Uncovering Heart Rate Response Patterns to Threat Pictures through Deep Latent Representation Learning with a Variational Autoencoder

Group-level averaging of psychophysiological data often obscures meaningful individual differences, masking response patterns that may explain variability in behavior and central nervous system activity. Identifying such patterns is particularly relevant in heart rate (HR) responses to threat, where subtle variations may reflect distinct coping mechanisms such as orienting or defense. Machine learning techniques that learn latent representations, particularly variational autoencoders (VAEs), offer powerful tools for revealing such hidden structures.This methodological report introduces a simple VAE-based approach for characterizing HR responses to threat pictures in 165 participants. To validate the method, simulations first demonstrated that the model accurately separated noisy sine and cosine waveforms. The VAE was then applied to empirical HR responses, mapping them into a two-dimensional latent space for subsequent cluster analysis, which was compared to clustering based directly on raw HR waveforms.The VAE revealed three distinct response profiles: (1) strong decelerators (fear bradycardia), (2) weak decelerators with late acceleration, and (3) immediate accelerators without a decelerative phase. In contrast, clustering raw HR waveforms identified only two groups. Clusters derived from the latent space were more coherent and exhibited greater within-group consistency. Finally, applying the pre-trained autoencoder to a small fear-conditioning dataset enabled characterization of distinct HR response patterns despite limited sample size. These findings show that even a basic autoencoder enhances the categorization of psychophysiological response patterns, offering a framework for linking individual autonomic variability to broader models of affective and defensive behavior.