Trends and Themes in Artificial Intelligence Applications for Obesity: A Bibliometric Mapping Study

Literature search strategy

This bibliometric analysis focused on publications between 2015 and 2025 related to AI applications in the context of obesity. A comprehensive query combining keywords related to both AI and obesity was used to search Web of Science (WoS) and Scopus databases. WoS and Scopus were queried on July 15, 2025; records dated up to this search date were included. The search was conducted within topic fields, and the same logic was adapted for each database’s syntax. WoS and Scopus were selected due to their comprehensive coverage of scientific literature across diverse disciplines, including health sciences, computer science, and engineering, which are highly relevant to the interdisciplinary nature of AI in obesity research. While other specialized databases exist, these two platforms offered the broadest interdisciplinary capture for a bibliometric overview. The core search string combined terms such as “artificial intelligence,” “machine learning,” “deep learning,” and “neural networks” with “obesity,” “bariatric surgery,” “weight management,” and “metabolic syndrome” using appropriate Boolean operators and wildcards to ensure broad yet focused retrieval. To ensure consistency and quality of included studies, the search was limited to publications in English and restricted to original articles and reviews. Details of the full search strategy are provided in Table 1.

OPEN IN VIEWER

Table 1.

Boolean Search Syntax for Each Database

Database	Search syntax
Web of Science	TS = (((obesit* OR overweight OR “body mass index” OR BMI OR “adiposit*” OR “weight loss” OR “weight management” OR “bariatric surgery” OR “metabolic surgery” OR “sleeve gastrectomy” OR “gastric bypass” OR “metabolic syndrome”)
	NEAR/10
	(AI OR “artificial intelligence” OR “machine learning” OR ML OR “deep learning” OR DL OR “neural network*” OR “natural language processing” OR NLP OR “computer vision” OR “reinforcement learning” OR “predictive model*” OR “data mining” OR “big data” OR “smart algorithm*” OR “large language model*” OR LLM OR “generative AI” OR chatbot* OR “conversational AI” OR “digital health” OR eHealth OR mHealth OR “wearable device*” OR sensor* OR IoMT OR “internet of medical things” OR “personalized medicine” OR “precision medicine” OR “decision support system*” OR “diagnostic tool*” OR “prognostic model*” OR “risk prediction” OR “health monitoring” OR “therapy adherence” OR “patient engagement”))
	OR
	((“obesity” OR “overweight” OR BMI) AND (“AI in health care” OR “medical AI” OR “clinical AI” OR “health AI”)))
Scopus	TITLE-ABS-KEY (((obesit* OR overweight OR “body mass index” OR BMI OR adiposit* OR “weight loss” OR “weight management” OR “bariatric surgery” OR “metabolic surgery” OR “sleeve gastrectomy” OR “gastric bypass” OR “metabolic syndrome”)
	W/10
	(AI OR “artificial intelligence” OR “machine learning” OR ML OR “deep learning” OR DL OR “neural network*” OR “natural language processing” OR NLP OR “computer vision” OR “reinforcement learning” OR “predictive model*” OR “data mining” OR “big data” OR “smart algorithm*” OR “large language model*” OR LLM OR “generative AI” OR chatbot* OR “conversational AI” OR “digital health” OR eHealth OR mHealth OR “wearable device*” OR sensor* OR IoMT OR “internet of medical things” OR “personalized medicine” OR “precision medicine” OR “decision support system*” OR “diagnostic tool*” OR “prognostic model*” OR “risk prediction” OR “health monitoring” OR “therapy adherence” OR “patient engagement”)) OR
	((“obesity” OR “overweight” OR BMI) AND (“AI in health care” OR “medical AI” OR “clinical AI” OR “health AI”)))

Data collection and processing

The initial search retrieved 5454 articles from Scopus and 2966 articles from Web of Science, resulting in a total of 8420 records. Of these, 107 records were excluded as they were not published in English, and 628 records were removed because they did not meet the study type criteria (i.e., not original articles or reviews). To ensure methodological integrity and consistency in applying the exclusion criteria, two independent authors randomly evaluated 5% of the initial retrieved records. This assessment focused on confirming the accuracy of classification based on publication type and relevance to AI in obesity. Discrepancies were resolved through discussion to reach a consensus, ensuring a uniform application of criteria. While a formal interrater reliability statistic like Cohen’s Kappa was not calculated for this preliminary quality check, this consensus-based approach aimed to mitigate potential biases in the screening process. We did not compute kappa, as the intent was calibration rather than formal dual screening. Following this review, an additional 295 records were excluded for being unrelated to the context of AI in obesity. The remaining 7390 articles were screened for duplicates, and after removing 1497 duplicates, a final set of 5893 unique articles was included for analysis. The selection and screening process is summarized in a Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram (Fig. 1). Bibliometric science-mapping is appropriate for synthesizing large, multidisciplinary corpora at scale. We therefore used standard performance and network analyses to objectively characterize nearly 6000 records without duplicating narrative review methods. We used a triangulated approach: (1) descriptive bibliometrics (production and citations), (2) science-mapping (coauthorship and keyword co-occurrence), and (3) brief thematic synthesis of clusters. This combination provides both structural and interpretive insight.

OPEN IN VIEWER

FIG. 1.

PRISMA flow diagram of study selection process.

Tools and analysis

All analyses were performed using R version 5.1. Bibliometric mapping and statistical summaries were generated using the following tools: •

Bibliometrix (v5.0): used for performance analysis and science mapping.

Performance analysis and science mapping were conducted with bibliometrix/Biblioshiny (R) and VOSviewer, standard toolchains for bibliometric synthesis in health sciences. ¹³ Coauthorship and co-occurrence networks used fractional counting; layout followed LinLog/modularity optimization with default stabilization.

Author keywords were normalized and ranked by frequency. The rank–frequency curve exhibited a Zipfian heavy tail; therefore, to improve network readability and ensure stable co-occurrence estimates, we retained only head terms (≥50 occurrences) for keyword mapping. ¹⁴ Theme labels and summaries were reviewed by clinical coauthors with bariatric/obesity expertise to ensure field relevance.

The dataset was divided into three periods: 2015–2018, 2019–2021, and 2022–2025. Author and index keywords were lowercased, cleaned (punctuation and duplicates removed), and counted within each period, excluding generic or noninformative terms. Period-specific keyword frequencies were compared to identify dominant themes and highlight the emergence of representative concepts such as machine learning, deep learning, bariatric surgery, computer vision, and large language models.

Author productivity followed Lotka’s law (few authors contribute most output); to avoid sparse ego-networks and label overlap, we restricted the coauthorship map to authors with ≥5 publications. ¹⁵

Annual scientific production and citations

A total of 5893 unique articles were identified from Web of Science and Scopus databases between 2015 and 2025, covering AI applications in obesity-related research. As shown in Figure 2, there was a consistent upward trend in the number of annual publications, with a marked acceleration beginning around 2019. Between 2018 and 2023, the number of publications more than tripled, indicating growing interest in the field.

OPEN IN VIEWER

FIG. 2.

Annual scientific production and average citations per year (2015–2025). The bar plots represent the number of articles published each year on artificial intelligence applications in obesity research between 2015 and 2025. The line graph (right Y-axis) displays the average citations per article for that specific year. The observed decrease in average citations for recent years reflects the expected citation lag for newer publications.

Citation trends followed a similar pattern. Older studies accumulated more citations over time, consistent with the time lag inherent to citation dynamics. Notably, articles published in 2020 exhibited the highest average citation rate per article, an observation likely reflecting a surge in digital health and AI research during the early stages of the COVID-19 pandemic, which spurred rapid engagement with newly published work.

The mean number of citations per article was 14.3, with a median of 6, and 17.4% of articles had not been cited at the time of analysis. Only a small number of articles exceeded 400 citations, indicating a concentration of influence among a few key publications.

Author collaboration network

Across 34,033 authors, productivity followed Lotka’s law; we, therefore, mapped only authors with ≥5 publications (n = 66) to focus on stable collaboration structure (Fig. 3). Most authors were clustered in small, institutionally affiliated groups, although a few larger hubs emerged in domains such as bariatric surgery, machine learning, and surgical informatics. OPEN IN VIEWER

The most prolific authors identified in the dataset were as follows: •

Mohan, Viswanathan, nine publications.

Author disambiguation using the AF field improved accuracy, especially for common surnames. While a few high-output authors were prominent, the overall network remained relatively sparse and decentralized, suggesting an emerging and still-fragmented field.

Thematic structure of keywords

Using VOSviewer, we identified 9868 keywords. In line with Zipf’s law, only the head terms (≥50 occurrences; n = 32) were included in the map to prune the Zipfian long tail. A co-occurrence map of keywords appearing in at least 50 publications revealed four major thematic clusters (Fig. 4):

OPEN IN VIEWER

FIG. 4.

Keyword co-occurrence network and thematic clusters. This network illustrates the co-occurrence of author keywords in AI-driven obesity research. Node size indicates keyword frequency, and edge thickness reflects co-occurrence strength. Four thematic clusters were identified: a red cluster focusing on surgical applications of AI, including bariatric surgery, risk prediction, and decision support; a green cluster centered on lifestyle and remote care through digital health, mobile apps, and wearables; a blue cluster highlighting conversational technologies such as chatbots, natural language processing, and large language models; and a yellow cluster representing precision health concepts like personalized medicine and predictive analytics. The network captures the multidimensional integration of AI into obesity and metabolic health research. AI, artificial intelligence.

These clusters reflect the multidimensional growth of the field, spanning both technical development and clinical integration.

Temporal evolution of research themes

When grouped chronologically, keyword patterns revealed a distinct thematic progression within the literature. During 2015–2018, research primarily addressed metabolic and epidemiological aspects of obesity, with emphasis on risk factors, glucose metabolism, triacylglycerol levels, and cross-sectional analyses. Mentions of “machine learning” were limited and exploratory, while “deep learning” and “natural language processing” appeared only sporadically. Between 2019 and 2021, AI-related terms became more frequent, indicating a transition toward clinical and perioperative applications. The keywords “machine learning,” “deep learning,” and “bariatric surgery” gained prominence, suggesting the adoption of predictive modeling and early computer-vision efforts in surgical contexts. In 2022–2025, the field expanded rapidly, with “machine learning” and “deep learning” dominating the keyword landscape, alongside the emergence of “ChatGPT,” “large language model,” and “natural language processing,” reflecting the integration of conversational and generative AI tools into obesity-related research. Across all periods, the increasing frequency of bariatric surgery–related terms demonstrated a sustained and growing clinical interest.

Most influential articles

•

Top five most cited works are listed below with their first authors and verified citation counts.

These publications span methodological overviews, bibliometric frameworks, and AI integration in both clinical and population health contexts.

Equity and representativeness

Because obesity care spans diverse ages, sexes, and comorbidity profiles, models developed on narrow samples may not generalize well to important subgroups; routine subgroup reporting and basic bias checks are advisable. ²¹ For vision models, deliberate testing in challenging conditions (e.g., smoke, bleeding, and uncommon anatomy) is informative. ²²

From prediction to care pathways

Models are most useful when tied to a clear next step—linking preoperative risk to optimization plans, using intraoperative prompts to support safety checks, or connecting discharge risk to structured follow-up; simple usability testing with clinical teams, followed by pragmatic evaluations, helps determine whether such tools change outcomes. ²³

Interpretation and implications

Field map: A corpus-level view linking four applied areas—surgical/perioperative AI, digital/remote care, conversational/NLP tools, and precision/phenotyping—to their specific evidence needs.

Collaboration signal: Demonstrates fragmentation of coauthorship networks, motivating shared benchmarks and simple, reusable analysis pipelines.

Temporal perspective: Trend-topic slices show a shift from epidemiologic foundations (2015–2018) to clinical/perioperative ML/DL (2019–2021) and then to scaled use with early LLM-adjacent work (2022–2025).

Emerging directions

Combining data types (imaging, text, and sensor/device data), adapting foundation or vision-language models to surgical tasks, and explaining results in ways that support clinical decisions are promising; these advances benefit from straightforward governance: clear data permissions, audit trails for model updates, and transparent reporting of computing demands. ²⁴

In conclusion, the bibliometric landscape of AI and obesity research is characterized by rapid growth and diverse thematic exploration. While the field is still maturing in terms of collaborative networks, the increasing volume of publications and the impact of highly cited articles underscore AI’s pivotal and expanding role in addressing the global obesity epidemic. Continued research, particularly fostering greater interdisciplinary collaboration, will be crucial in translating these advancements into impactful public health interventions.

Limitations

Despite its comprehensive scope, this bibliometric analysis has several inherent limitations as follows: •

Searches were performed on July 15, 2025; work indexed after this date is not included, and very recent articles will have had less time to accrue citations.