Structural differences between REM and non-REM dream reports assessed by graph analysis

Abstract

Dream reports collected after rapid eye movement sleep (REM) awakenings are, on average, longer, more vivid, bizarre, emotional and story-like compared to those collected after non-REM. However, a comparison of the word-to-word structural organization of dream reports is lacking, and traditional measures that distinguish REM and non-REM dreaming may be confounded by report length. This problem is amenable to the analysis of dream reports as non-semantic directed word graphs, which provide a structural assessment of oral reports, while controlling for individual differences in verbosity. Against this background, the present study had two main aims: Firstly, to investigate differences in graph structure between REM and non-REM dream reports, and secondly, to evaluate how non-semantic directed word graph analysis compares to the widely used measure of report length in dream analysis. To do this, we analyzed a set of 133 dream reports obtained from 20 participants in controlled laboratory awakenings from REM and N2 sleep. We found that: (1) graphs from REM sleep possess a larger connectedness compared to those from N2; (2) measures of graph structure can predict ratings of dream complexity, where increases in connectedness and decreases in randomness are observed in relation to increasing dream report complexity; and (3) measures of the Largest Connected Component of a graph can improve a model containing report length in predicting sleep stage and dream report complexity. These results indicate that dream reports sampled after REM awakening have on average a larger connectedness compared to those sampled after N2 (i.e. words recur with a longer range), a difference which appears to be related to underlying differences in dream complexity. Altogether, graph analysis represents a promising method for dream research, due to its automated nature and potential to complement report length in dream analysis.

The Language of Dreams: Application of Linguistics-Based Approaches for the Automated Analysis of Dream Experiences

“…This approach allows the researcher to examine the overall structure of a graph, quantifying the pairwise relationships among its units and evidencing features that could not be captured by considering only its elements/nodes. Hence, when applied to language processing, graph analysis determines the non-semantic word-to-word structural organization of speech [ 27 ]. While this method could be applied to both written and spoken dream reports, previous studies mainly used it on oral speech samples.…”

“…In healthy subjects, graph analysis of mentation reports was used to investigate potential stage-specific structural properties. Specifically, Martin and colleagues investigated both differences in graph structure between REM and NREM dreams and the relationship between non-semantic graph attributes and dream report length, defined here as the total number of words used to describe a dreaming experience after excluding redundancies, repetitions, interjections, corrections and dreamer’s comments about the mentation [ 27 ]. By analyzing 133 reports collected from 20 participants in controlled laboratory awakenings from REM and N2 sleep, the authors found that REM dream reports were characterized by higher word count and graph connectedness (expressed in terms of LCC and LSC; see Table 1 ) as compared to N2 dreams.…”

“…In this respect, it is crucial to define and describe specific rules for the identification of such elements for the purpose of reproducible results. However, the removal of elements such as speakers’ comments or corrections is (at present) not automatable and must be performed manually, and preferentially by operators blind to the hypotheses and experimental conditions of interest [ 27 ].…”

“…In fact, word embedding algorithms necessitate large amounts of data—possibly in the order of several million words—to produce highly stable and reliable results and, as such, are more computationally demanding. In this respect, it should be noted that free dream reports typically include ~100 words on average (although values may change considerably depending on several factors, such as the type of report or the sleep stage the oneiric experience was conceived in [ 24 , 27 , 76 ] and thus, several thousand or even tens of thousands reports may be necessary. Therefore, when relatively small datasets are analyzed, dictionary-based approaches may represent the only viable option.…”

We dream about typical themes in both REM and non-REM sleep.

“…There appear to be some differences in the content of REM and NREM mentation. REM mentation reports display more connectivity and more convoluted structures than do NREM (Stage II) mentation reports (Martin et al, 2020). In addition, the dream self in REM mentation reports resembles the waking self, whereas the dream self in NREM (slow wave sleep) mentation reports can take various forms, ranging from a thinking agent, through to a passive observer and an entity similar to the waking self (Occhionero et al, 2005).…”