What are we learning with Yoga? Mapping the scientific literature on Yoga using a vector-text-mining approach

Abstract

The techniques used in yoga have roots in traditions that precede modern science. Research shows that yoga enhances quality of life and well-being, positively impacting physical and mental health. As yoga gains acceptance in Western countries, scientific studies on the subject increase exponentially. However, many of these studies are considered inconsistent due to the diverse methodologies and focuses in the field, which creates challenges for researchers and hampers progress. This study aims to develop a comprehensive framework for existing literature on yoga, facilitating multidisciplinary collaboration and bringing new light to relevant aspects. Given the complexity of the subject, advanced modeling techniques are necessary. Contemporary artificial intelligence methods have advanced Bioinformatics, including text mining (TM), allowing us to employ vector representations of texts to derive semantic insights and organize literature effectively. Based on TM resources, we provided a better general understanding of yoga and highlighted the relationships between yoga practice and various domains, including biochemical parameters and neuroscience. It also reveals that practitioners can learn to engage with their bodies and environments actively, enhancing their quality of life. However, there is a lack of research exploring the mechanisms behind this learning and its potential for further enhancement. Vector TM has made it possible to bolster and improve human analysis. The set of resources developed allowed us to determine the mapping of the literature, the analysis of which revealed 4 dimensions (exercise, physiology, theory and therapeutic) divided into 9 cohesive groups, representing the trends in the literature. The resulting platforms are available to Yoga researchers to evaluate our findings and make their forays into the existing literature.

Fig 1. A simple representation of a phylogenetic tree.

Each letter (A-E) represents a leaf, while the lines depict branches. The points where branches meet are called nodes, and the most basal node in the tree is referred to as the root.

Fig 2. Schematic of the methodology.

The first stage consisted of surveying the literature on PubMed. Only texts with ‘YOGA’ in the title were filtered for the subsequent stages. The texts (title and abstract) were processed, and a vector representing each word, document, and logical expression was obtained. This representation permitted phylogenies, clustering, visualizations, the HTML-Text Mining (HTML-TM), and subsequent analyses.

Fig 3. Mapping of Yoga literature with all 113 logical expressions (LOGEXP) studied.

The phylogenetic tree presents the relationship of the 113 LOGEXP and divides the findings into nine groups with related topics. The tree characterizes the field of yoga as a reflection of the literature. Each branch presents a LOGEXP and its identification number. It also indicates the number of studies/documents (nDcs) and the number of titles (nTit) hit by the logical expression. Group II.BRAIN-RESPIRATORY-HEART is subdivided into four groups: A. Respiratory, B. Metabolic and Cardiovascular, C. Biomarkers, D. Neurological.

Fig 4. The circles correspond to the groups in Table 1. We vectorized each group by the mean of logical expressions (LOGEXP) vectors inside it.

To define the centers of the circles, a t-SNE dimension reduction is applied, and the area is the total of different papers accessed by the corresponding group. The observed overlapping is illustrative and represents possible interconnections between them.

Fig 5. Dimensional Scatterplot Visualization of Word Proximity.

It is possible to obtain the closest words or studies within the same vector space from a vectorized word. From the word ‘PARASYMPATHETIC’ as an example, we can receive: A. phylogenetic tree associating the word with the 40 closest words according to Yoga literature and B. list of the closest papers to the word. Note that the tree shows words related to the subject, such as ‘NERVOUS’, ‘AUTONOMIC’, and physiological characteristics modulated by the PARASYMPATHETIC system, such as ‘HEART’, ‘RATE’, ‘CARDIAC’, ‘PRESSURE’, ‘HEMODYNAMIC’, related to the cardiovascular system, and ‘RESPIRATORY’. Similarly, the listed papers present the cardio-respiratory and autonomic nervous systems as the central subject of the study. The references for articles are: [,,,,,–76].

Fig 6. t-SNE visualization of vectorized documents.

Each dot represents the ‘TITLE+ABSTRACT’ for each document. A. Semantic searches for ‘HEART’ (green), ‘CARDIOVASCULAR’ (magenta), and ‘BLOOD’ (yellow). The search considers 40 hits closest to each query. B. Distribution of 300 nearest hits for six semantics searches listed in the legend. Note that in B, the overlap between the topics of ‘NEURO’ (black) and ‘LEARNING’ (green) is visible, as well as between ‘BREATH’ (blue) and ‘CARDIO’ (yellow), and more subtly between ‘MOOD’ and ‘BALANCE’.