• 2-way mixed ANOVA
• affective computing
• anxiety
• aracnophobia
• biofeedback
• EDA
• fear
• HRV
• phobia
• physiological habituation

Mental and emotional disorders are revealing as a major threat to health and well-being in modern societies. Real-time biofeedback is a promising therapy for many affective disorders, such as phobias. But it requires accurate features from mega-short recordings of 1 minute or less. This is a challenge for current HRV features.
In this work, HRV and EDA signals are analyzed from aracnophobic patients. Two different datasets are used, belonging to past research. The first dataset uses subliminal stimulation and compares the response between spiders and non-phobic animals, for subjects with varying SPQ. The second dataset uses 1-min exposure with phobics and assess the effect of training with HRV-biofeedback and having or working memory task during exposure, as well as comparing the response between a cognitive focus and a positive emotional focus.
In both datasets, the signals are preprocessed. Linear and nonlinear features are extracted. 2-way mixed ANOVA is performed for each feature and differences across the between-subjects and within-subject effects are reported. Any potential interactions between the effects are also taken into account.
Dataset1 results are in accordance with previous analysis. Additionally, Visibility Graph features are employed, yielding novel results on the factors effect and the detection of physiological habituation. It is also concluded the EDA has no remarkable latency and the useful information can be captured in the first 5 seconds. Moreover, the effect of conscious perception of the stimuli and biases due to the experimental protocol design in the fear response is discussed and its effect on the conclusions drawn is considered.
Analysis of dataset2 is novel and reveals that making the subjects focus on positive emotional facets of the spider clips has a greater effect in fear reduction than establishing a cognitive focus. In addition, some hypotheses made by the dataset experimenters are tested: partial evidence is found that subjects training with HRV-Biofeedback before exposure have greater fear reduction than those training with pseudo-biofeedback. Moreover, results suggest that doing a low-load working memory task during exposure to reduce fear could be counterproductive.
Based on the results from both datasets, VG features are found to be promising predictors for classifiers/decoders. Besides, we formulate 3 hypotheses regarding their potential role as ANS correlates.
Finally, some recommendations are given regarding which features to employ in a real-time HRV/EDA-Biofeedback system.