Hybrid Decoders for Marked Point Process Observations and External Influences

Abstract

Objective: Internal physiological processes govern multiple state variables within the human body. Estimating these from point process-type bioelectric and biochemical observations is a challenge. Here we seek to estimate cortisol-related energy production and sympathetic arousal based on point process and continuous-valued data while permitting an external influence to affect the state estimates. Methods: Traditional point process state-space methods, such as those used for estimating the aforementioned quantities from cortisol and skin conductance measurements respectively, suffer from the inability to permit the state estimates to also fit to an external influence (e.g. labels) or be guided by it. Here we modify an existing recurrent neural network (RNN) approach for state-space estimation through a weighted cost-function to enable a hybrid estimator that has this capability. Results: Results on cortisol data based on a hypothetical sleep-wake influence term show how energy production can be estimated by permitting the estimates to fit to the external influence as much as desired. We further show how overfitting may be reduced by using circadian rhythm-based influence terms. Results on skin conductance data also indicate how the method can be used to estimate sympathetic arousal in an experiment containing stressors and relaxation, and permit an external influence as well. Conclusion: The RNN-based hybrid method is thus able to recover internal physiological states from point process and continuous-valued observations while permitting an external influence to guide the estimates. Significance: The hybrid estimator could be embedded within wearable monitors that can be tailored based on domain expertise or individual feedback.