Embedded biomedical systems for assistive technology and state-of-vigilance forecasting
Open Access
- Author:
- Graybill, Philip
- Graduate Program:
- Electrical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 16, 2022
- Committee Members:
- Vijaykrishnan Narayanan, Major Field Member
Minghui Zhu, Major Field Member
Bruce Gluckman, Co-Chair & Dissertation Advisor
Mehdi Kiani, Co-Chair & Dissertation Advisor
Andrew Geronimo, Outside Unit & Field Member
Kultegin Aydin, Program Head/Chair - Keywords:
- embedded systems
biomedical
microcontroller
FPGA
assistive technology
unscented Kalman filter
state of vigilance
sleep
forecasting
eyelid
optimization
wearable
epilepsy - Abstract:
- Embedded systems have become widespread in biomedical applications. They offer the benefits of portability or wearability and the constant monitoring of physiological signals. However, the design of embedded biomedical systems also poses challenging design constraints related to the limited energy budget, computing power, and memory inherent in embedded systems. These challenges can be approached through the exploration of tradeoffs between computational complexity and system output accuracy. Here, we describe design efforts for two embedded systems for biomedical applications that explore these tradeoffs. First, the design and testing of the Eyelid Drive System (EDS) is detailed. The EDS is a hands-free assistive technology that uses inductive sensing to detect patterns of blinks and winks to control other electronic devices. The EDS comprises a pair of glasses outfitted with a printed circuit board, two pairs of transmitting coils wound around the frames, and a pair of passive receiver coils that are adhered to the user’s eyelids. The EDS offers a unique combination of advantages over existing eye- and eyelid-based assistive technologies: it requires neither ambient light nor exaggerated motions; is able to distinguish among up to eight different commands; demonstrates relatively high accuracy and information transfer rate; avoids the use of electrooculographic electrodes; and is portable, wearable, and inexpensive. Trials with human subjects yielded a group mean accuracy of 96.3% for a set of four different commands at a response rate of 3 s. A mean information transfer rate (ITR) of 56.1 bits/min over all subjects was achieved with a set of six different commands at a response rate of 1.5 s. Machine learning algorithms were applied to a database of difficult-to-classify EDS signals and assessed for computational complexity and classification accuracy. We also describe design efforts toward an embedded state-of-vigilance (SOV) forecasting system that employs an unscented Kalman filter. Data assimilation refers to methods used to synchronize a dynamical model to sparse or noisy measurements related to model states. The synchronized model can then be used to gain insight into the modeled system or make forecasts of future system states. The unscented Kalman filter (UKF) is one such data assimilation method that can be applied to nonlinear models. As a first step toward designing a UKF-based system for an embedded platform, we developed a generalizable method for doing so. The method seeks to minimize both computation time and error in UKF state reconstruction and forecasting. As a case study, we applied the method to a model for the rat sleep-wake regulatory system in which 432 variants of the UKF over six different variables are considered. The optimization method is divided into three stages that assess computation time, state forecast error, and state reconstruction error. We apply a cost function to variants that pass all three stages to identify a variant that computes 27 times faster than the reference variant and maintains required levels of state estimation and forecasting accuracy. We draw the following insights: 1) process noise provides leeway for simplifying the model and its integration in ways that speed computation time while maintaining state forecasting accuracy, 2) the assimilation of observed data during the UKF correction step provides leeway for simplifying the UKF structure in ways that speed computation time while maintaining state reconstruction accuracy, and 3) the optimization process can be accelerated by decoupling variables that directly impact the underlying model from variables that impact the UKF structure. Finally, we describe the optimization of a UKF-based forecasting system for the prediction of transitions into REM sleep in a behaving rat. We first establish a framework for assessing the REM transition forecasts. This assessment framework includes random forecasters against which the system under design is assessed, rules for making positive and negative forecasts, and a cost function for quantifying forecaster performance. Next, we detail the optimization of the UKF forecaster over multiple design options, including the magnitude of the entries in the process noise (covariance inflation) matrix, the magnitude of the entries in the observation noise matrix, and the number of REM-positive sigma points required to forecast a transition. Finally, we compare the performance of the optimal UKF forecaster and random forecasters on four days of physiological recordings. The UKF forecaster achieves a lower cost function than all of the random forecasters and achieves levels of precision and sensitivity suitable for an in vivo stimulation experiment.