AI for Healthcare

Clinical data are complex, multimodal, and dynamic. Electronic health records, physiological waveforms, laboratory measurements, and imaging data together form a longitudinal portrait of patient health. We develop multimodal, real-time machine learning (ML) models that integrate these diverse data streams to assess patient risk and monitor disease progression.

A central focus of our lab is safe AI. Even highly accurate models are imperfect, and in medicine, a single incorrect decision can have serious consequences. We develop computational approaches that incorporate uncertainty quantification and confidence estimation at the individual patient level. By explicitly modeling predictive uncertainty, we aim to facilitate safer clinical deployment and responsible implementation.

Beyond predicting the patients’ outcome, our lab also is interested in causal ML models that estimate the potential effect of treatment strategies. Unlike traditional predictive models that forecast outcomes, causal models aim to support decision-making by evaluating what may happen under different interventions. This moves AI closer to personalized treatment planning and real-world clinical decision.

Our work is deeply embedded within the healthcare system through close collaboration with MaineHealth Maine Medical Center: 1) In the cardiovascular ICU, we are developing multimodal real-time models to monitor recovery following cardiac surgery (Figure 1). 2) In the medical ICU, we are building causal machine learning models to support treatment strategies for severe hyponatremia (Figure 2).

Through these translational partnerships, our lab seeks to bridge the gap between algorithm development and bedside implementation.