((A-190)) (A-190) Using causal trees to explore the heterogeneity of effects

Background: Causal tree allows to explore the complex relationships between treatments and clinical outcomes, by identifying possible heterogeneous effects of treatments in specific subgroups of patients. This methodology will be used to compare the effectiveness of a remote monitoring system according to patient characteristics in Real World Evidence (RWE) data.

Objectives: To use the causal tree methodology on RWE data

Methods: Patients remotely monitored by device A (cases) were matched to patients remotely monitored by device B (controls) to compare the occurrence of adverse events since the inclusion. Causal trees were constructed using the following characteristics: gender, age, year of inclusion, comorbidities, treatment history, social deprivation index and history of hospitalizations of interest. Several parameters were explored and their impact on the structure of tree was studied, in particular the size of the leaves. Honest causal tree were also implemented. The effect of treatment within each leaf was assessed using a generalized linear model with a binomial distribution and a logit link and results were expressed in Odds-Ratios (OR) and their confidence interval (CI). All analyses were weighted according to the weight of cases and controls.

Results: The overall comparison of the occurrence of adverse events between case and controls was not significant (OR [CI95%] = 0.94 [0.78;1.12], p-value = 0.46). The causal tree identified three patient profiles using the history of hospitalizations of interest and the social deprivation index, including a profile with a significant treatment effect (see Figure). Patients with no hospitalization history of interest and a high index of social deprivation showed a deleterious effect of treatment A compared with treatment B, with an odds ratio of 1.81 [1.11 ; 2.98] and a p-value of 0. 02. The size of the leaves had small impact on the results obtained, with the algorithm converging on the same tree above a certain threshold. The honest trees were unstable, due to the small number of patients.

Conclusions: Using a limited number of patients, the causal trees allow to explore sub-group analyses that would not have been possible manually. This methodology highlights key causal factors, for a better understanding of complex relationship between exposure and patients characteristics.