerebrospinal fluid leakage following spinal anesthesia; automated retrieval of relevant procedural histories embedded within patient notes can trigger early diagnostic consideration and targeted clinical evaluation.

Structured laboratory data. Our research develops deep learning models for clinical classification tasks with special emphasis on interpretability, enabling rigorous performance while providing clinically meaningful explanations. A central challenge we address is how to represent heterogeneous numeric variables in a semantically coherent manner that supports meaningful inference.

Omics data. Our research focuses on high-dimensional and ultra-sparse single-cell transcriptomic datasets. We study early unsupervised clustering methods for single-cell RNA sequencing to facilitate downstream biological interpretation and discovery. Our collaborative work extends to disease-gene identification and mechanistic reasoning, particularly in the context of Alzheimer’s disease and neurodegeneration.

Our goal is to develop innovative AI methodologies that integrate diverse clinical data sources to enable earlier, more accurate, and clinically actionable disease diagnosis.