Specific phobia affects 7.7% of the population at some point in their lives. Virtual reality exposure therapy has demonstrated potential as a viable treatment; tracking patients’ progress, however, remains a challenge. Recent work has demonstrated that by employing machine learning, multimodal physiological sensing can be leveraged to address this need. Specifically, features extracted from the electrocardiogram (ECG) and electrodermal activity (EDA) can be used to accurately discriminate periods of phobic anxiety from rest. Herein, we extend the state of the art by demonstrating that features extracted from the respiratory effort (RSP) signal can further enhance this capability. This is investigated by adding respiratory markers to the classifier and evaluating feature importance via permutation and removal. We thereby find that respiration significantly contributes to classifier performance. In particular, five respiratory markers rank within the top 10 features based on permutation importance; notably, four of which are respiratory variability metrics. Additionally, respiration features contribute significantly to model generalizability, corroborated by the 7% decrease in accuracy caused by solely using ECG and EDA features (i.e., removing respiration).