[Regression analysis to calculate the time point of ROSC-A feasibility study]

A regression model to estimate the duration from the onset of resuscitation efforts to the return of spontaneous circulation (ROSC) could help improving both resuscitation care and the quality control of registries. This study aims to evaluate the prediction accuracy and to identify challenges for future model development.

Regression models based on M5P, random forest (RF) algorithms and a linear regression (LR) modified using M5P were retrospectively developed using a Belgian cohort of 84 individuals in whom ROSC was achieved. Model performance was assessed using quality metrics, such as the correlation coefficient (CC), coefficient of determination (R²), and root mean square error (RMSE) in a cross-validation approach.

In the cohort 61.9% were male with a mean age of 65.7 years. A shockable rhythm was present in 27.7% of cases and the bystander cardiopulmonary resuscitation (CPR) rate was 48.2%. The no-flow time averaged 5.13 min. The mean time from CPR onset to first defibrillation was 7.81 min and to first medication administration 11.31 min. The ROSC occurred after an average of 16.8 min, the LR showed the highest correlation (0.73, 95% confidence interval, CI 0.72-0.74) and R² (0.53 [0.52-0.55]) along with the lowest RMSE (6.76 min [6.63-6.90]). The M5P yielded similar not significantly different values (CC 0.72 [0.70-0.73], R² 0.52 [0.50-0.53], RMSE 6.84 min [6.69-6.99]). In contrast, RF performed significantly worse (CC 0.62 [0.61-0.63], R² 0.38 [0.37-0.40], RMSE 7.89 min [7.82-7.96], all p < 0.01). Only LR showed no significant difference between predicted and actual values in terms of mean (p = 0.75) and variance (p = 0.15). The proportion of potentially prematurely terminated resuscitation attempts, defined as cases with actual ROSC occurring later than predicted ROSC plus RMSE, ranged from 13% (M5P) to 18% (LR).

The duration from the start of CPR to ROSC appears to be a process that is suitable for modelling with machine learning algorithms. At this early stage of development, the individual regression models did not demonstrate sufficient validity possibly due to low sample size and simplified data structure; however, the findings indicated potential for an application as a quality assurance tool to compare actual vs. predicted time to ROSC. Therefore, to increase the robustness the results require further evaluation in a larger cohort with additional variables and improved data quality based on the Utstein criteria.