Prediction of potential organ donation after circulatory death in Scottish patients
Background: The Scottish government has set an ambitious target to increase authorisation rate for donation after circulatory death (DCD) to 80% from 57% in 2016. The process is complex and involves referral, screening, estimation of time to asystole after withdrawal of life saving treatment (WLST) and family approach. There is evidence that if we streamline this process by using a screening tool initially and then accurately estimate time of death after WLST, we can make family discussions more meaningful, increase referrals and reduce DCD stand downs.
Aim: To improve the DCD process by increasing the accuracy of estimation of imminent death, a crucial factor that determines whether the referral proceeds or not.
Methods: We performed a retrospective observational study looking at all DCD referrals for a period of a year in Scotland. The data is collected routinely by the specialist nurses in organ donation at referral and thereafter. The variables analysed were: age, FiO2, spontaneous respiration rate, presence of cough, presence of gag reflex, if the patient was extubated, need for sedation, sex, vasopressor use and ventilation mode. All statistical analyses were carried out in R version 3.5.1 (2018-07-02), using the Regression Modelling Strategies (RMS) R package version 5.1-2 and the Tangram R package version 0.6.2.
All simple statistical comparisons were carried out using Kruskal-Wallis testing for the grouped numeric analysis and Pearson chi-squared analysis for any categorical data. The survival analysis was carried out using Cox proportional hazards modelling.
The main model was built using a standard unadjusted, adjusted strategy. Thirteen unadjusted Cox regression models were then constructed. Wald test statistics with a p-value < 0.1 were used to test whether a candidate covariate should be taken forward into a full adjusted model. A p-value of < 0.05 was used to test if a covariate was significant in the final model. There was a total of 76 referrals with 64 becoming asystolic after WLST and 55 proceeding to successful donation.
Results: A total of 76 referrals were made where the median age amongst the adult patients was 59 and there were 47 males and 29 females. After adjustment, it was found that all modes of ventilation other than CPAP significantly increased the likelihood of asystole within 240 minutes of WLST.; χ2 = 6.438, p = 0.011 with an adjusted hazard of 2.375. Of note, fraction of inspired oxygen (FiO2) was significant in the unadjusted model (p = 0.01), however when adjusted for variables it lost its significance (p = 0.072).
Discussion: The validity of physiological variables to predict asystole within the time frame of 240 minutes was studied. When simply comparing spontaneous versus mandatory modes of ventilation there was no difference between the two groups. However, when comparing all other modes of ventilation to CPAP, there was a 137.5% increase in likelihood of asystole within 240 minutes (hazard estimate 2.375 – see figure 2). Despite using a year of national data, these results likely represent an underpowered study. This is evidenced by the covariate FiO2 which initially showed significance in the unadjusted model. The results suggested the individual contribution of the covariate was to increase likelihood of asystole at 240 minutes by 30% for every increase of 0.15 in FiO2 (Figure 1). However, when corrected for confounders, the significance was lost due to low power. There was a disparity in our cohort of how many patients became asystolic within 240 minutes with far more achieving this than not which may reflect clinician bias.
Conclusion: We were able to demonstrate that ventilation modes other than CPAP significantly increase the likelihood of asystole within 240 minutes of WLST. We also found that an increasing FiO2 trended towards significance as a predictive parameter. The limitations of this study include small numbers and possibility of bias (i.e. patients not being referred because they are deemed unlikely to die within the time frame). A larger study looking at all deaths in intensive care, not just referrals and using machine learning would help in predicting time to asystole, increasing not just successful DCD numbers but reduce DCD stand downs and improve family approach for authorization.