Time-dependent propensity scores and sequential risk-set matching

We performed time-dependent propensity score and risk-set matching analysis in each cohort of the initial cardiac rhythm to evaluate the association between ECPR and outcomes. The details of the time-dependent propensity scores and sequential risk-set matching are explained in the Additional file (Additional file 1: Method 3). To estimate the time-dependent propensity score of receiving ECPR after hospital arrival, we applied the fine-gray regression model in the presence of competing risk with time-dependent covariates [30]. In this model, we dealt with ROSC and death (termination of resuscitation) before ECPR as the competing risk. We also set 120 min after arrival at the hospital as censoring in the model. Time-dependent covariates were defibrillation, intubation, and adrenaline administration in the hospital. The time-independent covariates in the calculation of the propensity score were as follows: patient characteristics (sex and age), pre-hospital information (witness, bystander CPR, initial cardiac rhythm, defibrillation by bystander, physician-staffed ambulance or helicopter, defibrillation by paramedics, type of advanced airway, number of administrations of adrenaline, pre-hospital ROSC, and time from call to hospital), and hospital capability of ECPR. Hospital capability for ECPR was defined as “high,” “middle,” or “low,” according to the tertiles of the number of patients who received ECPR in the last two years. Although we adopted the same variables to calculate the time-dependent propensity score for both shockable and non-shockable cohorts, the scores were calculated separately for each rhythm cohort.

Based on the time-dependent propensity scores, we performed 1:1 sequential matching of the patients who received ECPR at any given minute from minute 0 to minute 120 (patients treated with ECPR) to a patient who was at risk of receiving ECPR within the same minute (control) with replacement. In this sequential matching, at-risk patients (controls) included those who were still undergoing resuscitation at the emergency department and had not received ECPR before or within the same minute. Matching with replacements of unexposed patients meant that matched controls with no ECPR at each time were allowed to match again later until they received ECPR. Therefore, at-risk patients also included those who received ECPR later, as matching was not dependent on future events. In sequential matching, we set the caliper width for nearest-neighbor matching at 0.2 standard deviations of the propensity score, as recommended in previous literature [31]. We calculated standardized differences to evaluate the balance of variables in each propensity score-matched cohort. We considered a standardized mean difference of < 0.25, as suggested in the literature [31].