Bobak Jack Mortazavi, PhD

My research goals are to take common statistical pattern recognition techniques, found in many computer science and electrical engineering applications, and apply them to clinical data sets. The goal of such an application is to use advanced analytics to develop new health metrics and interventions based upon relationships previously unknown to researchers. By developing personal monitoring systems, it may be possible to intervene before illnesses reach a point of hopsitalization, improving patient care and reducing economic burden of such care.

My research began in traditional computer science systems. This included architecture development for processing elements as well as Field Programmable Gate Array (FPGA) reconfigurable networks for multimedia and smart home environments. These FPGAs allow for the design and prototype of reconfigurable hardware systems ideal for signal processing. From this I moved into the embedded systems realm, applying my electrical engineering hardware background with my computer science math and algorithms background to develop full working systems. These embedded systems are application oriented, where each end product/goal introduces new challenges to the hardware platforms needed as well as the software algorithms that run on those platforms or the data produced by them.

With the advent of small, wearable sensors, it became possible to apply reserach in sensor networks to the human through the use of Body Sensor Networks (BSN). These BSN allow for the development of human-oriented embedded systems, with a wide range of possible applications, many of which span the realm of health care. My Ph.D. focussed on such a system for physical activity monitoring. By placing multiple inertial measurement units (IMUs - sensors containing accelerometers, gyroscopes, and magnetometers) on various parts of the body, it becomes very simple to collect lots of time-series data on human motion. However, determining more detailed specifics of such motion are often more complicated. Many traditional human activity goals look at either developing a simple pedometer application or identifying common cyclical patterns (such as walking, climbing stairs, etc.). These patterns use common supervised learning techniques to identify a type of motion over a period of time. What if, however, you wanted to know detailed movements, such as those in sports, and as they were happening?

My research in physical activity monitoring aimed at focusing on such a question. In particular, could you monitor sports motinos, where the actions are often singular intense actions with little to no repetition, and apply such a recognition system to be the input to a new video gaming system? If so, you could develop an exercise-based game (called exergaming) that would allow for users to play games with increased energy expenditure. I looked at all the challenges in developing such a system, applied to a soccer gaming environment, from the sensor design for ubiquitous monitoring, to the machine learning algorithm at the core of the work. I analyzed the feature extraction and classificaiton techniques necessary to make the recognition of fine-grain activities possible. Furhter, I investigated the trade-offs necessary in a gaming environment, which include possibly sacrificing accuracy in an attempt to arrive at a classification in a timely fashion.

The application of such algorithms and systems can, however, be extended beyond the exergaming environment. My research goals at Yale include applying such supervised and unsupervised learning techniques to medical data in an attempt to develop better risk models for hospital re-admissions, with an eye toward cardiology and heart failure. Such data analytic techniques may be able to identify patterns for risk models not known through common biostatistic methods.

Machine learning and Systems Development for Medical/Clinical Data and Applications