JMIR Publications

ABSTRACT

Background:

Social isolation and loneliness (SIL) are associated with critical health consequences but are difficult to measure in healthcare settings because they typically appear in unstructured text. Natural language processing (NLP) offers a promising approach to identify these constructs at scale, but its current applications in this domain have not been systematically characterized.

Objective:

To investigate how NLP is being used to study SIL, identify gaps, and outline priorities for advancing rigorous NLP-based measurement of these constructs in health research.

Methods:

A scoping review was conducted following the PRISMA-ScR guidelines. Six bibliographical databases (Ovid MEDLINE, Embase, Scopus, Web of Science, APA PsycINFO, ProQuest Dissertations & Theses Global) and two preprint servers (bioRxiv, medRxiv) were searched from inception to June 18, 2025. Reviewers independently screened abstracts and full texts, data were double-charted using a standardized form, and final results were synthesized using structured template analysis.

Results:

A total of 63 studies published between 2019 and 2025 met the inclusion criteria. Most were conducted in the US (27/63, 42.9%) and used cross-sectional designs (37/63, 58.7%). Studies mostly targeted older adults (31/63, 49.2%), used survey data (26/63, 41.3%), and focused on loneliness (32/63, 50.8%). Most studies (44/63, 69.8%) did not use a validated loneliness scale; among those that did, the UCLA Loneliness Scale was most common (16/63, 25.4%). Classification (24/63, 38.1%) was the most frequent NLP application. Rule-based (32/63, 50.8%) and traditional machine learning (18/63, 28.6%) approaches predominated, but large language models (16/63, 25.4%) and transformer-based models (14/63, 22.2%) increased over time. External validation was rare (2/63, 3.2%), code was shared in only 19% of studies (12/63), and 25.4% (16/63) addressed bias in their data or analysis.

Conclusions:

NLP applications for SIL are expanding rapidly but rest on narrow methodologies with limited validation measures and demographic groups. Advancing the field requires tying model development to validated measurement of the target constructs and adopting established reporting frameworks such as TRIPOD+AI and MI-CLAIM.