Current Fellows
Objective and Equitable Identification of Delirium: An Artificial Intelligence Approach Based on Big Clinical Data (EEG & Video)
Delirium is an acute neuropsychiatric disorder characterized by disturbances of attention and awareness that impacts more than 1 in 5 hospitalized older adults. Delirium causes substantial burden and distress for patients and their caregivers and is strongly associated with poor outcomes, including long-term cognitive decline, prolonged hospitalization, and significantly increased mortality. While delirium can be managed and even prevented in 40% of patients with appropriate assessments, it remains routinely underdiagnosed across diverse clinical conditions. Efforts to identify delirium more consistently have created standardized interview-based tools such as the Confusion Assessment Method (CAM). Yet such tools are still limited by their intrinsic subjectivity, inter-rater variability, intermittent application, and are time-consuming. Additionally, these tools are challenging to use in certain contexts, such as with non-English-speaking patients. We have found that Spanish-speaking patients are 5.2 times less likely to be fully screened for delirium compared to English-speaking patients. Similarly, patients of color are less likely to receive proper assessments, despite facing higher risks. This highlights an urgent need for equitable and accurate methods to identify delirium across diverse clinical populations. Therefore, our goal is to develop novel, highly sensitive and generalizable tools to assess delirium objectively and equitably. To achieve this, I will leverage our lab’s prior work to develop 1) quantitative electroencephalography (qEEG)-based and 2) video-based tools, combined with machine learning and artificial intelligence (AI) techniques, such as human pose estimation, to detect and track the development of delirium. By focusing on diverse patient groups—including individuals from varied racial, ethnic, and language backgrounds—and employing non-language-based methods such as neurophysiological and behavioral evaluations through EEG and video analysis, we can provide a unique resource to objectively assess delirium. This approach can ensure high generalizability across diverse populations and aims to improve clinical care while addressing healthcare disparities.
Automated Assessment of Selective Motor Control in Preterm Infants Using Computer Vision
Cerebral palsy (CP) is the most common motor disorder in infants born preterm, often caused by early injury to the developing corticospinal tracts (CST). Early injury to the CST causes impaired selective motor control (SMC) in children with CP, affecting their ability to isolate movement at one joint at a time. While impaired SMC is the greatest contributor to gross and fine motor ability in children and adults with CP, little is known about the development of SMC in infants with CP, and no quantitative tools exist to measure this construct. In this project, we will use human pose estimation algorithms applied to video recordings of spontaneous infant movements to develop a reliable method for determining joint kinematics and SMC in preterm infants. Our overall objective is to create a machine learning tool which automatically quantifies and characterizes early SMC behavior to produce translatable and clinically actionable insights. In the long term, this work will improve the outcomes of children with CP by providing more timely and targeted interventions.
Developing Genetic Models of Childhood Hyperphagia and Obesity
Bardet-Biedl syndrome (BBS) is a rare, autosomal recessive disorder. BBS proteins are involved in the function of cilia, cellular organelles that are essential for cellular signaling. Obesity, polydactyly, retinitis pigmentosa, renal anomalies, and learning difficulties are among the main features of BBS, wherein obesity starts in early childhood. Obesity in BBS has been attributed to the melanocortin 4 receptor (MC4R) pathway in the hypothalamus, which plays a major role in appetite control and energy balance. BBS proteins in the proopiomelanocortin (POMC) neurons activate MC4R neurons via the leptin receptor to initiate satiety. However, the contribution of ciliary dysfunction to obesity remains poorly understood. The goal of this project is to understand how BBS proteins play a role in hyperphagia and obesity using three complementary approaches in zebrafish: (1) monitoring the eating behavior; (2) quantifying lipid accumulation; and (3) characterizing relevant transcriptional changes. By leveraging these tools, we aim to (1) Determine whether different BBS gene mutants display variable leptin signaling and (2) Use zebrafish BBS models to screen new drugs. This work will inform further the understanding of MC4R disruption caused by BBS genes and could be used as a therapeutic testing platform.
Decoding Cellular Logic Through Virtual Perturbation Screens: An AI-Driven Approach to Complex Disease
Complex diseases such as cancer and immune dysfunction emerge from high-dimensional interactions among genes, pathways, and cellular environments. Yet, despite the explosion of single-cell datasets, we still lack mechanistic models that can predict how specific genetic or chemical changes drive transitions between healthy and diseased cell states. My research develops virtual functional screens that integrate statistical physics, machine learning, and genomics to infer the causal logic of these molecular changes directly from large-scale gene expression data. Using modern AI and physics frameworks, I model how genetic and chemical changes (perturbations) reshape cellular states, revealing synergistic and antagonistic interactions that underlie cancer therapy resistance and phenotypic reprogramming. This framework treats perturbations and the gene expression changes they affect as training data, enabling prediction of unseen combinations and identification of driver genes that best explain new observed cellular transitions—for example, from naïve to resistant states in cancer. By combining interpretable generative models with optimization and statistical inference, this approach provides a scalable, quantitative foundation for AI-driven virtual perturbation screens—allowing us to predict more effective interventions, and ultimately bridge the gap between molecular perturbations and complex disease phenotypes.
