The systematic review was registered in PROSPERO (CRD42024578082), and the protocol was published in JMIR Research Protocol [37]. The review followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [38], and the search strategy was reported in accordance with PRISMA-S [39]. We conducted a narrative synthesis without meta-analysis following the Synthesis Without Meta-analysis guidelines [40] due to substantial methodological heterogeneity across studies. Formal subgroup, heterogeneity, sensitivity analyses, reporting bias analysis, or certainty of evidence assessment were not conducted as the review did not involve meta-analysis and instead examined differences across studies descriptively. Synthesis focused on identifying patterns and trends in NLP methods, SDOH domains, data sources, and reported performance rather than pooling quantitative estimates. The interdisciplinary review team consisted of a librarian and researchers with expertise in health informatics, public health, and social sciences.

This work does not involve human participants and does not require approval from the IRB.

We included peer-reviewed literature (journal articles and full conference papers) published between January 2014 and November 2025. The selection of this timeframe was deliberate, as it strategically captures the entire modern deep learning era in NLP, from the rise of word embeddings (eg, Word2Vec) around 2014 to the development of the transformer architecture in 2017 and the subsequent explosion of LLMs. Only studies in English were included, although the underlying NLP data could be in any language. Eligible studies met the following specific criteria: (1) addressed a research question involving the design, development, and application of NLP in health data analysis for SDOH; (2) used text-based data; and (3) used NLP techniques or LLMs (open-source or commercial) related to SDOH. Studies using NLP in conjunction with other methods were eligible, provided NLP was a core component. Preprints (arXiv/bioRxiv), forewords, prefaces, table of contents, programs, schedules, indexes, call for papers or participation, lists of reviewers, lists of tutorial abstracts, invited talks, appendices, session information, obituaries, book reviews, newsletters, lists of proceedings, lifetime achievement awards, erratum, systematic reviews, scoping reviews, and notes were excluded.

We systematically searched 7 databases: PubMed, Scopus, Web of Science, PsycINFO, Health Source: Nursing/Academic, ACL Anthology, and IEEE Xplore. Searches were conducted independently in each database rather than through a multi-database platform. No study registries, manual browsing of web-based resources, citation chaining, or direct contact with authors or experts were undertaken as part of the search methodology. The search strategies were developed specifically for this review and were not adapted from previous reviews. No additional information sources or search methods were used.

The search spanned publications from January 1, 2014, to November 2, 2025. A health sciences librarian (HR) developed and iteratively refined the search strategy in consultation with the research team, using both controlled vocabulary and free-text terms. The final PubMed search strategy, including filters, is detailed in Textbox 1. Comprehensive search strategies for all information sources are provided in Multimedia Appendix 1.

Three reviewers (SR, ZZ, and YC) independently screened each study for eligibility by marking it as a “yes” (for inclusion), “no” (for exclusion), or “maybe” (in case of uncertainty about relevance) on the Covidence platform [41]. Two senior reviewers (AS and YX) resolved potential discrepancies during any screening step. Blinded voting ensured that reviewers did not view others’ decisions during the screening process. The reviewers retrieved eligible studies for second-stage review (full-text) using the same inclusion criteria and removed those that did not meet them. The final set of studies to include was approved by consensus of all reviewers.

Before formal data extraction, 3 reviewers (SR, ZZ, and YC) piloted a structured data extraction form using 5 sample studies to ensure clarity and consistency. Six independent reviewers (SR, ZZ, YC, AKP, ML, and SD) then conducted the full data extraction in teams of 2, with each member independently extracting half of the assigned studies. Ambiguities were resolved within the teams to maintain accuracy and consistency. The final data captured study metadata (eg, year of publication and type), dataset characteristics (eg, source, sample size, and type), NLP approaches (eg, models used and task type), SDOH domains addressed, and performance metrics (eg, precision, F₁-score, and recall). We treated these performance metrics as effect measures for outcomes of interest (ie, model performance). Due to heterogeneity in study design, NLP tasks, datasets, and evaluation frameworks, no single standardized effect size metric was applicable, and reported metrics were extracted and summarized descriptively without statistical transformation. Model performance metrics reported by individual studies were extracted as reported and treated descriptively. We summarized study characteristics in structured tables and used visualizations, including a PRISMA flow diagram, frequency plots, heatmaps, and bubble charts, to display study selection, SDOH domains, and NLP model performance patterns. We assessed risk of bias using a short checklist adapted from the Joanna Briggs Institute [42] and covered 6 items: reporting of population demographics (Q1), clarity of study aim (Q2), relevance to SDOH and NLP (Q3), use of a reference standard (Q4), reporting of evaluation measures (Q5), and acknowledgment of study limits (Q6). Each item was rated as yes, no, or unclear. Studies were classified as low (≥5 yes), moderate (3‐4 yes), or high (≤3 yes) risk of bias.

To assess visibility and identify recurring methodological, collaborative, and reporting features, we calculated a normalized citation ratio (NCR) for each study:

This approach adjusts for citation differences based on publication year.

Figure 1 shows the number of articles included in each phase of the screening process, resulting in the eventual inclusion of 142 studies (Multimedia Appendix 1 shows the search strategy results across databases). Figure 2 presents publication trends over the years, illustrating a growing research interest over time, with the majority of studies being published between 2023 and 2025 (89/142, 62.7%). The majority of the studies were published in journals (109/142, 76.7%), while the remaining were in conference or workshop proceedings (Figure 2C). Also, as shown in Figure 2B, single-site studies were more common overall, but there has been a recent rise in multi-site studies (data from multiple institutes). Most studies were rated as low risk of bias (116/142), 26 as moderate, and none as high risk (Table 1). Table S5 in Multimedia Appendix 1 presents a summary of studies. Also, because this review used descriptive narrative synthesis across highly heterogeneous studies, formal heterogeneity analyses, sensitivity analyses, reporting-bias assessment, and certainty-of-evidence grading were not conducted.

Across the 142 studies reviewed, over half used private datasets (81/142, 57.04%), while publicly accessible datasets were used in 29 studies (20.42%). A smaller number of studies used datasets that were available via specific data use agreements (7/142, 4.93%), accessible only for shared tasks (5/142, 3.52%), or available in partial form (subset available, 5/142, 3.52%). The most common data type was EHRs, reported in 93 studies. Eight studies used national datasets (Primary Land Use Tax Lot Output [PLUTO], CalEnviroScreen 4.0, etc). Social media sources, including Twitter, Reddit, and Facebook, were leveraged in 9 studies (5.1%). Other, less frequently reported types included research abstracts, web content, and interviews.

The majority of private datasets included institution-specific EHRs (eg, Columbia University Irving Medical Center; University of California, San Francisco; Johns Hopkins; Eskenazi; Medical University of South Carolina; UF Health; UNC; Kaiser Permanente Southern California) and clinical data warehouses tied to academic or health systems. The Veterans Affairs health system was particularly prominent, covering datasets such as: Veterans Affairs Corporate Data Warehouse [46,47,75,83,89,90,94,135], Veteran Health Administration notes [114], Supportive Services for Veteran Families [75], administrative data [136], Veterans Aging Cohort Study [141]. Several datasets incorporated augmented sources to enrich available information, such as LexisNexis, geospatial SDOH (from diverse government sources [47]), and patient-reported tools (eg, Timeline Follow-Back [83]). Publicly available datasets included Medical Information Mart for Intensive Care III (MIMIC-III) [31,32,55,57,58,74,77,78,82,85,86,95,97,104,118,125,128,130], n2c2 2018/2022 [55,58,80,85,128], SemEval-2015 [128], LGBTQ+ Minority Stress on Social Media (MiSSoM+) dataset [92], and others. MIMIC-III and its derivatives (eg, social and behavioral determinants of health-MIMIC) were frequently used for training [85,86,95], validation [82], or annotation purposes [32].

Figure 3 illustrates the most commonly studied SDOH factors across the years. We show factors that appeared in at least 15 papers. Table 2 shows the SDOH type (individual vs structural) studied in our cohort of papers. The most studied SDOH factors were housing instability (50.7%) and employment (45.8%), appearing in nearly half of the 142 papers, followed by financial context (44.4%), substance use (37.3%), and social isolation (34.5%). Education and living circumstances each appeared in 27.5% and 28.9% of studies, respectively. At the lower end, justice system involvement and language literacy each appeared in approximately 8% of studies. The least studied factors included immigration status (3.5%) and 4 factors that appeared in less than 3 papers each: access to lethal means, acculturation, digital divide, and military sexual trauma. Collectively, these results show that individual SDOH (substance use, social connection/isolation, etc) received more research attention than structural SDOH (transportation, insurance, etc) over the years. The uneven focus on certain SDOH could be due to data and methodological constraints rather than importance alone. Individual-level factors such as housing and employment are more explicitly documented and easier to annotate, whereas structural determinants (eg, immigration status or the digital divide) are often inconsistently recorded or absent from clinical text.

Figure 4A presents a heatmap illustrating the temporal distribution and frequency of the dominant NLP methods identified in this review. Among the included studies, the majority (n=83) quantified model performance using standard metrics, including F₁-score, precision, and recall. Figure 4B delineates the distribution of F₁-scores across the 5 categories, highlighting significant performance variations between rule-based, traditional machine learning, and deep learning approaches. Complementing this, Figure 5 provides a bubble chart depicting the mean precision and recall values for each model category. Collectively, the visualized performance trends show the general capabilities of different modeling approaches for SDOH tasks, though performance may vary based on task complexity, dataset characteristics, and annotation quality.

Overall, the results highlight that transformer-based encoder-only models such as BERT and generative models (encoder-decoders) trained on specialized clinical tasks currently represent the state of the art in terms of SDOH classification and extraction performance. LLMs (19.7% of reviewed studies, including encoder-decoders) and their capabilities represent an emerging area of research. The limited adoption of LLMs likely reflects practical and ethical concerns rather than a lack of capability. Challenges related to hallucinations, limited reproducibility, fairness and bias risks, and uncertainty around data leakage and governance (especially in clinical settings) may constrain their broader use despite strong performance in controlled evaluations. The majority of studies included in this review (≈90%) use a cross-sectional design, focusing on extracting or classifying SDOH information from a single point in time, providing a valuable snapshot. While useful, such approaches cannot capture the changes in SDOH factors among study subjects. Only a handful of studies [66,75,83] adopt a longitudinal approach, following individuals’ records (like EHRs) over time to track changes in SDOH and their impact on health outcomes.

Model interpretation is essential for understanding not just how well a model performs, but also the basis for its decisions. A total of 82 studies reported on using interpretability techniques to gain insights into model performance. These included qualitative content analysis [29,59,131], Shapley additive explanations [47,71,163], Local Interpretable Model-Agnostic Explanations [23], attention visualization or neuron activation analysis [54], and ablation studies [57,87]. One common approach found across studies was error analysis, which plays a key role in understanding how and why models make specific mistakes. It involves examining misclassifications or incorrect outputs to identify patterns of failure. Several studies in our review performed manual (domain expert-based) error analysis to identify model weaknesses. For instance, one paper found that their model struggled to distinguish between general and specific substance-related terms by incorrectly extracting the generic verb “smokes” as the key information when the actual target was the more specific noun “cigarettes” [77]. Another study showed that soft prompting (using trainable continuous vectors instead of manually crafted text instructions) improved the extraction of overlapping or nested SDOH concepts, especially when prompt length was tuned [58]. Error analyses were particularly prominent in studies using LLMs such as ChatGPT, Llama, and Gemini [26,31,57,59,61,155-157,159] because these models (being black-box) can produce outputs that appear coherent and convincing even when they contain mistakes. In addition, confusion matrices were used to visualize class-level misclassification patterns and highlight systematic errors [62].

Analysis of the 20 influential studies published between 2018 and 2025 reveals a methodological landscape dominated by advanced NLP architectures, including BERT [60,78,80,118,130] and LLMs [57], alongside some rule-based [22,127] and hybrid approaches [27,109,168]. Notably, three-quarters (15/20) of these studies relied on manual annotation or clinician input to establish ground truth, highlighting a continued reliance on human expertise to ensure data quality and study robustness. The prevalence of multi-institutional collaborations in these studies (n=11) suggests that diverse teams and broader data sources are key drivers of research impact, likely facilitating increased generalizability and resource access. Regarding performance, the majority of studies (13/20) reported strong F₁-scores, though results varied by task complexity. High benchmarks included concept extraction scores up to 0.9118, and relation extraction at 0.8332 [58], with Seq2Seq models achieving approximately 0.889 [80] and other systems reaching 0.86 [78]. Performance heterogeneity was evident in class-specific tasks; for example, one study reported attribute performance (eg, substance use and employment) ranging from 0.81 to 0.93 [118], while another reported a wider spread from 0.491 for non-SDOH factors to 0.774 for the best-performing class (occupation) using the same model [130]. Beyond standard metrics, one study involving EHR data reported that NLP-based methodology captured SDOH in 80.03% of cases compared with 38.17% using structured fields only [89]. Surprisingly, only 9 out of these 20 studies provided publicly accessible code. Although this represents a higher proportion than the overall set (49/142, 34.5% papers shared their code), public code availability remains limited even among the most influential studies.

The most frequent publication venue was the Journal of the American Medical Informatics Association (12/142, 8.5%), followed by Journal of the American Medical Informatics Association Open (4.9%), Journal of Medical Internet Research (4.2%), and Journal of Biomedical Informatics (4.2%). Conference contributions were most often from the IEEE and AMIA Annual Symposium Proceedings. A majority of studies (110/142, 77.46%) acknowledged funding. The National Institutes of Health and its affiliated institutes (National Library of Medicine, National Center for Advancing Translational Sciences, National Cancer Institute, National Institute on Aging, and NIDA) were the leading funders (64/110, 58.2% of funded studies). Other key funders included the National Science Foundation (n=10), Agency for Healthcare Research and Quality (n=6), and international agencies such as the National Institute for Health Research and Canadian Institutes of Health Research. There is an increase in the number of studies reporting federal funding for SDOH-related NLP research in recent years. The percentage of funded studies more than doubled from 33% to 67% during 2018‐2020, to 78%‐83% during 2021‐2025.

This systematic review synthesizes evidence from 142 studies, addressing 2 primary objectives: to characterize NLP techniques, including LLMs, used to analyze SDOH in text-based data, and to assess their reported effectiveness, gaps, and future research needs. Overall, the review reveals rapid methodological advancement and improvement in NLP performance in recent years, marked by a shift toward transformer-based architectures and emerging use of LLMs, with most applications focused on extracting or classifying individual-level SDOH from clinical and social text. Despite technical progress, the review shows that challenges related to data availability, performance reproducibility, model interpretability, and translation into real-world clinical or public health practice remain. Key findings are elaborated in the following paragraphs.

Across studies, model performance in identifying individual SDOH from unstructured text has generally improved over time, reflecting advances in representation learning and modern NLP systems’ ability to capture abstract and context-dependent social concepts [171]. These advances are likely driven by both increased recognition from the research community of the role of SDOH in shaping health outcomes [172] and by the maturation of transformer-based models that perform well on complex linguistic tasks [173]. Strong reported performance, however, does not consistently translate into reproducible findings or application-ready systems [174]. Many studies relied on private, institution-specific EHR datasets, which limit independent replication and cross-site validation [175]. Although manual and clinician-led annotation was common across the reviewed studies, detailed documentation of annotation guidelines and procedures was often lacking, making it difficult to assess consistency in how SDOH concepts were defined and applied across studies. Limited transparency around data and annotation practices also has important equity implications [176]. Reliance on proprietary EHR data favors well-resourced institutions and restricts broader participation in SDOH-focused NLP research [177].

This review also sheds light on the gap between technical development and practical implementation, and limitations on model interpretability. While many studies report promising model performance, few detail pathways for integrating NLP-derived SDOH insights into clinical workflows or public health interventions, which may be due to systemic, technical, and regulatory barriers [178]. As a result, the potential of these methods to inform equity-focused decision-making remains largely unrealized, a finding supported by prior research [176]. NLP/LLM model interpretability has received more research attention in recent years [179]. A little over half of all reviewed studies used some variant of interpretability or explainability methods, but their use was inconsistent and often inadequately discussed. This gap is relatively more critical for SDOH applications, where language may reflect complex social, cultural, and structural nuances, with uninterpretable errors disproportionately affecting marginalized groups [180]. Limited interpretability also undermines trust in systems’ ability to inform care or policy, both within and outside the sphere of SDOH [181].

Based on the review, we outline several priorities for advancing NLP and LLM applications in the study of SDOH. Most of the existing work used cross-sectional designs, offering only static snapshots of social context. Future research should adopt longitudinal designs capable of capturing the dynamic nature of social risk factors, their accumulation over time, and their interaction with health trajectories [182]. Such longitudinal approaches can include continuous monitoring as well as repeated snapshot or point-in-time measures collected at predetermined intervals [183], for tracking meaningful changes in SDOH. Future work may combine unstructured narratives with structured EHR variables, imaging, and community-level data to produce more holistic and actionable models [184,185]. Effective multimodal approaches must account for heterogeneity in SDOH-containing data [186] to maximize information capture and analytical utility. Advancing the field will also require stronger commitments to transparency and reproducibility. Establishing open benchmarks, shared tasks, and community challenges has driven NLP method development on many topics [80,128], and should be the focus of future work to promote transparency and reproducibility in SDOH-related NLP. Greater attention is also needed for structural determinants of health. Fewer than one-fifth of reviewed studies focused on structural SDOH factors such as policy, racism, or socioeconomic stratification [187]. Capturing structural determinants presents unique methodological challenges, such as patient-reported data on systemic factors being limited by awareness and subjectivity [188]. Linking individual-level health narratives with community- and policy-level data may offer a more comprehensive approach to understanding how structural forces shape health outcomes in future research [189]. Bias and ethical considerations also remain underexplored. Few studies systematically examined algorithmic bias, despite well-documented risks of inequities in automated decision-making [190,191]. Future systems must embed safeguards for fairness, inclusivity, and accountability to avoid reinforcing disparities. Recent work has proposed actionable frameworks and techniques for mitigating bias in artificial intelligence systems [192-194]. Adopting such approaches will be essential for ensuring SDOH-focused models promote rather than perpetuate health inequities. In addition, while clinical concepts are supported by structured hierarchies such as the Unified Medical Language System [195], no equivalent framework exists for social (or nonclinical) determinants. Establishing such frameworks would reduce inconsistencies, enable interoperability, and support the creation of benchmark datasets. Finally, improving model generalizability remains critical, as models trained on single-institution EHRs often underperform in external settings, reflecting the lexical, demographic, and contextual variability of clinical narratives [174,196]. Future work should therefore move beyond single-site training by explicitly benchmarking models in cross-institution evaluations [197]. Training on pooled multi-site data and testing on held-out institutions, with performance reported by SDOH category and population subgroup, may provide actionable evidence of real-world generalizability.

This review has several limitations. Determining whether studies were directly relevant to SDOH required subjective judgment, as many did not explicitly frame their research in SDOH terms. This introduces potential selection bias, despite a structured review process. To balance comprehensiveness and feasibility, we adhered to predefined selection criteria, focusing primarily on studies with a clear connection to SDOH. Although this approach helped maintain relevance, it may have led to the exclusion of studies with indirect but meaningful contributions to the field. Future work could explore more systematic methods, such as automated screening tools or expert consensus frameworks, to enhance the consistency and reproducibility of the selection process. We also acknowledge that our synthesis is constrained by heterogeneity across study designs, SDOH definitions, annotation schemes, datasets, and evaluation metrics, which limits direct comparability of reported performance across studies.

In this systematic review, we provide a comprehensive synthesis of NLP and LLM applications for SDOH by systematically surveying methodologies, data sources, evaluation practices, and translational readiness across 142 studies. In contrast to prior reviews that covered subdomains of SDOH, clinical contexts, or individual modeling techniques, ours provides a comparison of tasks, architectures, performance patterns, interpretability practices, and reproducibility considerations across the research spectrum, spanning both individual and structural SDOH. By discussing NLP/LLM technical performance in the broader context of data accessibility, bias, generalizability, and implementation, this review advances the field beyond proof-of-concept model development toward a clearer understanding of what is required for real-world deployment. The findings highlight that, while NLP-based SDOH extraction is technically mature for certain use cases, its impact remains limited by fragmented SDOH representations, a paucity of public benchmarks, and insufficient evaluation in clinical and public health workflows. In light of our findings, future work should prioritize public benchmarks, clearer reporting, and more inclusive datasets to advance this emerging field. Addressing these gaps has direct real-world implications by facilitating inclusive data practices, interpretable and reproducible methods, and implementation-focused evaluations.