State-space Decoders for Wearable Healthcare Applications

Homeostatic processes govern multiple latent state variables within the human body. While many of these states remain largely unobserved, they frequently give rise to bioelectric and biochemical phenomena that can be measured. The measured signals provide a window into estimating the unobserved states. In a number of instances, the observed electrical and chemical phenomena are pulsatile or impulse-like in nature. This dissertation describes state-space methods for estimating latent variables tied to changes in skin conductance, heart rate and cortisol secretion – all of which have a characteristic point process nature. In the first two sections, state-space methods are developed for estimating sympathetic arousal from skin conductance and heart rate features. Estimation involves Bayesian filtering applied within an expectation-maximization framework. Results are provided on experiments involving different types of mental stressors and Pavlovian fear conditioning. The results agree with general expectations with high arousal levels typically occurring during stressors and lower values occurring during relaxation. General agreement with expectations is also found with different trial averages in fear conditioning. Skin conductance-based estimates are also validated with blood flow signals in the brain in a separate experiment. In the third section, state-space methods are developed to estimate energy production from blood cortisol measurements. The methods are applied to simulated and experimental data from patients suffering from Cushing's disease, chronic fatigue syndrome and fibromyalgia syndrome. The results help shed light on why patients with hypercortisolism may experience daytime fatigue and nighttime sleeping difficulties. Circadian-like behavior is also seen with higher energy estimates occurring towards morning awakening and lower values at bedtime. In the final section, machine learning methods are used for state-space estimation. Traditional Bayesian filtering methods do not have the ability to permit external influences such as domain knowledge or labels to affect the state estimates. We develop a hybrid estimator that enables this possibility and apply it to both skin conductance-based arousal estimation and cortisol-related energy estimation. The hybrid estimator permits the enforcement of circadian rhythms on to the state estimates and the customization of the level to which the external influence is permitted to affect them.