Cannabis Use Is Associated With a Small Increase in the Risk of Postoperative Nausea and Vomiting

A Retrospective Machine-learning Causal Analysis

Wendy Suhre; Vikas O'Reilly-Shah; Wil Van Cleve


BMC Anesthesiol. 2020;20(115) 

In This Article


This study was a retrospective cohort analysis of general anesthesia cases lasting 30 min or longer conducted at the University of Washington Medical Center (UWMC) from July 1, 2016 until September 30, 2018. Data from Harborview Medical Center (HMC) from the same time period were used for model validation. Inclusion criteria were general anesthesia cases for patients aged 18 years and older with a documented pre-anesthetic evaluation who also received post-operative care in the Post Anesthesia Care Unit. Data regarding anesthetic management were obtained from the hospital Anesthesia Information Management System (Merge AIMS, Hartland, WI). Obstetric and cardiac cases were excluded, as were cases with an American Society of Anesthesiologists Physical Classification 4 or greater. Data regarding risk factors for PONV and pattern of ongoing cannabis use were gathered from the pre-anesthetic evaluation documented for the case. Data regarding the occurrence of PONV were abstracted automatically from nursing documentation in the post-anesthesia care unit. Severity of PONV was not considered in this analysis. The dataset was obtained from a central repository of perioperative and anesthetic data maintained by the UW Perioperative and Pain initiatives in Quality and Safety Outcome Center, which performed data extraction, validation, and de- identification prior to providing it to our research team. Because of patient de-identification, this study was exempted from review by the University of Washington Institutional Review Board as non-human subjects research. This manuscript was prepared in accordance with STROBE guidelines for improved reporting of observational studies.[11]

Primary Predictor

Plain text from the preoperative evaluation note regarding the use of non-prescribed substances/drugs was extracted and manually reviewed by one of the investigators (WS). cannabis use as described by the patient was classified by the investigator as "daily" (used on a daily basis), "current" (used at present, but less often than daily) or "none" (i.e. past use was not considered).

Primary Outcome

A composite variable constituted by PONV of any severity as recorded by the recovery room nurse, or the administration of an antiemetic drug in the PACU (ondansetron, promethazine, perphenazine, or metoclopramide), was used to indicate the presence of PONV in our analysis.


Following the strategy employed by a recent published study examining associations between perioperative medication use and PONV, we picked a set of a priori covariates we expected to be associated with PONV.[12] These included (a) age less than 50 years, (b) ASA classification, (c) exposure to nitrous oxide (defined as exposure to nitrous oxide for greater than 5% of surgical time), (d) exposure to a potent volatile anesthetic agent (defined as age adjusted MAC > 0.5 for greater than 15% of surgical time), (e) surgical duration in minutes (log transformed), (f) female sex, (g) history of PONV or motion sickness, (h) absence of patient reported tobacco use, (i) receipt of an opioid drug in the PACU, and (j) the total number of prophylactic anti-emetic drugs given pre- or intraoperatively (drugs considered included dexamethasone, gabapentin, haloperidol, meclizine, metoclopradmide, ondansetron, prochlorperazine, promethazine, and transdermal scopolamine). Notably, many of these covariates were unlikely to be associated with both cannabis use and PONV, classifying them as potential effect modifiers rather than confounders. In some of our analyses, we combined the PONV risks commonly summed to create the "simplified Apfel score" (i.e. female sex, history of PONV or motion sickness, absence of patient reported tobacco use, and receipt of an opioid drug in the PACU) and stratified our analysis by the number of PONV risks.[13,14]

Statistical Analysis

Our primary analysis estimated the causal effect of cannabis use on PONV. Realizing that cannabis use was not randomly distributed throughout our sample, we employed a statistical method known as Bayesian Additive Regression Trees (BART). BART combines flexible nonparametric regression tree methods with a "Bayesian backfitting" algorithm that minimizes the amount of overfitting that can occur in similar machine learning algorithms.[15] BART has been demonstrated to generate valid causal effect estimates without the well-described weaknesses of propensity score estimation or matching, which include the potential for improper specification of the propensity model, problems handling large numbers of covariates, and proper modeling of non-linear relationships and variable interactions.[16]

Analyses were performed using R version 3.6.2 (R Core Team, Vienna, Austria) within the RStudio platform 1.2.1335 (R Studio Team, Boston, MA). Probit BART models for the primary analysis were created using the BART package v2.7.[17] We calculated 2 formulations of the causal effect estimate: the relative risk of PONV and an absolute increase in the probability of PONV associated with no use of cannabis (referent group), current use, and daily use. Counterfactual sample estimates were generated by artificially assigning all members of the sample to each condition and comparing the probability of the outcome of interest under each condition, allowing us to calculate the sample average treatment effect (sATE). Because BART provides a true Bayesian posterior estimate, we generated 95% credible intervals by carrying out the aforementioned analyses for each of 1000 Markov Chain Monte Carlo (MCMC) estimates, and then extracting the appropriate quantile from the resulting population of parameter estimates. As an internal validation of our initial result, we performed a propensity score analysis: we first created a Bayesian logistic regression model to model the probability of using any cannabis using the brms package v 2.11.1, followed by a second Bayesian logistic regression model that included our estimated probability of cannabis use as a covariate (e.g. propensity score adjustment) alongside parameters otherwise identical to those in our BART model.[18] We then examined both the parameter estimates and sample average treatment effects of this model.[18,19] Finally, as an external validation of our findings, we created a second BART model with parameters identical to those used in our initial model using data collected at HMC, and again assessed the sATE for cannabis use on the risk of PONV. All statistical analyses were conducted by the primary research team.

Statistical Significance

Bayesian posterior estimates differ fundamentally from frequentist parameter summaries, and therefore no a priori statement about binary p-value thresholds representing statistical significance can be offered. We report 95% credible intervals for our parameter estimates, which represent the numeric interval in which 95% of the posterior probability density lies. Further, when estimating relative risk, we calculated the posterior probability that the relative risk exceeded 1.