Assessment of the conditional exchangeability assumption in causal machine learning models: a simulation study