| 14:30-15:30 | Hessel Peters-Sengers (AUMC) - ICU-acquired complications (slides)
Acquired complications in the intensive care unit (onset > 48 hours after admission) are common, for example acute respiratory distress syndrome (ARDS), infections, and acute kidney injury (AKI). While its presence is associated with higher mortality, the adage “correlation does not imply causation” suggests many patients may die with, rather than from, such complications. Appropriate quantification of the impact and burden of these complications are imperative to understanding its severity and the importance of additional preventive measures and timely treatment. This study aimed to estimate the attributable effect of an ICU-acquired complication on mortality, with acute kidney injury as an example. We studied consecutive adult patients with a length of stay of at least 48h in which AKI was not present in two tertiary intensive care units in the Netherlands from 2011 to 2022. We adjusted for the evolution of disease severity and possible nephrotoxic medications until onset of ICU-acquired AKI using marginal structural modeling via inverse probability weighting, and calculated the time-dependent population-attributable fraction of ICU mortality. In all ICU-acquired AKI cases and a random selection of controls (who are at least 48h in the ICU), we also measured sequential plasma proteins across 6 different pathophysiological domains to gain insight in the predisposition to an ICU-acquired event. |
| 15:30-16:30 | Richard Post (TU/e) - Flexible Machine Learning Estimation of Conditional Average Treatment Effects (slides)
Causal inference from observational data requires untestable identification assumptions. If these assumptions apply, machine learning (ML) methods can be used to study complex forms of causal effect heterogeneity. Several ML methods were recently developed to estimate the conditional average treatment effect (CATE). If observed features cannot explain all heterogeneity, the individual treatment effects (ITEs) can still seriously deviate from the CATE. In this talk, we will illustrate the possible difference between the ITE distribution and the individualized CATE distribution by presenting scenarios with varying conditional ITE variance. If the distribution of the ITE equals that of the CATE, the observed difference in conditional variance between treated and controls should be small. If they differ, an additional causal identifiability assumption is necessary to quantify the heterogeneity not captured by the distribution of the CATE. The conditional variance of the ITE can be identified when the ITE is independent of the outcome under no treatment given the measured features. Under this assumption, we extend the causal random forest algorithm to illustrate how ML methods may be used to estimate the conditional variance of the ITE from observational data. For the scenarios where the ITE and CATE distributions differ, we show that the extended causal random forest can appropriately estimate the variance of the ITE distribution, while the traditional causal random forest fails to do so. Finally, we will discuss the impact of the violation of the untestable identifiability assumption on the performance of the extended random forest. |