This research seeks to develop novel machine learning algorithms that enable real-time video and sensor data analysis on large data streams given limited computational resources. The work focuses on healthcare as an application domain where real-time video analysis can prevent user-errors in operating medical devices or provide immediate alerts to caregivers about dangerous situations. The research will develop algorithms to automatically adapt data analysis approaches to maximize accuracy of analysis within a short time period despite limited available computing resources. Today's healthcare environment is significantly more technologically sophisticated than ever before. Many medical devices are now frequently used in patient's homes, ranging from simple equipment such as canes and wheelchairs to sophisticated items such as glucose meters, ambulatory infusion pumps and laptop-sized ventilators. The rapidly growing home health industry raises new safety concerns about devices being used inappropriately in the home setting. The proposed research is designed to reduce medical device related use-errors by developing computational algorithms that perform real-time video analysis and alert the patient or caregiver when medical devices are not used appropriately. The real-time video and sensor data analysis is also critical to the healthcare systems that monitor the activities of the elderly or those with disabilities in order to allow a caregiver to react immediately to an incident.
New machine learning theories and algorithms will automatically adapt to hardware limitations, with the aim to learn from a large number of training examples, a prediction function that (i) is sufficiently accurate in making effective predictions and (ii) can be run efficiently on a specified computer system to deliver time critical results. Three types of prediction models are studied to address the problem of automatic hardware adaptation, including a vector-based model, a matrix-based model, and a prediction model based on a function from a Reproducing Kernel Hilbert Space (RKHS). A general framework and multiple optimization techniques are being developed to learn accurate prediction models that match limited memory and computational capacity. The new learning algorithms will be evaluated in several medical scenarios through real-time prediction of a patient's activities from observations in the large video archives collected by several healthcare related projects. The intellectual merit of the proposed work is in bridging the gap between the high complexity of a prediction model and limited computational resources, a scenario that is encountered in many application domains besides healthcare. The proposed research in machine learning algorithms and theories will make it possible to run complicated prediction algorithms on big data within the limitation of a given computing infrastructure. The developed techniques for automatic hardware adaptation will be applied to a large dataset of continuous video and sensor recordings for medically-critical activity recognition. The project's broader impacts include providing medical experts with algorithms and tools supporting novel approaches to analyzing observational data in their quest to recognize and characterize human behavior. Surveillance systems with continuous observations will be able to categorize salient events with co-located, limited hardware. Researchers with complex data from continuous streams will be able to explore their domains with greater accuracy within constrained time using their available computing resources. Similarly, large archives can be exploited as rapidly as possible with limited hardware.
