Research

A central theme of my research is development of mathematical and computational tools that enable a mechanistic understanding of complex diseases, and allow for phenotyping of patients, with the goal of monitoring disease progression and titration of therapies. To accomplish these, my research brings together concepts and tools across signal processing, information theory, control theory, optimization, and machine learning to design physiologically interpretable models and predictive analytic algorithms. A major focus of my postdoctoral research has been in the intensive care unit (ICU) where I have been developing advanced machine learning algorithms capable of meaningfully summarizing large volumes of continuously measured patient data, with the goal of timely prediction of potentially life threatening clinical events and early risk assessment.

Brief Description of Research Interests

Working with Professor Ryan Adams at Harvard, I recently developed a novel machine learning algorithm for automatic discovery of a collection of ‘phenotypic’ dynamical patterns in a database of patient time series. These dynamics, which capture the directional couplings among the vitals signs, can be then correlated with short-term (onset of hypotension, positive blood culture, and sepsis) and long-term (survival vs. mortality) outcomes, arriving at a notion of ‘healthy’ and ‘unhealthy’ dynamics. To accomplish this, we showed it is possible to transform a switching state-space model of the underlying physiology (and more generally, any dynamic Bayesian network) to an equivalent deep recurrent neural network architecture for supervised pattern recognition. This approach has two main advantages. First, it allows for a ‘model-based’ (as apposed to ‘black-box’) approach to supervised pattern discovery in multivariate physiological time series. Second, it enables a principled approach for guiding medical interventions to restore healthy dynamics via application of optimal control methodologies. The working hypothesis of this research is that guiding patients’ trajectory (via pharmacological interventions) towards healthy dynamics provides more effective treatment strategies, and results in faster recovery and better long-term outcomes. The ultimate goal of this research is to change the current standard of care by guiding interventional strategies based on data-driven severity of illness scores, and model-based assessment of mechanisms underlying physiological deterioration. 

Coming soon ...