A complex relationship is there between movement dynamics and related physiological responses in chronic disease patients. To comprehend the complexity, advanced analytical techniques is required for putting effect into high-dimensional, time-series data. To meet this need, this work introduces a novel method based on spectral-temporal features derived from physiological signals for improving chronic pain state recognition. Particularly, we employ spectrogram analysis to transform body movement and surface electromyography (sEMG) signals of the EmoPain dataset into deep learning-friendly representation. To better capture the latent patterns in such spectrogram representations, we adopt an autoencoder-based approach. The model learns a compressed lower dimensional representation of the input, which is used for classification. Our results show that the learned features of the autoencoder consistently outperform the learned features of a convolutional neural network (CNN) in terms of high classification accuracy for capturing chronic condition variance. This is indicative of the ability of autoencoders to learn informative information from spectral representations of physiological signals towards better chronic pain state estimation.