Hybrid Decoders for Marked Point Process Observations and External Influences

Area of Science:

• Physiological monitoring
• Biomedical signal processing
• Machine learning in healthcare

Background:

• Estimating internal physiological states from bioelectric and biochemical data is challenging.
• Existing methods struggle to incorporate external influences into state estimations.
• Cortisol and skin conductance are key indicators of physiological stress and arousal.

Purpose of the Study:

• To develop a novel hybrid estimator for physiological state estimation.
• To enable state estimates to be guided by external influences.
• To improve the accuracy of estimating cortisol-related energy production and sympathetic arousal.

Main Methods:

• Modification of a recurrent neural network (RNN) for state-space estimation.
• Implementation of a weighted cost-function for a hybrid estimator.
• Application to point process and continuous-valued physiological data.

Main Results:

• Demonstrated accurate estimation of energy production using cortisol data with external sleep-wake influences.
• Showcased reduced overfitting by incorporating circadian rhythm-based influences.
• Successfully estimated sympathetic arousal from skin conductance data during stress and relaxation experiments.

Conclusions:

• The RNN-based hybrid method effectively recovers internal physiological states.
• External influences can be integrated to guide state estimates from bio-observations.
• This approach enhances physiological monitoring capabilities.