A Brief Overview of Burst Suppression: Causes and Effects on Mortality in Critical Illness

Home , Burst suppression

Brigham Young University BYU ScholarsArchive

Undergraduate Honors Theses

2020-03-20

A BRIEF OVERVIEW OF BURST SUPPRESSION: CAUSES AND EFFECTS ON MORTALITY IN CRITICAL ILLNESS

Jake Hogan Brigham Young University

Haoqi Sun Massachusetts General Hospital

Hassan Aboul Nour Henry Ford Health System

Stephen Piccolo Brigham Young University

Jacob Crandall Brigham Young University

See next page for additional authors

Follow this and additional works at: https://scholarsarchive.byu.edu/studentpub_uht

BYU ScholarsArchive Citation Hogan, Jake; Sun, Haoqi; Nour, Hassan Aboul; Piccolo, Stephen; Crandall, Jacob; Peck, Steven; and Westover, M Brandon, "A BRIEF OVERVIEW OF BURST SUPPRESSION: CAUSES AND EFFECTS ON MORTALITY IN CRITICAL ILLNESS" (2020). Undergraduate Honors Theses. 121. https://scholarsarchive.byu.edu/studentpub_uht/121

This Honors Thesis is brought to you for free and open access by BYU ScholarsArchive. It has been accepted for inclusion in Undergraduate Honors Theses by an authorized administrator of BYU ScholarsArchive. For more information, please contact [email protected], [email protected]. Author Jake Hogan, Haoqi Sun, Hassan Aboul Nour, Stephen Piccolo, Jacob Crandall, Steven Peck, and M Brandon Westover

This honors thesis is available at BYU ScholarsArchive: https://scholarsarchive.byu.edu/studentpub_uht/121 1 Honors Thesis 2 3 4 5 6 7 A BRIEF OVERVIEW OF BURST SUPPRESSION: 8 9 CAUSES AND EFFECTS ON MORTALITY IN CRITICAL ILLNESS 10 11 12 13 14 15 by 16 Jake Hogan 17 18 19 20 21 22 Submitted to Brigham Young University in partial fulfillment 23 of graduation requirements for University Honors 24 25 26 27 28 29 Biology Department 30 Brigham Young University 31 April 2020 32 33 34 35 36 37 Advisor: Stephen Piccolo 38 39 Honors Coordinator: Stephen Peck 40

42 ABSTRACT 43 44 45 46 47 A BRIEF OVERVIEW OF BURST SUPPRESSION: 48 CAUSES AND EFFECTS ON MORTALITY IN CRITICAL ILLNESS 49 50 51 Jake Hogan

52 Biology Department

53 Bachelor of Science in Bioinformatics

57 This thesis examines the effects of brain inactivation (burst suppression) on

58 mortality in mechanically ventilated ICU patients. Past research has associated burst

59 suppression with increased mortality. However, the effects of burst suppression on ICU

60 patients while taking into account the relative contributions of propofol (a sedative), and

61 critical illness to mortality, and whether preventing burst suppression might reduce

62 mortality, has not yet been quantified. This thesis explores this relationship, and the effect

63 of critical illness and propofol infusion on burst suppression to understand what drives

64 burst suppression.

65 To measure the relationship between burst suppression, critical illness, propofol

66 infusion, and mortality, we use a “counterfactual” randomized controlled trial. This is a

67 form of causal inference that allows one to estimate causal effects from retrospective

68 data. We create a causal graph and train independent multivariate regression models

69 using random forests, XGBoost, and logistic regression to estimate causal effects.

iii

70 These results show that patients with mild vs. severe illness are expected to have a

71 39% difference in burst suppression burden, 95% CI [8-66%], and a 35% difference in

72 mortality [29-41%]. Assigning patients to maximal (100%) burst suppression burden is

73 expected to increase mortality by 12% [7-17%] compared to 0% burden.

74 Our analysis clarifies how important factors contribute to mortality in ICU

75 patients. Burst suppression appears to contribute to mortality but is primarily an effect of

76 critical illness rather than propofol infusion. Thus, limiting burst suppression through

77 clinical intervention may be more difficult than previously thought.

80 ACKNOWLEDGEMENTS 81 82 A special thanks to my Principal Investigator Dr. Brandon Westover and my

83 mentor Dr. Haoqi Sun. Also, these members of my lab at Massachusetts General Hospital

84 are co-authors on an expanded manuscript which is published in Neurocritical Care.

86 Haoqi Sun, PhD, Hassan Aboul Nour, MD, Jin Jing, PhD, Mohammad Tabaeizadeh MD, 87 Maryum Shoukat, MD, Farrukh Javed, MD, Solomon Kassa, MD, Muhammad M. 88 Edhi,4, MD, Elahe Bordbar, MD, Justin Gallagher, BSc, Valdery Moura Junior, MSc, 89 Manohar Ghanta, MSc, Yu-Ping Shao, MSc, Oluwaseun Akeju, MD, Andrew J. Cole, 90 MD, Eric S. Rosenthal, MD, Sahar Zafar, MD, M. Brandon Westover, MD, PhD 91 92

vii

94 TABLE OF CONTENTS 95 Title ...... i 96 Abstract ...... iii 97 Acknowledgements ...... vi 98 Table of Contents ...... viii 99 Background Definitions ...... 1 100 Introduction...... 4 101 II. Methods ...... 5 102 II.I Ethics Approval ...... 5 103 II.II Patients ...... 8 104 II.III EEG Recording ...... 8 105 II.IV Measures ...... 8 106 II.V Effect Estimation ...... 9 107 II.VI Sensitivity Analysis ...... 9 108 II.VI Confidence Intervals...... 10 109 III. Results ...... 10 110 III.I Predicting Burst Suppression Burden and Mortality ...... 10 111 III.II Estimated Effects ...... 11 112 III.III Sensitivity Analysis ...... 12 113 IV. Discussion ...... 13 114 VI. Conclusions...... 15 115 Works Cited ...... 16 116 Supplementary Materials ...... 18 117

viii

118

119 LIST OF TABLES AND FIGURES 120 121 122 123 124 FIGURE 1. Normal electroencephalogram recording (EEG)...... 1 125 126 FIGURE 2 Burst-suppressed EEG...... 1 127 128 FIGURE 3 Example Causal Graph...... 3 129 130 TABLE 1 Patient Characteristics...... 5 131 132 FIGURE 4 Histogram of burst suppression burden...... 5 133 134 FIGURE 5 Causal graph with effect estimates...... 7 135 136 TABLE 2 Prediction performance using different machine learning models...... 9 137

138 Background Definitions

139 Clinical Background:

140 Electroencephalogram (EEG) Monitoring- EEG Monitoring is used to track 141 electrical brain activity. EEG recordings can determine whether a patient is undergoing a 142 seizure, determine which stage of sleep a patient is in, or detect the presence of unusual 143 phenomena such as burst suppression.

144 145

146 Figure 1. Example of a normal (not burst-suppressed) EEG1. Note the continual 147 oscillation of brain waves, which indicates brain activity.

148

149 Burst Suppression- a phenomenon in sedated, mechanically ventilated patients 150 where their brain undergoes short bursts of activity followed by complete suppression for 151 a few seconds. Some studies have shown that burst suppression increases patient 152 mortality. This is the claim I will investigate in my Honor’s Thesis.

153 154 Figure 2. Example of unintentional burst suppression in an EEG recording from a 155 mechanically ventilated ICU patient receiving propofol. There are two periods of higher 156 amplitude oscillatory EEG activity (bursts), surrounded by relatively flat periods 157 (suppressions).

158

159 Propofol- A sedative used for general anesthesia for patients undergoing surgery 160 and for sedation in the Intensive Care Unit (ICU).

161 Mechanical Ventilation- A breathing tube used for heavily sedated patients.

162 Status Epilepticus- a seizure lasting longer than 5 minutes or two seizures within 163 5 minutes without regaining consciousness. This constitutes a medical emergency2.

164 APACHE II, CCI, premorbid mRS- Four different overall measures of how 165 sick a patient is. Acute Physiologic Assessment and Chronic Health Evaluation II, 166 Charlson Comorbidity Index.

167 initial GCS- The degree of a patient’s comatose state at the beginning of their 168 Intensive Care Unit stay (Glasgow Coma Scale).

169 premorbid mRS- A measure of neurological disability (premorbid/before-disease 170 Modified Ranklin Score for Neurologic Disability).

171 172 Machine Learning:

173 Random Forests- A machine learning method that creates multiple decision trees 174 based on row and column subsets of a dataset. These trees form a “forest,” which votes 175 on each new observation to classify it into a certain category. For an intuitive, non- 176 mathematical understanding of decision trees and random forests, explore these videos:

177 https://www.youtube.com/watch?v=eKD5gxPPeY0 (decision trees)3, and 178 https://www.youtube.com/watch?v=J4Wdy0Wc_xQ&t=320s (random forests video)4.

179 Extreme Gradient Boosting- Tree-based model that creates simple models, 180 analyzes the errors in the model and creates another simple model to fix these errors. 181 These trees are combined to form a more complex, more highly predictive model.

182 Causal Inference:

183 Statistical Inference- The ability to say that two variables are associated with one 184 another (for example, the rooster crows right before the sun rises). However, the fact that 185 A and B are associated does not mean that A has a causal effect on B (for example, a 186 rooster crowing does not cause the sun to rise).

187 Causal Inference- The ability to say that a certain outcome (B) is due to another 188 variable (A). Causality is usually established using a Randomized Controlled Trial; 189 however, causal inference allows one to calculate causal effects using a “counterfactual” 190 Randomized Controlled Trial.

191 Randomized Controlled Trial- a Randomized Controlled Trial is the gold 192 standard for determining causality. This is because by randomly assigning treatments 193 among the two groups you remove the effects of confounding variables (technically, 194 confounding variables have the same distribution in both populations so their effect is 195 nullified). This means that any observed difference in the outcome is due purely to the 196 treatment. An example is the famous Salk polio vaccine experiment in the 1950s, where it 197 was proven that the polio vaccine helped prevent the disease5.

198 “counterfactual” Randomized Controlled Trial- a counterfactual is a “what if,” 199 or an alternate reality. For example, “what would have happened if I didn’t stop at the 200 stop sign?” As humans, we use counterfactuals all the time: if I didn’t go to work for a 201 month then I would be fired, so I’m going to work today. Even though we don’t skip 202 work for a month, it is easy to imagine what would happen in the counterfactual case.

203 Counterfactual randomized controlled trials take data from patients who were not 204 randomly assigned treatment and calculate what would have happened if all the patients 205 in that experiment were randomly assigned to receive the treatment or not receive the 206 treatment. It allows one to calculate causal effects in cases where Randomized Controlled 207 Trials are not feasible or are not ethical (for example, investigating the effects of smoking 208 on lung cancer).

209 Causal Graph- a causal graph is a group of variables, or factors, with arrows 210 going from possible causes to effects. One major limitation of causal inference is that, 211 while causal inference relies on causal graphs, these graphs cannot be empirically 212 validated and are created using expert knowledge. Graphs must be Directed, Acyclic 213 Graphs (DAGs), which means there are no double-sided arrows and no loops (if A causes 214 B then B cannot also cause A). An example of a causal graph is shown below.

215 216 Figure 3. Example Causal Graph. This graph has arrows from causes to potential effects. 217 If it rains, the lawn will be wet and you likely will not use your sprinklers at the same 218 time. If you use your sprinklers, the lawn will be wet, but this will not affect whether or 219 not it will rain. If the lawn is wet, this will not cause it to rain or cause the sprinklers to 220 turn on (this is why no arrows are going from Wet Lawn to either Rain or Sprinkler).

221 Pearl’s Do-Calculus- A method that takes a causal graph (shown above) and 222 calculates which variables to consider when estimating a causal effect.

223 Statistics:

224 10-fold cross-validation- 10-fold cross-validation is a way to see if your model 225 works on data it has not seen before. It starts by dividing the data into 10 equally-sized 226 segments (folds). After dividing the data, a model is trained using 9 of the folds, and 227 tested on the remaining fold. This procedure is repeated 10 times total so that each fold is 228 used as the test set exactly once. The average performance on all 10 test sets is used to 229 estimate model performance.

230 Bootstrapping- a technique that allows one to estimate how accurate a model is 231 by resampling a large number of times with replacement and training the model 232 separately on each resampled dataset.

233

234 I. Introduction

235 Burst suppression is a pattern of brain activity where an electroencephalogram (EEG) is 236 intermittently interrupted by “suppressions,” i.e., periods of reduced voltage (Figure ). 237 Prior studies in mechanically ventilated ICU patients found that burst suppression is 238 associated with increased 30-day, 6-month, and one-year mortality, and with increased 239 incidence and duration of delirium6,7. Since it has been thought that oversedation may 240 contribute to burst suppression, prior studies have recommended measures reducing the 241 use of sedatives, particularly propofol.

242 However, the complete picture remains unclear: among patients not receiving intentional 243 burst suppression for therapeutic reasons or anoxic encephalopathy (lack of oxygen in the 244 brain), how much do propofol and critical illness contribute to the occurrence of burst 245 suppression? How important is burst suppression as a mediator in the chain of causes and 246 effects that lead to in-hospital death? With these questions unanswered, whether reducing 247 propofol use will reduce burst suppression and whether reducing burst suppression will 248 save lives remains unclear.

249 While determining effects with a causal interpretation from observational studies is not 250 always possible, recent work has shown that this can be accomplished under certain 251 assumptions8,9. Here we estimate the effects: (1) of propofol dosage and patient critical 252 illness severity on burst suppression; (2) of burst suppression on in-hospital mortality; (3) 253 of propofol exposure on in-hospital mortality; and (4) of critical illness severity on in- 254 hospital mortality. Sensitivity analysis is used to quantify the robustness of the results to 255 unmeasured or unmeasurable confounding by computing the E-value10. These approaches 256 offer a step towards a mechanistic understanding of the causes and effects of burst 257 suppression, where randomized trials are impossible or unethical to carry out.

258

259

260 II. Methods

261 II.I Ethics Approval

262 This observational study was approved by the Institutional Review Board of Partners 263 HealthCare, which waived the need for informed consent for this study. The study also 264 complies with the World Medical Association Declaration of Helsinki regarding the 265 ethical conduct of research involving human subjects.

266

Category Variable Value Number of patients 471 Length of hospital stay: days, median (IQR) 16 (9 – 26)

Length of ICU stay: days, median (IQR) 9 (5 – 16) M In-hospital mortality: n (%) 128 (27.2%) B Burst Suppression Burden: median % (IQR) 9% (0% – 60%) D Propofol: mg/kg, median (IQR) 230.5 (86.7 – 570.4) Age: year, median (IQR) 61 (50 – 72) Female 233 (49.5%); Male 238 Gender: n (%) (50.5%)

Smoker: n (%) 225 (47.8%) Alcohol: n (%) 125 (26.5%) Patient APACHE II: median (IQR) 23 (20 – 27) covariates including CCI: median (IQR) 3 (1 – 6) critical Initial Glasgow Coma Scale score: median illness 7 (3 – 13) (C) (IQR) Premorbid modified Rankin Score: median 0 (0 – 2) (IQR)

Structural brain injury: n (%) 295 (62.6%) Hypertension: n (%) 252 (53.5%) Respiratory failure: n (%) 182 (38.6%)

267 Table 1- Patient Characteristics (covariates affecting less than 25% of the population are 268 included in Supplementary Table 1).

269

270 271 Figure 4. Histogram of the burst suppression burden using 10% bins. These data were 272 obtained by calculating the BSB for each patient (percent of epochs containing burst 273 suppression). It is notable that a little over half of the patients (50.3%) experienced little 274 or no burst suppression while in the ICU.

275

276 II.II Patients

277 Patient data were collected from seven Intensive Care Units (ICUs) at Massachusetts 278 General Hospital (MGH) between January 2014 and December 2017. The inclusion 279 criteria were: (1) age ≥18 years; (2) underwent continuous EEG monitoring for at least 18 280 hours; (3) on mechanical ventilation at the time EEG monitoring was initiated. The 281 exclusion criteria were: (1) anoxic encephalopathy due to cardiac arrest; (2) intentional 282 burst suppression for therapeutic reasons, such as refractory intracranial hypertension 283 (high fluid pressure in the skull), status epilepticus, or acute respiratory distress syndrome 284 (ARDS) / extracorporeal membrane oxygenation (ECMO, a life-support machine) / 285 mechanically ventilated patients that needed paralysis and paralytics for any other reason; 286 (3) ongoing status epilepticus during EEG monitoring (patients with status epilepticus as 287 the presenting complaint, however, were not excluded). The final cohort contained 471 288 patients that received anesthetics or anxiolytics purely for sedation purposes.

289 Patients in this dataset had a median age of 61 years with interquartile range 50 years to 290 72 years with almost equal numbers of genders (233 Female and 238 Male). 27% of the 291 patients (128 /471) died before discharge. Half of the patients (238/471) experienced 292 some degree of unintentional burst suppression. The histogram of the burst suppression 293 burden (defined below) is shown in Figure 4. Notably, over half of the patients in the 294 cohort had structural brain injuries (62.7%). Patient characteristics are described in Table 295 1. Potential confounding variables were selected based on clinical grounds from a list of 296 120 factors. Inclusion criteria for potential confounding variables were: (1) the possibility 297 (based on pharmacologic considerations) of affecting both unintentional burst 298 suppression and in-hospital mortality11; and (2) that data was available for more than 90% 299 of patients.

300

301 II.III EEG Recording

302 All EEG recordings were obtained following the standard international 10-20 system for 303 electrode placement. Raw EEG data were reviewed by two clinical neurophysiologists, 304 and the findings were reported in the electronic medical record every 8 to 12 hours, per 305 institutional protocol. All EEGs were performed on clinical grounds, to evaluate 306 suspected brain dysfunction such as subclinical seizure activity or encephalopathy.

307 II.IV Measures

308 This dataset is cross sectional, i.e. taken from a single hospitalization for each patient. For 309 every patient, we have the following variables: burst suppression (B), in-hospital 310 mortality (M), cumulative dose of propofol (D), and covariates (C). Note that, although 311 the data is cross sectional, the exposure variables of interest (B, D, C) precede the 312 outcome of interest in time (M), a necessary condition for causal inference. Beyond this 313 requirement, to allow estimating causal effects from our cross sectional dataset, we make 314 the assumptions encoded in the causal graph (Error! Reference source not found.). 315 These assumptions are based on domain knowledge and clinical experience. The 316 justification of the assumptions represented by this graph is detailed in the Supplementary

317 Methods. The estimated effects are valid to the extent that these assumptions are 318 reasonable.

319  In-hospital mortality (M): This is the outcome. Mortality is 1 if the patient was 320 deceased during hospital stay; otherwise 0. 321  Burst Suppression (B): Burst suppression burden (BSB) is defined as the percentage 322 of epochs for each patient containing a generalized/global EEG pattern with 323 intermittent periods of suppression ≤ 10 µV, with these periods of suppression 324 occupying 50% or more of an epoch12. Burst suppression was assessed visually by 325 clinical neurophysiologists as part of clinical care and reported in the medical record 326 in epochs of 8-12 hours duration. 327  Propofol infusion (D): We computed the body-weight-normalized cumulative 328 propofol dose (mg/kg) during ICU stay. 329  Patient covariates (C): Demographics and critical illnesses were included on the 330 basis that they can potentially affect both the propensity for burst suppression and 331 mortality (Supplementary TABLE S1).

332

333 II.V Effect Estimation

334 Effects derived from causal statistical inference can be understood as estimates of the 335 differences (e.g. in mortality, or burst suppression burden) that one should expect to 336 observe in various “counterfactual” randomized controlled trials (cRCTs), in which one 337 factor is randomly assigned for each patient while other factors are held constant. These 338 effects are counterfactual because they are estimated based on observational data, rather 339 than directly measured in actually-performed trials.

340 The first step in estimating the effects is to fit two prediction models. One for mortality 341 while adjusting for burst suppression, comorbidities, and propofol exposure, P(M|B, C, 342 D); and another for burst suppression, adjusting for comorbidities and propofol exposure, 343 P(B|C, D). To be robust to model misspecification, we evaluated two different 344 nonparametric models: gradient boosted tree and random forest in addition to logistic 345 regression. We selected the model with the best prediction performance. Each model was 346 fit using nested 10-fold cross-validation, consisting of external and internal folds. The 347 internal folds were used to select the model hyperparameters with the best average 348 performance on the validation folds. The external loop was used to obtain the final 349 performance by comparing the true values and the pooled predictions from all held-out 350 testing sets.

351 To further protect against possible model misspecification, we used doubly robust 352 estimation (DRE) in addition to ordinal regression (OR). The results presented in the 353 main text use DRE due to its theoretically proved lower bias to model misspecification.

354 II.VI Sensitivity Analysis

355 The E-value was recently proposed as a way to assess the robustness of effect estimates 356 to unmeasured confounding10. This value, calculated as 퐸 = 푅푅 + 푅푅(푅푅 − 1),

357 quantifies the minimum strength of unmeasured confounding on either the cause or effect 358 that would be needed to explain away the estimated effect, expressed on the risk ratio 359 (RR) scale.

360

361 II.VI Confidence Intervals

362 Confidence intervals were estimated nonparametrically using bootstrapping, which 363 simulates the sampling process from an infinite population. Specifically, the effect of 364 each arrow is calculated 200 times. In each calculation, a bootstrapped dataset the same 365 size as the actual dataset was obtained by random sampling with replacement. The 95% 366 confidence interval was estimated using the 2.5% and 97.5% percentiles of the 367 bootstrapped observations. P-values are provided for univariate associations. P-values of 368 the estimated effects are not stated, in accordance with recent literature11.

369 III. Results B | C, D M | B, C, D Prediction Model Spearman’s rho Area under ROC correlation Logistic regression 0.158 (0.074 – 0.247) 0.634 (0.581 – 0.689) Gradient boosted 0.174 (0.084 – 0.264) 0.684 (0.632 – 0.737) trees Random forest 0.188 (0.097 – 0.277) 0.682 (0.631 – 0.732) 370

371 Table 2. Prediction performance of each prediction model from the external cross- 372 validation

373

374 III.I Predicting Burst Suppression Burden and Mortality

375 In Table , we show the cross-validation performance of predicting burst suppression 376 burden using patient covariates and propofol dose (B | C, D); and the cross-validation 377 performance of predicting mortality using patient covariates, propofol dose, and burst 378 suppression burden (M | B, C, D). Overall, the best performance is achieved by the 379 random forest model, where the spearman’s correlation is 0.188 (0.097 – 0.277) for 380 predicting burst suppression, and the area under ROC is 0.682 (0.631 – 0.732) for 381 predicting mortality. The results suggest a nonlinear relationship between these variables. 382 In the next subsection, we show the effect estimates based on random forest.

383 The ten most important variables for predicting burst suppression burden based on 384 random forest are (descending order): Age, propofol, APACHE II, initial GCS, CCI, 385 premorbid mRS, seizure / status epilepticus, structural brain injury, gender, and smoking. 386 The ten most important variables for predicting mortality are (descending order): Age, 387 CCI, propofol, APACHE II, septic shock, initial GCS, burst suppression burden, 388 premorbid mRS, hemorrhagic stroke, and seizure/status epilepticus.

389

390

391 III.II Estimated Effects

392

393 394 Figure 5. Estimated direct and mediated effects on mortality of burst suppression burden, 395 clinical condition, and drug exposure. B: burst suppression burden measured by the percentage 396 of 8-12 hour EEG epochs with burst suppression (see methods). M: in-hospital mortality (1, 397 deceased; 0, alive). D: drug, i.e. cumulative propofol exposure. C: patient characteristics 398 including demographics and critical illnesses (Table 1). The estimated effect is shown above each 399 arrow, measured on the difference scale. The width of each arrow is proportional to the 400 estimated effect. Effects are measured as the difference in expected outcome between groups 401 in a counterfactual randomized clinical trial (cRCT), which two groups are randomized to two 402 different levels of B, D, or C. (A) Difference in mortality (39%) and difference in Burst 403 Suppression Burden (36%) between ICU patients randomized to severe vs mild critical illness 404 severity. (B) Difference in mortality (-.5%) and difference in Burst Suppression Burden (6%) 405 between patients randomized to high vs. low cumulative exposure to propofol. (C) Difference in 406 mortality between patients randomized to 0% vs 100% burst suppression burden (12%). (D) 407 Sensitivity analysis, measuring the minimum risk ratio of an unmeasured confounding factor to 408 either the exposure or the outcome that would be needed to explain away the estimated effect 409 (E-value).

410 We assessed the effects of the burst suppression burden (B), exposure to propofol (D), 411 and clinical condition (C) on mortality, by contrasting the expected mortality rate under 412 two different levels for each of the explanatory variables B, D, and C. These effects 413 represent the estimated difference we would expect to see in a series of counterfactual 414 randomized clinical trials (cRCTs), each with two groups: one randomized to each of the 415 two levels of the explanatory variables.

416 Effect of burst suppression on mortality (Figure 2A): To quantify the effect of burst 417 suppression on in-hospital mortality, we estimated the expected difference in outcome for 418 a cRCT in which one group had their burst suppression burden set to 0%, and the other to 419 100%. The effect is a 12% [95% CI 7-17] increase in mortality, or a 1.45 times higher 420 mortality risk ratio.

421 Effect of propofol exposure on burst suppression burden and on mortality (Figure 2B): 422 To measure the effect of propofol exposure on mortality, we estimated the expected 423 difference in outcome for a cRCT in which one group received a weight-normalized 424 cumulative propofol dose over their ICU stay of 284.1 mg/kg (“low dose”), while the 425 other group received 1315.5 mg/kg (“high dose”). In our analysis, propofol can affect 426 mortality either directly, or indirectly by changing the burst suppression burden. The 427 average direct effect is to decrease mortality slightly, by 0.5% [0.05-1].

428 Effect of clinical condition on burst suppression burden and on mortality (Figure 2C): To 429 measure the effect of clinical condition on mortality, we estimated the expected 430 difference in outcome for a cRCT in which one group of ICU patients is “assigned” to 431 have mild critical illness, while the other group is assigned to be more severely ill. In our 432 analysis, clinical condition can affect mortality either directly, or indirectly by changing 433 the burst suppression burden. The average direct effect is to increase mortality by 35% 434 [29-41] (8.42 [4.66-44.17] times increased risk ratio).

435 The complete list of estimated effects is provided in the Supplementary Tables (Tables 436 S3 and S4).

437

438 III.III Sensitivity Analysis

439 There are many potential unmeasured confounding variables in any cross-sectional study. 440 In our case, they are defined as any unmeasured variable that affects two or more 441 variables among D, B, and M, which include (1) the differences between the seven ICUs 442 such as room temperature and noise level; (2) other variables not included in the 443 covariates C such as genetic factors.

444 The E-values for each of the estimated effects are shown in Error! Reference source not 445 found.C. For example, to explain away the estimated effect of burst suppression on in- 446 hospital mortality as due to cofounding, there would need to be an unmeasured factor U 447 that makes either the expected burst suppression burden or the expected in-hospital 448 mortality rate 2.27 times more likely when U has a unit change (all other factors 449 remaining constant). The same explanation applies to other pairs.

450

451 IV. Discussion

452 In this study, we investigated the mechanisms by which burst suppression and mortality 453 occur in ICU patients. Our results suggest that critical illness itself, as opposed to 454 exposure to high doses of anesthetic/sedative drugs, is the dominant driver of both burst 455 suppression and mortality.

456 Prior studies have also shown that burst suppression is associated with increased 457 mortality, which is consistent with our finding. A prior study on 125 mechanically 458 ventilated adult ICU patients found that unintentional burst suppression (taken as a 459 dichotomized variable) was associated with 6-month mortality, with a hazard ratio of 460 2.04 (1.12 – 3.70) based on a time-dependent Cox proportional hazard regression model 461 after adjusting for clinically import covariates, including CCI, baseline dementia, 462 APACHE II, sequential organ failure assessment (SOFA), daily coma and delirium 463 status6. The difference between their results and ours suggests that the clinical setting and 464 patient population, i.e., ICU vs. operating room, may be an effect modifier.

465 Our results suggest that most unintentional burst suppression is probably not due to the 466 use of high propofol doses; rather, burst suppression is primarily attributable to critical 467 illness itself. Two possible explanations exist. The first is that critical illness itself is 468 primarily responsible for a large portion of burst suppression in the ICU. In support of 469 this possibility, burst suppression in the EEG has been reported as one manifestation of 470 severe sepsis13,14. Our analysis suggests that burst suppression is likely a manifestation of 471 a wide range of severe medical illnesses. A second, compatible explanation is that severe 472 illness alters the brain’s pharmacodynamic response to sedative drugs, such that low 473 doses that would ordinarily induce only sedation instead induce burst suppression. Under 474 either of these explanations, burst suppression serves as a biomarker of neurological and 475 medical “vulnerability”. This suggests that the problem is not primarily exposure to high 476 doses of sedative drugs and that avoidance of burst suppression by reducing sedative use 477 would not likely have a large effect on mortality. Note that our results still do not rule out 478 the possibility that closely monitoring a patient’s EEG to personalize propofol dosing 479 might reduce mortality to a modest degree. That is, while merely avoiding “over-dosing” 480 is unlikely to impact mortality, it is still possible that tailoring doses to the bare minimum 481 needed for an individual (recognizing that some individuals require substantially lower- 482 than-average doses to achieve the same pharmacodynamic effect) could be beneficial. A 483 recent randomized clinical trial appears to provide some evidence along these lines: The 484 trial was designed to test whether EEG-guided anesthesia could reduce the incidence of 485 intra-operative burst suppression; while intra-operative burst suppression was reduced, 486 this did not affect delirium. Nevertheless, the reduction in burst suppression was 487 associated with a small (2.4%) reduction in 30-day mortality15.

488 In terms of the validity of our effect estimates, the similarity in magnitude and direction 489 of estimated effects among multiple methods suggests that these estimates are robust to 490 model specification. The fact that one of the two nonparametric models was more 491 accurate than logistic regression in predicting burst suppression, and that both 492 nonparametric methods outperformed logistic regression in predicting mortality, suggests

493 a nonlinear relationship exists between the variables. Therefore, nonparametric methods 494 could be a more precise modeling method for predicting and effect estimation in a 495 clinical setting, albeit at the price of reduced clinical interpretability.

496 Our study has several important limitations. (1) Our results may be influenced by 497 selection bias. In current clinical practice, continuous EEG is not routinely performed in 498 mechanically ventilated ICU patients; rather it is performed primarily in cases with 499 suspected brain dysfunction such as subclinical seizure activity or encephalopathy, 500 conditions themselves associated with increased mortality. The fact that more than 60% 501 of patients in our dataset had structural brain injury reflects the fact that continuous EEG 502 use is generally symptom-driven as opposed to randomly assigned. Thus, the effect sizes 503 estimated in our study might not generalize to all ICU patients. (2) Our estimation of 504 burst suppression burden (BSB) is coarse, being based on relatively long epochs (8-12 505 hours in duration, see Methods). Continuous quantitative evaluation of burst suppression 506 would provide higher time resolution and might allow more granular insights into the 507 effects. (3) The way we delineated mild illness from severe illness using 10% and 90% 508 percentiles is arbitrary although reasonable. Due to its continuous nature, the effect size is 509 dependent on the contrast we pick. The patient closest to 10th percentile is a 30-year-old 510 female with no comorbidities who presented with CNS infection and respiratory failure, 511 representative of a typical young patient with acute illness in the ICU; while the patient 512 closest to the 90th percentile is an 83-year-old female with multiple comorbidities and 513 severe cerebrovascular abnormalities, representative of a typical elderly ICU patient with 514 chronic illness. (4) We do not have other post-illness morbidity scores. In-hospital 515 mortality considered here is an important and easily obtained variable, but ultimately, 516 neurocognitive morbidity is also important to investigate. (5) Our quantification of drug 517 exposure using cumulative dose normalized by body weight ignores the finer temporal 518 structure of drug administration history. Therefore, we do not know how prolonged low 519 exposure vs. brief high doses might differentially affect burst suppression and mortality; 520 knowing these details could provide more detailed guidance for drug dose management. 521 (6) We represented many of the clinical variables as binary (e.g. respiratory failure, 522 seizures or status epilepticus, toxic metabolic encephalopathy). This is a simplification 523 since in most cases these clinical problems can be given finer-grained characterizations 524 according to etiology (e.g. cause of seizures) and severity (e.g. size of ischemic stroke). 525 Nevertheless, given the variety of clinical problems in the cohort, we binarized variables 526 in order to avoid a profusion of small subsets of patients that would make the statistical 527 analysis problematic. Future studies focused on more homogenous but more “deeply 528 phenotyped” cohorts might yield further insights that were not accessible to our 529 approach. (7) Along the same lines, patients came from 7 different ICUs, which have 530 somewhat different patient populations. We did not explicitly adjust for ICU location, 531 thus this is a potential source of unmeasured confounding in our graph. This limitation is 532 mitigated to some extent by the broad range of medical condition variables that we did 533 adjust for, which should to a large extent capture the reasons for patients being in the 534 different ICUs. (8) Also in the same vein, seizures in ICU patients exist along a 535 continuum, ranging from rare intermittent seizures to frequent seizures, to continuous 536 seizures. There is an ongoing debate regarding whether seizures in critically ill patients 537 contribute to poor outcomes vs. directly cause harm; and if they do directly worsen 538 outcomes, there no consensus on the dose-response relationship between seizure burden

539 and neurologic outcomes. Thus, seizure burden is a potentially important confounder 540 omitted from our analysis and deserves further investigation.

541

542 VI. Conclusions

543 Unintentional burst suppression contributes substantially to in-hospital mortality. 544 However, unintentional burst suppression appears to be caused more by critical illness 545 itself, rather than by high exposure to propofol, and thus may be less modifiable than 546 previously believed.

547

548 WORKS CITED

549 550

551 1. BaHammam AS, Gacuan DE, George S, Acosta KL, Pandi-Perumal SR, Gupta R. 552 Chapter 25 Polysomnography I: Procedure and Technology. In: Synopsis of Sleep 553 Medicine. Apple Academic Press; 2016:443-453. doi:10.1201/9781315366340-26

554 2. Status Epilepticus | Johns Hopkins Medicine. 555 https://www.hopkinsmedicine.org/health/conditions-and-diseases/status- 556 epilepticus. Accessed January 18, 2020.

557 3. Victor Lavrenko. Decision Tree 1: how it works. 2014. 558 https://www.youtube.com/watch?v=eKD5gxPPeY0. Accessed January 18, 2020.

559 4. Starmer J. StatQuest: Random Forests Part 1 - Building, Using and Evaluating - 560 YouTube. https://www.youtube.com/watch?v=J4Wdy0Wc_xQ&t=320s. Published 561 2018. Accessed January 18, 2020.

562 5. History.com. Polio vaccine trials begin. https://www.history.com/this-day-in- 563 history/polio-vaccine-trials-begin. Published 2019. Accessed January 18, 2020.

564 6. Watson PL, Shintani AK, Tyson R, Pandharipande PP, Pun BT, Ely EW. Presence 565 of electroencephalogram burst suppression in sedated, critically ill patients is 566 associated with increased mortality. Crit Care Med. 2008;36(12):3171-3177. 567 doi:10.1097/CCM.0b013e318186b9ce

568 7. Andresen JM, Girard TD, Pandharipande PP, Davidson MA, Ely EW, Watson PL. 569 Burst suppression on processed electroencephalography as a predictor of postcoma 570 delirium in mechanically ventilated ICU patients. Crit Care Med. 571 2014;42(10):2244-2251. doi:10.1097/CCM.0000000000000522

572 8. Hernan MA, Robbins JM. Causal Inference. Boca Raton: Chapman & Hall/CRC, 573 forthcoming; 2019.

574 9. Pearl J. Causality. Cambridge University Press; 2009.

575 10. VanderWeele TJ, Ding P. Sensitivity Analysis in Observational Research: 576 Introducing the E-Value. Ann Intern Med. 2017;167(4):268-274. 577 doi:10.7326/M16-2607

578 11. Lederer DJ, Bell SC, Branson RD, et al. Control of Confounding and Reporting of 579 Results in Causal Inference Studies: Guidance for Authors from Editors of 580 Respiratory, Sleep, and Critical Care Journals. Ann Am Thorac Soc. September 581 2018:AnnalsATS.201808-564PS. doi:10.1513/AnnalsATS.201808-564PS

582 12. Hirsch LJ, Laroche SM, Gaspard N, et al. American clinical neurophysiology 583 society’s standardized critical care EEG terminology: 2012 version. J Clin 584 Neurophysiol. 2013. doi:10.1097/WNP.0b013e3182784729

585 13. Polito A, Eischwald F, Maho A-L, et al. Pattern of Brain Injury in the Acute 586 Setting of Human Septic Shock. Crit Care. 2013;17(5):R204. doi:10.1186/cc12899

587 14. K. H, N. G, F. S, et al. Clinical neurophysiological assessment of sepsis-associated 588 brain dysfunction: A systematic review. Crit Care. 2014. doi:10.1186/s13054-014- 589 0674-y

590 15. Wildes TS, Mickle AM, Ben Abdallah A, et al. Effect of Electroencephalography- 591 Guided Anesthetic Administration on Postoperative Delirium Among Older Adults 592 Undergoing Major Surgery. JAMA. 2019;321(5):473. 593 doi:10.1001/jama.2018.22005

594

595

596 SUPPLEMENTARY MATERIALS Category Variable Value Number of patients 471 Length of hospital stay: days, median (IQR) 16 (9 – 26) Length of ICU stay: days, median (IQR) 9 (5 – 16) M In-hospital mortality: n (%) 128 (27.2%) B Burst Suppression Burden: median % (IQR) 9% (0% – 60%) D Propofol: mg/kg, median (IQR) 230.5 (86.7 – 570.4) Age: year, median (IQR) 61 (50 – 72) Female 233 (49.5%); Male 238 Gender: n (%) (50.5%) Smoker: n (%) 225 (47.8%) Alcohol: n (%) 125 (26.5%) Substance abuse: n (%) 57 (12.1%) APACHE II: median (IQR) 23 (20 – 27) CCI: median (IQR) 3 (1 – 6) Initial Glasgow Coma Scale score: med (IQR) 7 (3 – 13) Premorbid modified Rankin Score: med (IQR) 0 (0 – 2) Structural brain injury: n (%) 295 (62.6%) Hypertension: n (%) 252 (53.5%) Respiratory failure: n (%) 182 (38.6%) Seizures / status epilepticus on presentation: n 106 (22.5%) (%) Hemorrhagic stroke: n (%) 94 (20.0%) Patient Cardiovascular accident: n (%) 92 (19.5%) covariates including Subarachnoid hemorrhage: n (%) 83 (17.6%) critical Septic shock: n (%) 76 (16.1%) illness (C) Toxic metabolic encephalopathy: n (%) 74 (15.7%) History of seizure: n (%) 72 (15.3%) Subdural hematoma: n (%) 49 (10.4%) CNS infection or inflammation: n (%) 48 (10.2%) Ischemic stroke: n (%) 40 (8.5%) Cardiac failure: n (%) 39 (8.3%) Renal failure: n (%) 35 (7.4%) Brain tumor: n (%) 30 (6.4%) Brain Surgery: n (%) 29 (6.2%) Malignancy solid tumors hematologic: n (%) 27 (5.7%) Liver failure: n (%) 24 (5.1%) Primary psychiatric disorder: n (%) 20 (4.2%) CNS Cancer: n (%) 15 (3.2%) Endocrine emergency: n (%) 10 (2.1%) Primary hematological disorder: n (%) 9 (1.9%) Systemic hemorrhage: n (%) 3 (0.64%) 597 Supplementary Table S1- full list of patient characteristics.

598 Variable Mild Severe Illness Illness Age (year) 30.5 83 Gender Female Female Smoker No No Alcohol No No Substance abuse No No APACHE II 15 26 CCI 0 8 Initial Glasgow Coma Scale score 12 3 Premorbid modified Rankin Score 0 5 Structural Brain Injury No Yes Hypertension No No Respiratory failure Yes No Seizure or Status Epilepticus No Yes Hemorrhagic stroke No No Cardiovascular accident No No Subarachnoid hemorrhage No No Septic shock No No Toxic metabolic encephalopathy No No History of Seizure No No Subdural hematoma No No CNS infection or inflammation Yes No Ischemic stroke No Yes Cardiac failure No Yes Renal failure No No Brain tumor No No Brain Surgery No No Malignancy solid tumors No No hematologic Liver failure No No Primary psychiatric disorder No No CNS cancer No No Endocrine emergency No No Primary hematological disorder No No Systemic hemorrhage No No Propofol (mg/kg) 284.1 1315.5 599 600 Supplementary Table S2. The patient characteristics and propofol dose of the 601 representative healthy and sick patients 602

Prediction C->B D->B C.I. B->M Method 0.05931 or Logistic 0.34882 0.02195 (-0.06478 -- 0.18988) regression (0.09196 -- 0.60369) (-0.00302 -- 0.06260) 0.17093 dr (0.09544 -- 0.24050) 0.02227 or 0.38792 0.05576 (-0.02129 -- 0.10017) Random forests (0.08335 -- 0.65971) (0.00812 -- 0.11760) 0.11757 dr (0.07123 -- 0.16770) 0.03684 or Gradient boosted 0.29477 0.03577 (-0.04691 -- 0.11993) tree (-0.02629 -- 0.85398) (-0.02129 -- 0.22492) 0.05309 dr (-0.01129 -- 0.09600) 603 Supplementary Table S3. Causal effect estimates using different prediction methods. C: 604 chronic illness, B: Burst suppression, D: Drug infusion (propofol), M: mortality. or: 605 ordinal regression. dr: doubly robust estimation. Doubly robust estimation was not used 606 to predict B (burst suppression) because burst suppression is a continuous outcome 607 between 0 and 1, not a dichotomous outcome. C.I. Causal inference method. 608 Causal Prediction Method Inference C→M D→M Method Ordinal 0.37162 -0.04517 Logistic regression regression (0.03553 -- 0.76009) (-0.09356 -- 0.00389) Doubly robust 0.38047 -0.00346 estimation (0.30688 -- 0.44685) (-0.00860 -- 0.00113) Ordinal 0.32201 -0.01109 regression (0.13797 -- 0.45161) (-0.04860 -- 0.04753) Random forests Doubly robust 0.34613 -0.00413 estimation (0.28842 -- 0.40269) (-0.00891 -- 0.00044) Ordinal 0.12241 0.00909 regression (0.04255 -- 0.65686) (-0.10485 -- 0.08366) Gradient boosted tree Doubly robust 0.37884 -0.00344 estimation (0.31275 -- 0.42866) (-0.00822 -- 0.00104) 609 Supplementary Table S4. Causal effect estimates using different prediction methods 610 and different causal inference methods. C: chronic illness, B: Burst suppression, D: Drug 611 infusion (propofol), M: mortality. 612 613

614 Computing the potential outcomes 피[퐁(퐜)] and 피[퐁(퐝)] 615 Based on the structure of the graph in Error! Reference source not found., we have 피[퐵(푐)] = 피[퐵|푑표(퐶 = 푐)] Effect of c on B is the probabilityf of B given C = c (c is a specific set of clinical covariates for a representative patient). = 푃퐵 = 1푑표(퐶 = 푐) B is the burst suppression burden, a fraction between 0 and 1. = 푃(퐵 = 1|퐶 = 푐) This step is based on Pearl’s do calculus9

= 푃(퐵 = 1|퐶 = 푐, 퐷 = 푑)푃(퐷 = 푑|퐶 = 푐)푑푑 C and D are independent based on = 푃(퐵 = 1|퐶 = 푐, 퐷 = 푑)푃(퐷 = 푑)푑푑 applying d-separation rule to Figure 2 (D_||_C). This is because there are no measured variables that effect both D and C and neither D nor C effect the other (no backdoor paths). 1 The integral is approximated using the ≈ 푃퐵 = 1푐, 퐷() . 푁 sample mean. 616 617 618 Similarly, we have 1 619 피[퐵(푑)] ≈ 푃퐵=1퐶(),푑 . 푁 620 621 Computing the potential outcome 피[푴(풃)] 622 We used several methods as below. 623 624 Outcome regression (OR) 피[푀(푏)] = 피[푀|푑표(푏)] The potential outcome, or probability of mortality given a certain level of Burst Suppression Burden (b). = 푃푀 = 1푑표(푏) M is binary

= 푃(푀 = 1|푑표(푏), 푐, 푑)푃푐, 푑|푑표(푏)푑푐푑푑 , Applying intervention to b, as = 푃(푀 = 1|푑표(푏), 푐, 푑)푃(푐, 푑)푑푐푑푑 expressed by do(b), does not change c , and d, since b does not affect either c

or d (Figure 2). Based on Pearl’s do calculus9 = 푃(푀 = 1|푏, 푐, 푑)푃(푐, 푑)푑푐푑푑 , 1 The integral is approximated using the ≈ 푃푀 = 1푏, 퐶(), 퐷() . 푁 sample mean. 625 626 Compute 퐌퐜퐛 using outcome regression 피[푀(푐, 푏)] = 피[푀|푑표(푐), 푑표(푏)] The potential outcome, or probability of mortality given a certain collection of covariates (c) and a certain level of Burst Suppression Burden (b). = 푃푀 = 1푑표(푐), 푑표(푏) M is binary

= 푃(푀 = 1|푑표(푐), 푑표(푏), 푑)푃푑|푑표(푐), 푑표(푏)푑푑 Intervening to change c and b = 푃(푀 = 1|푑표(푐), 푑표(푏), 푑)푃(푑)푑푑 does not change d, since c and b are not the parents of d (Figure 2) Based on Pearl’s do calculus9 = 푃(푀 = 1|푐, 푏, 푑)푃(푑)푑푑 1 The integral is approximated ≈ 푃푀 = 1푏, 푐, 퐷() . 푁 using the sample mean. 627