We conducted this cross-sectional study in two urban academic emergency departments in the USA—Beth Israel Deaconess Medical Center (BIDMC), Boston, Massachusetts and Cooper University Hospital (CUH), Camden, New Jersey—from September 2023 to May 2024. This study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement (Supplemental Table 1)10 and the Checklist for Reporting Of Survey Studies (CROSS) (Supplemental Table 1).11

Each hospital’s Institutional Review Board approved the study, and all subjects provided written informed consent prior to participation. Subjects were not compensated for their time.

We enrolled a convenience sample of adult ED patients at each site. Inclusion criteria were as follows: (1) age 18 years or older; (2) presenting as a patient to the ED; and (3) English or Spanish speaking. Exclusion criteria included the following: (1) having an acute psychiatric emergency; (2) inability to participate (i.e., history of dementia, critically ill); (3) previously participated in the study; and (4) prisoner.

Research assistants approached potential subjects in the ED for enrollment after completion of care by the ED clinician (i.e., either time of hospital admission or discharge from the ED). After obtaining written informed consent, the research assistants gave the subjects a computer tablet with the research questionnaire in electronic form. The Research Electronic Data Capture (REDCap) survey distribution tool was used to administer the research questionnaire and capture responses directly into the research database.12,13 Research assistants were then instructed to leave the patients’ bedside to allow for privacy while filling out the questionnaire, unless the patient requested they stay to help, and to return in 15 min to collect the tablet.

Our group previously performed a systematic review of compassion training for physicians.9 Based on this review, we derived a list of 27 physician actions previously suggested to increase patient experience of compassion (SupplementalTable 3). The research questionnaire queried patients if the ED physician performed each action during the ED encounter. Subjects answered, “yes,” “no,” or “unsure” for each action. The questionnaire collected patient demographic information. Using a standardized data collection form, we collected ED length of stay, and disposition from the ED (discharge vs. observation/hospital admission) from the medical record.

The primary outcome measure was patient experience of ED physician compassion using the 5-item compassion measure. This measure was previously psychometrically validated for use in the ED, outpatient, and inpatient settings in over 12,000 patients and consists of five items measured on a 4-point Likert scale (Fig. 1).14,15,16 The 5-item compassion measure was administered prior to questions pertaining to the actions and the items were administered in the order shown in Fig. 1 (there was no randomization). Scores for the five items were summed to obtain a composite compassion score. Potential scores range from 5 to 20, with higher scores indicating greater compassion. We entered all data into REDCap, and exported into Stata/SE 18.0 for Mac, StataCorp LP (College Station, TX, USA) for analysis.

The 5-item compassion measure. All item responses on a four-point Likert scale (1 = Never, 2 = Sometimes, 3 = Usually, 4 = Always).

We report continuous variables as median and interquartile range (IQR), and categorical variables as frequencies and percentages. We tested the internal reliability of the 5-item compassion measure using Cronbach’s alpha and display the full distribution of the composite compassion score in histogram form.

Data from BIDMC (site one) was used as the “training set” to identify the combination of actions that best predicted a higher 5-item compassion score. We performed separate pairwise univariable linear regression analyses to identify which actions were associated with the 5-item compassion score, with each action as a binary independent variable (yes vs. no/unsure) and the 5-item compassion measure as a continuous dependent variable. We used conservative robust standard errors to estimate the 95% CI to reduce the risk of type I error. We considered an action to be associated with the 5-item compassion score if the 95% CI for the beta coefficient did not cross zero. Given the potential for multicollinearity between measured actions, we used variance inflation factor (VIF) to test for high correlation between actions, which can produce unstable regression models. VIF < 5 indicates multicollinearity is not a problem. We used the machine learning algorithm LASSO to identify the group of actions that best predict the 5-item compassion measure. We entered the 5-item compassion measure as a continuous (ordinal) variable and performed a post hoc sensitivity analysis entering the 5-item compassion measure as a categorical variable: low (score < 15), moderate (score 15–19), and perfect (score 20) (See Supplemental Fig. 2 for details of LASSO models).

Data from CUH (site two) was used as the “test set” to validate the final model in a separate cohort. We calculated the post-selection model fit for both sites separately and report them as r2. We then performed a linear regression model with the final actions as independent variables and the 5-item compassion measure as the dependent variable to determine the overall r2 across both cohorts and used Shorrocks-Shapley decomposition to determine each action’s relative contribution to the overall r2.

We tested if, among the validation cohort, a greater number of model actions reported being performed was associated with a greater 5-item compassion score using linear regression analysis and display the median compassion score by the number of actions performed using box plots. We also report the proportion of subjects with a perfect 5-item compassion score of 20 by the number of model actions reported being performed.

We combined both data sets to test if the strength of the associations between individual actions and the 5-item compassion measure score differ by patient race (White vs. Black) and gender (male vs. female). We used separate pairwise multivariable linear regression models to test for an interaction between each action and race (reference: White) and gender (reference: male). We considered an interaction significant if the 95% CI (using robust standard errors) for the interaction term did not cross zero.

There is no current method for sample size calculation for LASSO. LASSO can assess 1000 s of possible predictors at once, but the number of final selected predictors must be equal to or smaller than the sample size.17