Map the themes: co-occurrence, thematic maps and trends

The conceptual structure shows what a field is about: its themes, how they connect and how they change. It is often the most interesting part of a bibliometric review, and the part where careful interpretation matters most.

How co-occurrence networks are built

Every keyword is a node. Two keywords are linked when they appear in the same document, and the link weight is the number of documents they share. Three choices shape the map:

• Items and thresholds: keep the 50 to 100 most frequent terms for a readable map, and raise the minimum link weight to remove chance links.
• Normalisation: raw counts favour frequent terms, which would link to everything. Association strength, the ratio of observed to expected co-occurrences, is the recommended default (van Eck and Waltman, 2009) and matches VOSviewer.
• Clustering: the Louvain algorithm groups terms so that links within clusters are stronger than expected by chance. A higher resolution gives more, smaller clusters.

Reading the co-occurrence network

Each colour is a cluster, which you interpret as a theme. In the demo data, look for clusters around health records and data quality (electronic health records, data quality, interoperability), mobile and remote care (mhealth, mobile apps, wearable devices, telemedicine, remote monitoring), technology adoption (technology acceptance, UTAUT, usability, trust) and artificial intelligence (machine learning, deep learning, explainable AI, large language models). Name each cluster by reading its most frequent terms and the most cited documents that use them. The overlay view colours terms by their average publication year, so recent terms stand out; the density view shows where research concentrates.

The thematic map

The thematic map (Cobo et al., 2011) places each cluster on two axes. Centrality (x-axis) is the strength of a theme’s links to other themes: how important it is to the field as a whole. Density (y-axis) is the strength of the links among its own terms: how developed it is internally.

The quadrant of a theme is a starting point for discussion, not a verdict. A theme in the lower left could be emerging (large language models, generative AI and ChatGPT in the demo data) or declining (older interoperability topics); trend topics and thematic evolution tell which.

Thematic evolution

Split the timespan into periods at meaningful cut years (for example, 2008–2015, 2016–2020 and 2021–2025), and the Studio builds a thematic map for each period and links themes that share keywords. The Sankey diagram shows themes splitting, merging, appearing and disappearing. Keep enough documents in each period (at least 50 to 100) for stable clusters.

Trend topics and burst detection

For each term, the trend topics chart shows the interquartile range of the publication years of the documents that use it. In the demo data, health information exchange (median year 2012) and electronic health records (2013) are early topics; technology acceptance and UTAUT peak around 2016; telemedicine (2022) and patient satisfaction (2022) rose with remote care; and deep learning and neural networks (median 2024), large language models and artificial intelligence (2023) are the newest fronts. Burst detection (Kleinberg, 2003) finds terms whose use rose sharply for a period and reports the start, end and strength of each burst.

Factorial analysis

Multiple correspondence analysis places keywords on a two-dimensional map so that terms used by similar documents are close together; hierarchical clustering then groups them into topics. Report the share of inertia explained by the two dimensions and interpret clusters rather than single positions.

Choosing the keyword field

Author keywords reflect how authors frame their own work; Keywords Plus (Web of Science) and index keywords (Scopus) are assigned by the database and give broader coverage; title and abstract terms help when keywords are missing. Use one field consistently and state which.