Valeriy Gavrishchaka

West Virginia University, Morgantown, WV, USA

Abstract Title: Multi-expert indicators for detection of emerging psychophysiological changes using low-resolution data from wearables and personalized optimization of lifestyle, training and rehabilitation strategies

Biography: Valeriy Gavrishchaka received his MS and PhD degrees in computational and theoretical physics from Moscow Institute of Physics and Technology (Russian Federation) and from West Virginia University (USA), respectively. He has 30 years of academic and industry experience in complex systems research and applications including fundamental and applied solutions in plasma / space physics based on physics models / simulations and wide range of machine learning (ML) approaches, quantitative models and AI/ML frameworks for algorithmic trading, market / credit risk analytics and biomedical applications. He is an author of more than 80 publications in multi-disciplinary scientific journals and conference proceedings.

Research Interest: Capability for early detection of emerging abnormalities and slowly developing psychophysiological changes is critical for personalized multi-objective optimization of lifestyles, physical and psychological training protocols in sports, rehabilitation procedures, prevention therapies and treatments. Physiological signals can quantify personal dynamical state aggregating impact from all relevant factors and used in clinical diagnostics. Variability analysis of physiological signals, including heart rate (HR) and gait, are effective not only for diagnostics but also for early detection of emerging abnormalities. Beat-to-beat heart rate variability (HRV) or stride-to-stride gait variability analysis provide affordable means of detecting medical abnormalities and wide range of psychological / psychiatric conditions. Wearable systems for real-time collection of physiological data provide new opportunities for diagnostics and monitoring with capabilities of recording beat-to-beat (RR) time series with accuracy comparable to clinical electrocardiogram (ECG) equipment. While the best variability indicators need long time series for accuracy, early detection of emerging and intermittent psychophysiological changes require variability calculations on short time series. Previously, we have demonstrated that these challenges could be overcome by combining complementary variability indicators using boosting-like ensemble learning. Continuous 24/7 data collection with high resolution is not technologically feasible in most non-clinical devices and data are collected with much lower sampling rate. Accuracy preservation of variability indicators computed from such 24/7 low-resolution data is critical for early detection of emerging psychophysiological patterns which are often intermittent and could be easily missed when occasional short-term high-resolution recordings are used. Although accuracy of indicators is rapidly reduced with lowering data resolution, we demonstrated that accuracy problems can be resolved by framework where boosting-like algorithms are used for enhancement of existing variability indicators with further non-linear combination of ensemble components via deep learning or tree-based gradient boosting. Here we confirmed robustness of our indicators by adding gait data and wider range of psychophysiological states.