AI Research on Eating Disorders Is Booming, but Almost Nothing Reaches the Clinic

Artificial intelligence research on eating disorders has expanded dramatically over the past decade, but almost none of it has made the leap from the laboratory to the clinic. That is the central finding of a scoping review with bibliometric analysis published in Eating and Weight Disorders – Studies on Anorexia, Bulimia and Obesity, which mapped every eligible empirical study published between 2016 and 2025. Of the 66 articles that met the inclusion criteria, 77.3 percent described models tested exclusively in research settings, with no external validation or clinical pilot. Only two studies, or 3 percent, described clinically deployed systems.

The review team searched PubMed (MEDLINE), the Web of Science Core Collection, and Scopus on 20 November 2025, using Boolean strings that combined artificial intelligence and machine learning terms with eating disorder diagnoses. Across the three databases, 837 records were identified. After deduplication removed 310 duplicates, 527 unique records underwent title and abstract screening, 306 were excluded, and 221 full-text articles were assessed. Sixty-six met all inclusion criteria: original empirical research using artificial intelligence or machine learning as the primary analytical method in an eating disorder population. Reviews, editorials, letters, protocols, and commentaries were excluded, as were studies where artificial intelligence served merely as a screening or administrative tool.

The growth curve tells the story of an emerging field. Publication output rose from a single article in 2016 to 16 in 2025. Output was lowest between 2016 and 2019, ranging from zero to two articles per year, then climbed to seven in 2020, six in each of 2021 and 2022, 14 in 2023, 12 in 2024, and 16 in 2025. The authors classified each study along predefined taxonomies covering methodology, clinical application, data source, and implementation readiness, with two authors independently double-coding the full corpus and reaching substantial agreement (Cohen’s kappa of 0.80).

Methodologically, traditional machine learning dominated overwhelmingly, accounting for 52 studies, or 78.8 percent of the corpus. Natural language processing came second with 13 studies (19.7 percent), and a single conversational artificial intelligence application made up the remaining 1.5 percent. Notably, no study was assigned to the pre-specified deep learning category, a zero count the authors attribute to their task-based classification rule: neural-network and transformer-based models applied to text data were coded as natural language processing, while those applied to structured, neuroimaging, or sensor data were coded as machine learning. Deep learning architectures were therefore present in the corpus but subsumed within other categories rather than counted separately.

The clinical applications clustered around prediction. Risk prediction was the most common purpose, appearing in 22 studies (33.3 percent), followed by diagnosis in 17 studies (25.8 percent) and treatment outcome prediction in 14 studies (21.2 percent). Detection and screening accounted for 15.2 percent, while symptom monitoring, despite its promise for real-time tracking through wearables and smartphones, received just 4.5 percent. Clinical data were the most frequent data source (42.4 percent), followed by neuroimaging (18.2 percent) and social media (15.2 percent). Ecological momentary assessment, wearable sensors, and survey data each contributed smaller shares.

The translational picture is the review’s most sobering dimension. Beyond the 77.3 percent of studies confined to laboratory settings, a structured reporting appraisal using the TRIPOD+AI statement revealed systematic weaknesses in exactly the elements needed for clinical deployment. Across the 58 appraisable prediction-model studies, a mean of 16.8 of 27 TRIPOD+AI items were fully reported. Reporting was strong for data sources, analytical methods, objectives, and limitations, but critically deficient elsewhere: none of the studies documented prospective registration, only 2 percent reported fairness or subgroup performance analysis, 3 percent provided sample-size justification, 7 percent reported full calibration assessment, and just 10 percent reported external validation on an independent dataset. Confidence intervals around performance estimates were fully reported by only 17 percent.

These gaps matter because models developed on small samples and validated only internally are particularly vulnerable to overfitting and optimistic performance estimates. The median sample size across the corpus was 281, with an interquartile range of 70 to 728, a wide spread that underscores the variability. The authors point to work by Forrest and colleagues, who found that machine learning models offered only minimal advantage over traditional statistical approaches in predicting binge eating disorder treatment outcomes, and that seemingly excellent performance in some analyses was attributable to methodological inflation. Promising findings are thus not uncontested, and the near-absence of external validation tempers prevailing implementation-readiness claims.

The geography of the field is strikingly concentrated. Contributing authors were affiliated with institutions in 14 countries, with the United States accounting for the largest share at 33.3 percent, followed by Australia (12.1 percent), China (10.6 percent), Spain (9.1 percent), and Germany (7.6 percent). No contributions were identified from Africa, South America, or Southeast Asia. Because attribution was based on corresponding author affiliation, the authors caution that contributions from underrepresented regions embedded within multinational collaborations may not be fully captured. Even so, models trained exclusively on high-income Western data may perform poorly elsewhere, making broader geographical representation a prerequisite for clinically valid models across diverse populations.

Population coverage shows parallel gaps. Adults were the primary target group in 59.1 percent of studies, while only 12.1 percent targeted adolescents exclusively, despite eating disorder onset typically occurring during adolescence and earlier intervention being associated with significantly better long-term outcomes. Mixed or general eating disorder samples were the most studied populations (36.4 percent), followed by anorexia nervosa (31.8 percent). Studies focusing exclusively on bulimia nervosa were rare at 3.0 percent, positioning bulimia-specific prediction and treatment models as an underexplored area.

The authors conclude that artificial intelligence research in eating disorders is growing rapidly but remains geographically concentrated, methodologically dominated by traditional machine learning, and largely confined to laboratory settings. Closing the distance between research output and clinical implementation, they argue, will require independent model validation, prospective evaluation, calibration reporting, subgroup fairness analysis, and sustained collaboration between computational researchers, clinicians, people with lived experience, and regulatory bodies. The findings are descriptive rather than evaluative: this review maps where the field is investing its attention, not whether any given artificial intelligence system actually works.

Subject of Research: A scoping review and bibliometric analysis of artificial intelligence applications in eating disorder research from 2016 to 2025

Article Title: Artificial intelligence in eating disorder research: a scoping review with bibliometric analysis of trends, methodologies, and translational readiness (2016–2025)

Article References: Caliskan, Y., Sari, M., & Binokay, H. (2026). Artificial intelligence in eating disorder research: a scoping review with bibliometric analysis of trends, methodologies, and translational readiness (2016–2025). Eating and Weight Disorders – Studies on Anorexia, Bulimia and Obesity, 31(1), Article 89. https://doi.org/10.1007/s40519-026-01905-w

Image Credits: AI Generated

DOI: 10.1007/s40519-026-01905-w

Keywords: artificial intelligence, machine learning, eating disorders, scoping review, bibliometrics, anorexia nervosa, natural language processing, clinical implementation, translational research, TRIPOD+AI, Artificial, intelligence