Multimodal Physiological Indices During Surgery Under Anesthesia

Abstract

Monitoring nociception, the flow of information associated with harmful stimuli through the nervous system even during unconsciousness, is critical for proper anesthesia care during surgery. Currently, this is done by tracking heart rate and blood pressure by eye. Monitoring objectively a patient’s nociceptive state remains a challenge, causing drugs to often be over- or under-dosed intraoperatively. Inefficient management of surgical nociception may lead to more complex post-operative pain management and side effects such as post-operative cognitive dysfunction, particularly in elderly patients. We collected a comprehensive and multi-sensor prospective observational dataset focused on surgical nociception (101 surgeries, 18,582 minutes, 49,878 nociceptive stimuli), including annotations of all nociceptive stimuli occurring during surgery and medications administered. Using this dataset, we developed indices of autonomic nervous system activity based on physiologically and statistically rigorous point process representations of cardiac action potentials and sweat gland activity. Next, we constructed highly interpretable supervised and unsupervised models with appropriate inductive biases that quantify surgical nociception throughout surgery. Our models track nociceptive stimuli more accurately than existing nociception monitors.

Background

General anesthesia consists of four simultaneous components: lack of pain processing or antinociception, unconsciousness, amnesia or lack of memory, and muscle relaxation or paralysis while maintaining physiologic stability [1]. In the case of general anesthesia, the unconscious processing of pain via primal reflexes at the level of the spinal cord and brainstem is termed nociception [2]. One of the primary purposes of general anesthesia is to prevent nociception during surgery [2-4]. The standard of care results in frequent underdosing/overdosing of intraoperative pain medication [2]. Underdosing can lead to the patient waking up in pain, delayed recovery and healing, postoperative cognitive dysfunction, and prolonged hospital stay. Overdosing can delay patient awakening, increase side effects such as nausea, vomiting, and constipation, exacerbate cognitive side effects such as delirium, and prolong hospital stay [2,4-5]. The bottom line is that improved management of surgical nociception can ensure antinociception during surgery while also reducing the prevalence of undesirable postoperative outcomes including cognitive impairment, postoperative pain [6], and nausea and vomiting.

A popular approach employed for tracking surgical nociception has been based on autonomic nervous system (ANS) responses. This is justified by the fact that general anesthesia disrupts conscious processing of pain, which leaves the brainstem and spinal cord as the primary processing centers of nociceptive inputs [1,7]. These regions govern the most primal reflexes to nociception, including autonomic responses (“fight or flight”) [1]. Therefore, autonomic outputs can be used to gauge the nociceptive state of the body, if the concurrent contributions of anesthetic and cardiac medications are also considered.

Several previous efforts have been made to define indices that track surgical nociception using autonomic responses. One of these is the Analgesic Nociception Index (ANI - now the HFVI), which uses a proprietary algorithm to extract information from the electrocardiogram (ECG) about heart rate variability (HRV) as a measure of cardiac autonomic activity [8].

This dataset contains not only autonomic indices during 101 surgeries, but annotations of when anesthetic drugs were administered or painful stimuli were present. In addition, it also includes ANI output for the same surgeries. 

Methods

In this study, we have collected prospectively a comprehensive observational dataset focused on surgical nociception, including over 100 subjects and recorded during 18,500 minutes of surgery. Using this dataset, we applied previously constructed indices and developed new models able to track nociception successfully throughout the course of surgery. The dataset includes annotations of the timing and type of almost 50,000 nociceptive stimuli, the timing and doses of multiple classes of anesthetic medications, and autonomic indices derived from physiological signals collected continuously throughout surgery (ECG, EDA). Our dataset includes a variety of drug classes across several types of surgery. The original data were collected using the Thought Tech Neurofeedback Expert System.

To extract the most physiologically relevant information from the acquired signals, we rely on statistically rigorous point process models that underlie the physiology of heartbeat dynamics and EDA that we have previously developed. We used one model for the generation of heartbeats, which allows for the computation of HRV indices, and another for sweat gland pulses, which cause EDA [9-13]. We feed the indices from these models into interpretable supervised and unsupervised frameworks with embedded inductive biases (physiological insight) to characterize surgical nociception. The supervised frameworks include logistic regression [14] and random forest [15], while the unsupervised framework is a state space model [16] assuming that the measurable observations (HRV and/or EDA indices) are driven by two hidden states related to the patient’s time-varying nociceptive state. Based on physiology, we postulate that there should be at least two components to the autonomic state (sympathetic and parasympathetic). Therefore, we hypothesized two hidden states. We also include information about the timing and dosage of anesthetic drugs in some of the supervised models.

Data Description

Statistics: 101 surgeries, 18,582 minutes, 49,878 nociceptive stimuli. Collected over a span of 3-4 years (from 2018-2022.

Data Format: The data is provided as both a Matlab structure titled OR_data.mat and as a zip folder of CSV files titled OR_data_CSVs. OR_data.mat is provided to feed directly into the accompanying code, while the CSVs replicate all of the same data in the Matlab structure for separate use in other applications.

.Mat formatted data

OR_data.mat is a 1x101 structure for each of the 101 surgeries.

• 101 distinct subjects/surgeries
• For each subject/surgery:
  • subj (#)
  • phys_ind_order: order of physiologic indices
  • W_5sec: substructure containing all the data
    • multiS_noDrug:
      • Fs: sampling rate (256)
      • w: 5 second window length
      • noc_stim: binary vector of whether or not there is a nociceptive stimulus in that time window
      • X_unnorm: unnormalized (in actual units) physiologic features for each window - there are 15 features and 15 additional features which are derivatives of the original 15 features
      • X_norm_noNan: normalized physiologic features for each window with Nans interpolated
    • ANI: Data from the other monitor during this surgery
      • Fs: sampling rate (256)
      • w: 5 second window length
      • noc_stim: binary vector of whether or not there is a nociceptive stimulus in that time window
      • X_unnorm: raw ANI index value for each window
      • X_norm_noNan: normalized ANI index value for each window
    • DrugInfo: information about drug timing and dosing administered
      • drugClass: the 9 classes of drugs tracked
      • timeSince: the time since the last dose of each class of drug was administered
      • cumulDose: the (normalized) cumulative dose of that class of drug up on board at that point

.csv formatted data

OR_data_CSV is a folder with CSVs of all of the vectors and matrices for each subject (noc_stim, X_unnorm, and X_norm_noNan for ANI and multiS_noDrug and timeSince and cumulDose) saved under the following file names for each subject:

• S###_W5sec_multiS_noDrug_noc_stim.csv
• S###_W5sec_multiS_noDrug_X_unnorm.csv
• S###_W5sec_multiS_noDrug_X_norm_noNan.csv
• S###_W5sec_ANI_noc_stim.csv
• S###_W5sec_ANI_X_unnorm.csv
• S###_W5sec_ANI_X_norm_noNan.csv
• S###_W5sec_DrugInfo_timeSince.csv
• S###_W5sec_DrugInfo_cumulDose.csv

The fixed parameters, including Fs (sampling rate), w (window size), phys_ind_order (order of physiologic indices), and drugClass (order of 9 classes of drugs tracked) are not saved individually because they are the same for all surgeries. Their values are below:

• Fs: 256 Hz
• W: 5 seconds
• Phys_ind_order (first all indices, then all derivatives): 'MuHR', 'SigmaHR', 'TotalPower', 'HF', 'LFnu', 'HFnu', 'LFnu_noVLF', 'LF/HF', 'TonicEDA', 'MuAmp', 'SigmaAmp', 'MuPR_LogN', 'SigmaPR_LogN', 'MuPR_IG', 'SigmaPR_IG'
• drugClass: 'Sedative', 'Antinoc', 'MuscleRelaxant', 'Pressor', 'BetaBlocker', 'Alpha2Ag', 'NMDA-Ant', 'Hydralazine', 'Sevo'

Associated code files (Matlab):

The Matlab scripts takes OR_data.mat as direct input.

• LSSMspxem.zip
  • State space modeling package (pre-existing open source code [17])
  • Linear Gaussian state space modeling
• LogReg_modeling_nociception.m
  • Logistic regression modeling with leave-one-subject-out cross-validation
• RandomForest_modeling_analysis.m
  • Random forest modeling with leave-one-subject-out cross-validation
• StateSpace_modeling_initialization.m
  • Initialization step of state space modeling - running each individual subject backwards and saving the initial steps
• StateSpace_modeling_nociception.m
  • Training state space model after initialization - implemented leave-one-subject-out cross-validation with state space models
• Compute_ANI_performance.m
  • Compute performance of the comparison monitor
• Replicate_ANI_study.m
  • Replicate methodologies from other studies for the comparison monitor data

Usage Notes

This data has been used to design new algorithms for tracking nociception in the operating room (upcoming publication). It can be reused to understand the variation in different anesthetic strategies, as well as do secondary post-hoc analysis relating anesthetic strategy/drug choice to physiologic effects. It can also be used to further stratify the patients based on nociceptive response/anesthetic strategy etc and characterize antinociception in more detail.

Known limitations that users should be aware of include that as with all of pain research, there is no objective ground truth. We never know the absolute ground truth of when a patient actually experienced nociception; the annotations of nociceptive stimuli are a proxy and usually a superset. In addition, all manual annotations are subject to human error, especially with respect to exact time.

To better understand the electrodermal activity models that were used, as well as other examples of this kind of autonomic analysis, we suggest reviewing our perviously published datasets [18, 19, 20]

Release Notes

This is version 1.0.0 (first release)

Ethics

The study protocol was approved by the Massachusetts General Hospital (MGH) Human Research Committee (Protocol 2017P002591). The Human Research Committee is the Institutional Review Board for MGH. All participants provided written informed consent.

Acknowledgements

We would like to thank the Department of Anesthesia at Massachusetts General Hospital and the many anesthesiologists who helped with our study.

This work was generously supported by the JPB Foundation; the Picower Institute for Learning and Memory; George J. Elbaum (MIT '59, SM '63, PhD '67), Mimi Jensen, Diane B. Greene (MIT, SM '78), Mendel Rosenblum, Bill Swanson, Cathy and Lou Paglia annual donors to the Anesthesia Initiative Fund. This research was also supported by a NSF Graduate Research Fellowship to S.S. and an MIT Office of Graduate Education Collamore-Rogers Fellowship to S.S.

Conflicts of Interest

The authors have no conflicts of interest to declare.

References

1. E. N. Brown, R. Lydic, and N. D. Schiff. General anesthesia, sleep, and coma. New England Journal of Medicine, 363(27):2638–2650, December 2010.
2. E. N. Brown, K. J. Pavone, and M. Naranjo. Multimodal general anesthesia. Anesth Analg, 127:1246–1258, 2018.
3. J. P. Desborough. The stress response to trauma and surgery. Br J Anaesth, 85(1):109–117, July 2000.
4. E. N. Brown, L. A. Santa Cruz, and S. Subramanian. Multimodal general anesthesia in practice. Anesthesiology News Special Edition, 2019.
5. H. Kehlet and J. B. Dahl. The value of ”multimodal” or ”balanced analgesia” in postoperative pain treatment. Anesth. Analg., pages 1048–1056, 1993.
6. L. A. Santa Cruz Mercado, R. Liu, K. M. Bharadwaj, et al. Association of Intraoperative Opioid Administration With Postoperative Pain and Opioid Use. JAMA Surg. 2023, Jun 14. doi:10.1001/jamasurg.2023.2009
7. E. N. Brown, P. L. Purdon, and C. J. Van Dort. General anesthesia and altered states of arousal: a systems neuroscience analysis. Annu Rev Neurosci, 34:601–628, 2011.
8. R. Logier, M. Jeanne, B. Tavernier, and J. De jonckheere. Pain / analgesia evaluation using heart rate variability analysis. 2006 International Conference of the IEEE Engineering in Medicine and Biology Society, pages 4303–4306, 2006.
9. R. Barbieri, E. C. Matten, A. A. Alabi, and E. N. Brown. A point-process model of human heartbeat intervals: new definitions of heart rate and heart rate variability. Am. J. Physiol. Heart Circ. Physiol., 288(1):H424–435, January 2005.
10. S. Subramanian, R. Barbieri, and E. N. Brown. Point process temporal structure characterizes electrodermal activity. PNAS, 117(42):26422–26428, October 2020.
11. S. Subramanian, P. L. Purdon, R. Barbieri, and E. N. Brown. Elementary integrate-and-fire process underlies pulse amplitudes in electrodermal activity. PLOS Computational Biology, 2021.
12. D. Daley and D. Vere-Jones. An Introduction to the Theory of Point Processes. Springer, New York, 2003.
13. D. Daley and D. Vere-Jones. An Introduction to the Theory of Point Processes: Volume II: General Theory and Structure. Springer, New York, 2007.
14. C. R. Shalizi. Advanced Data Analysis from an Elementary Point of View, chapter 11, pages 234–261. Cambridge University Press, 2017.
15. L. Breiman. Random forests. Machine Learning, 45:5–32, January 2001.
16. R. H. Shumway and D. S. Stoffer. Time Series Analysis and Its Applications: With R Examples, chapter 6, pages 319–404. Springer, New York, NY, third edition, 2011.
17. Sebastien Blais (2017). EM algorithm for linear state-space models (https://www.mathworks.com/matlabcentral/fileexchange/64585-em-algorithm-for-linear-state-space-models), MATLAB Central File Exchange. Retrieved April 16, 2020.
18. Subramanian, S., Purdon, P., Barbieri, R., & Brown, E. (2021). Electrodermal Activity of Healthy Volunteers while Awake and at Rest (version 2.0). PhysioNet. https://doi.org/10.13026/136t-7g98.
19. Subramanian, S., Purdon, P., Barbieri, R., & Brown, E. (2021). Behavioral and autonomic dynamics during propofol-induced unconsciousness (version 1.0). PhysioNet. https://doi.org/10.13026/2rbc-1r03.
20. Subramanian, S., Purdon, P., Barbieri, R., & Brown, E. (2021). Pulse Amplitudes from electrodermal activity collected from healthy volunteer subjects at rest and under controlled sedation (version 1.0). PhysioNet. https://doi.org/10.13026/r9p1-bk90.