I am a cardiologist and researcher in Division of Research at Kaiser Permanente Northern California. I am also an adjunct Assistant Professor at the Smidt Heart Institute at Cedars-Sinai. As a physician-scientist and statistician with focus on cardiology and imaging, I majored in statistics at Rice University, obtained my MD at UCSF, and received post-graduate medical education in internal medicine, cardiology, and a postdoc in computer science and biomedical data science at Stanford University. I am also a Deputy Editor at NEJM AI and Chair of the ImageGuideRegistry for the American Society of Echocardiography. Our mission is to improve cardiovascular diagnosis and disease detection using computer vision and artificial intelligence. Our group works on applications of deep learning, computer vision, and the statistical analysis of large datasets within cardiovascular medicine. We focus on AI model development (see section 'Precision Phenotyping'), AI model understanding (see section 'Interpretable Medical AI'), and AI model deployment and implementation (see section 'AI Clinical Trials'). We are interested in all forms of cardiovascular data, including ECG, MRI, radiological imaging, and particicularly interested in novel insights from pre-existing data. Our work has been published in Nature, Nature Medicine, NEJM AI, Circulation, JAMA Cardiology, EHJ, and other scientific venues. Our group works on multi-modal datasets, leveraging EHR, ECG, echo, and MRI data for a broad perspective on cardiovascular disease. Our lab members bring expertise from a wide range of backgrounds ranging from physics, mechanical engineering, computer science to cardiology, anesthesia, and internal medicine.

MRI is increasingly being used to detect prostate cancer, yet its interpretation can be challenging due to subtle differences between benign and cancerous regions. Many deep learning models have been developed to automatically detect prostate cancer, yet these methods can be further improved by incorporating useful anatomical information about prostate tumors (e.g. ~70% of prostate cancers occur in the peripheral zone). We quantified the heterogeneity of prostate cancer occurrence using a registration pipeline to compute voxel-level cancer distribution, and incorporated it as a prior distribution for a Denoising Diffusion Probabilistic Model (DDPM). Our method achieved statistically significantly higher accuracy and Dice coefficient than alternative models. As many AI methods have been developed for similar problems, a selective feature ensemble method is needed to combine their strength and eliminate incorrect predictions. We developed a multi-teacher distillation pipeline with a reinforcement-learning agent to dynamically determine distillation weights, thereby enriching the student model with useful knowledge and outperforming all teacher models.

The rigorous evaluation and stress-testing of clinical AI models is frequently constrained by a lack of diverse, precisely controlled data, especially for rare pathologies or subtle clinical edge cases. This talk presents a framework for generating high-fidelity, controllable 3D CT volumes by conditioning latent diffusion models on clinical text, anatomical segmentations, and demographics to create targeted evaluation datasets. We demonstrate this approach across two distinct domains: generating 3D chest CTs conditioned on radiology reports to simulate specific findings like pulmonary embolism and medical devices, and creating counterfactual head CTs to model varied manifestations of intracranial hemorrhage. By evaluating these synthetic volumes through downstream diagnostic tasks and sensitivity analyses, we show that synthetic data can effectively preserve clinically relevant signals and provide a path for the rigorous evaluation and stress-testing of AI models before clinical deployment.

 In current clinical practice, treatment decisions often rely on broad demographic factors and standardized markers that miss individual patient nuances. In this talk, I explore how uncertainty-aware causal deep learning—informed by medical images and demographics—can tailor treatments to a patient’s unique profile, improving the accuracy, fairness, and safety of personalized medicine. This framework is grounded in large-scale MRI data from randomized controlled trials for neurological disease treatments.

To ensure these personalized models are truly equitable and trustworthy, we must rigorously expose and mitigate hidden biases. I will highlight how fine-tuned Vision-Language Foundation Models (e.g. based on Stable Diffusion) offer interpretable, patient-specific explanations by generating precise medical image counterfactuals—powerful tools for uncovering and mitigating biases driven by spurious correlations. Building on this, I will present recent strategies for correcting calibration biases across population subgroups in Multimodal Large Language Models (MLLMs). Finally, I will offer a glimpse into the enormous potential of Agentic AI in advancing clinical care. Transitioning toward dynamic clinical agents holds immense promise for unlocking transparent reasoning and advanced explainability, ultimately  paving the way for a truly interactive and trustworthy era of clinical decision support.

Most healthcare AI research asks a narrow question: can the algorithm perform the task well? Increasingly, the answer is yes. 

But task performance is insufficient when that algorithm enters a living clinical system - one shaped by human cognition, professional identity, institutional culture, and trust.

This talk shares a case study within the NHS of how an AI entering a system succeeded and also failed (depending on your definition) - not because of the technology, but the people and structures it encounters. 

Healthcare AI's real risks are not technical - they are increasingly behavioural. Automation bias, deskilling, and the quiet erosion of clinical reasoning do not announce themselves; they accumulate. And we are not routinely detecting them.

Much of the work is early and speculative, but I make the case for reflecting on human factors evaluation for AI deployment.

Current AI in oncologic imaging largely reads single time-point scans, while clinical decisions depend on how disease evolves: how lesions change in size, shape, density, and distribution across serial imaging, and how that aligns with the rest of a patient's record. In this talk, I will give an overview of how my group is working to close this gap.

I will cover three connected research lines. The first develops methods for longitudinal tumor assessment, combining lesion detection, segmentation, and image registration into validated pipelines for treatment response, with multi-reader studies showing reduced reading time and lower inter-reader variability. The second builds shared benchmarks through challenges such as Learn2Reg, ULS23, MAST, and UNICORN. The third tackles the annotation bottleneck by learning efficiently from clinical routine data, from LLM-Extractinator for turning radiology reports into structured labels, to weak-supervision and self-learning strategies that reduce reliance on dense manual annotation.

Together, these directions aim at AI that captures not just what a scan shows, but how disease is evolving.

TBA

TBA

TBA