Quantitative assessment of the relationship between behavioral and autonomic dynamics during propofol-induced unconsciousness

Figures

• 

  Fig 1. Schematic of incorporating longer-range history using lagged windows. This figure shows how lagged windows for each index were added to incorporate longer-range history. For example, if w is 30 seconds and h is 120 seconds, that results in M = 4 lagged windows for each index, resulting in a total of 5 features in the model from each index.

• 

  Fig 2. EDA indices from Subject 4. From top to bottom, mean pulse rate, mean pulse amplitude, correlation between mean pulse rate and pulse amplitude, standard deviation of pulse rate, standard deviation of pulse amplitude, and correlation between standard deviations of pulse rate and pulse amplitude. This subject shows several periods of high correlation between rate and amplitude information. However, at approximately 70–90 minutes, pulse rate and amplitude information become strongly anti-correlated. This is because mean and standard deviation of pulse rate remain elevated while mean and standard deviation of pulse rate decrease. This indicates lingering sympathetic tone in the form of sparser but still large pulses.

• 

  Fig 3. EDA indices from Subject 6. From top to bottom, mean pulse rate, mean pulse amplitude, correlation between mean pulse rate and pulse amplitude, standard deviation of pulse rate, standard deviation of pulse amplitude, and correlation between standard deviations of pulse rate and pulse amplitude. This subject shows both periods of high correlation and high anti-correlation between pulse rate and amplitude information. Specifically, from 60–100 minutes, they go from highly correlated to highly anti-correlated. This reflects a “handoff” of information, going from both rate and amplitude to largely rate alone. In this subject, the rate of pulses is heavily modulated, while most pulses are small except for short periods when pulse amplitude increases sharply.

• 

  Fig 4. EDA indices from Subject 7. From top to bottom, mean pulse rate, mean pulse amplitude, correlation between mean pulse rate and pulse amplitude, standard deviation of pulse rate, standard deviation of pulse amplitude, and correlation between standard deviations of pulse rate and pulse amplitude. This subject has few periods where pulse rate and amplitude information are highly correlated or highly anticorrelated. They are mostly decorrelated, with the correlation hovering around 0. In this case, pulse rate and amplitude are largely complementary; excluding either would exclude valuable information.

• 

  Fig 5. Autonomic predictions by subject from the best logistic regression model for Question 1. For each subject, the blue trace is the prediction of the model using autonomic indices, while the magenta asterisks indicate the state of consciousness based on behavior, where conscious is 0 and unconscious is 1.

• 

  Fig 6. Autonomic predictions by subject from the best logistic regression model for Question 2. Fig 6 allows for better visualization of the nuanced autonomic dynamics immediately around loss of consciousness that were not visible in Fig 5. With these nuanced autonomic, it is clear that the autonomic shifts are rarely aligned exactly with behavioral changes. While it is difficult to annotate behavioral changes on a continuum, autonomic dynamics allow for exactly that. Subject 4 shows the best alignment between autonomic and behavioral changes. Subjects 2, 3, and 6 exhibit delayed autonomic shifts compared to behavioral changes, while Subjects 5, 8, and 9 exhibit earlier autonomic changes compared to behavioral changes. Subjects 1 and 7 exhibit a more complex and variable dynamic that may have also been affected by phenylephrine administration around the time of loss of consciousness.

• 

  Fig 7. Autonomic predictions by subject from the best logistic regression model for Question 3. Unlike in Figs 5 and 6, with respect to the behavioral markers, unconsciousness is 0 and consciousness is 1.