Meet The Speaker
“Quickest Detection of Hidden Clinical and Behavioral State-Transitions: an Optimal Control Approach”
Sabato Santaniello is a postdoctoral fellow in the department of Biomedical Engineering and the Institute of Computational Medicine at the Johns Hopkins University. He received a M.S. degree in computer engineering from Università di Napoli “Federico II” (Italy) in 2004, and a Ph.D. degree in information and systems engineering from Università del Sannio (Italy) in 2007. His research interests are in systems and control and computational neuroscience with application to the study of deep brain stimulation, Parkinson’s disease, and epileptic seizure prediction.
Seminar Abstract
“Quickest Detection of Hidden Clinical and Behavioral State-Transitions: an Optimal Control Approach”
Accurately detecting hidden clinical or behavioral states from sequential measurements is an emerging topic in neuroscience and medicine, with a potentially dramatic impact in the field of neural prosthetics, brain-computer interface and drug delivery. For example, it is critical for clinical intervention to quickly detect when an epilepsy patient transitions from a benign (seizure-free) state to a problematic (pre-seizure) state from intracortical EEG recordings. As well, predicting an approaching ventricular fibrillation event from ECG measurements even a few seconds in advance could potentially saves lives.
We present a novel paradigm which employs optimal control to design a detection policy that precisely minimizes average delay and probability of false alarm. This “quickest detection” (QD) problem is posed by introducing a cost function which trades off detection delay and probability of false alarm. QD has been previously solved when the measurements are assumed independent, but this assumption is not valid for patho-physiologic activity where temporal dependencies exist. Here, we (i) extend the QD theory for dependent measurements, (ii) relax the constraints on the cost function to allow for nonlinear functions of the delay, and (iii) solve the QD optimization via dynamic programming.
Three clinical applications are provided: 1) detection of the movement onset in Parkinson’s disease patients from sequential recordings of subthalamic neurons (spike trains); 2) prediction of an approaching arrhythmia event (either ventricular tachycardia or fibrillation) from measures of beat-to-beat variability in ECG recordings of resting CAD (Coronary Artery Disease) patients; 3) prediction of an approaching epileptic seizure from thalamocortical and hippocampal field potentials in freely moving rats treated with chemoconvulsant pentylenetetrazol. For each application, we show that, with our approach, the average delay and the probability of false alarms are significantly lower (30-50% less) than a chance level predictor and a standard Bayesian estimator.
