Searching Smarter, Not Harder: Leveraging AI to Enhance Literature Searches for Theory-Driven reviews. A methodological case study

Searching Smarter, Not Harder: Leveraging AI to Enhance Literature Searches for Theory-Driven reviews. A methodological case study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Searching Smarter, Not Harder: Leveraging AI to Enhance Literature Searches for Theory-Driven reviews. A methodological case study Rebecca Hunter, Andrew Booth, Lianne Wood This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7924327/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 04 Mar, 2026 Read the published version in BMC Medical Research Methodology → Version 1 posted 9 You are reading this latest preprint version Abstract Background: Integrating artificial intelligence (AI) into literature searching has the potential to enhance research synthesis by improving the identification of conceptually rich or otherwise difficult-to-locate evidence. Theoretical or conceptual literature reviews, including realist reviews, often involve resource-intensive searches because they aim to trace nuanced ideas, mechanisms, or conceptual relationships across multiple sources. This case study illustrates the use of AI-powered tools to support and streamline such literature searching, using a realist review as an example. Methods: We applied AI tools—Scite and Undermind—in the context of a realist review to facilitate the identification of relevant studies. Seed papers and key informant papers guided the search, and a novel classification system (grandparent, parent, and child papers) was used to systematically organise studies for developing and refining theoretical constructs. Transparent screening procedures and decision-making frameworks were employed to ensure methodological rigour and reproducibility. Results: The integration of AI tools improved the feasibility of identifying conceptually relevant literature and helped manage complex datasets. The classification system enabled structured organisation of studies, supporting iterative testing and refinement of theoretical constructs. The workflow demonstrated flexibility and adaptability, suggesting potential applicability beyond realist review Conclusions: Our findings suggest that AI-powered tools can support and streamline literature searching, improving efficiency in identifying conceptually relevant studies. However, these tools do not replace the critical interpretive work required by researchers. Human judgement remains essential to assess relevance, evaluate nuanced concepts, and make informed decisions throughout the search process, with AI serving as a valuable adjunct rather than a substitute. Literature searches review methodology realist reviews evidence appraisal literature screening artificial intelligence Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Background The advancement of artificial intelligence (AI) capabilities presents an opportunity to enhance research processes, particularly in literature searching[1]. Capturing nuanced concepts, identifying theoretically relevant papers, or tracking related studies can be time- and resource-intensive using conventional methods. AI-powered tools offer opportunities to streamline these processes while maintaining conceptual depth and rigour. This is particularly relevant for literature reviews with a theoretical or explanatory focus, such as realist reviews. These reviews aim to understand complex interventions by examining how context interacts with underlying mechanisms to produce outcomes[2]. To do this, realist reviews often require broad, inclusive searches to identify conceptually rich material to develop and test programme theories[3]. To ensure the process is both thorough and credible, the literature search in a realist review must be systematic and transparent. While a realist review is not an exhaustive literature search, it is intended to be conducted in a way that ensures transparency and rigour in identifying relevant sources. Well-established guidance exists to support researchers to conduct these searches well[4–7]. However, the rapid development of innovative literature search tools—particularly those powered by artificial intelligence (AI)—has driven an emerging need for contemporary methodological guidance to help researchers in effectively utilising these tools. This case study aims to address this need by exploring how AI-driven tools can enhance the literature search process within realist reviews ensuring that research methodologies evolve alongside technological advancements. As a specific type of interpretive mixed methods review of complex interventions, realist reviews share a need for the identification of sufficiently numerous and rich reports to build a complete understanding of both context and of the theory underpinning the development of the intervention[3]. Multiple reports of the same intervention provide crucial contextual details that may be absent from a single source[8], allowing reviewers to develop robust programme theories explaining which mechanisms operate in different contexts to produce outcomes[9]. This thoroughness in identifying related reports is especially critical in realist syntheses, where understanding the interplay between context, mechanisms, and outcomes forms the methodological foundation[10]. Without this comprehensive approach, reviewers risk developing incomplete or misleading theoretical explanations that fail to capture the complexity of interventions as they operate in real-world settings. Researchers have long used techniques such as snowballing, a process of searching the references and/or citations of articles to identify additional relevant material[11]. Databases like Scopus and Web of Science facilitate this process by allowing researchers to track references and citations efficiently. After many years of variations in practice, published guidance now exists to steer the application and reporting of such methods[12]. AI-powered tools expedite and automate this process, making it faster and more efficient. In our own work, we found AI tools particularly useful for the second stage of “a mushroom search”—a two-stage search strategy where a broad, general search (the "cap") is first conducted to gather a wide range of studies, followed by targeted, focused searches (the "stalk") to narrow down and identify studies relevant to key concepts[4]. In our case, we manually entered key concepts drawn from our Context-Mechanism-Outcome configurations (CMOCs) and used AI to conduct a search based on these concepts. However, current AI applications can also automate stages through the search process by generating search strategies (AI-powered search engines) and directly identifying relevant studies (AI reference finders), removing the need for manual query entry and individual review of results. This case study describes the innovative approach the authors employed in conducting a realist review using AI-powered literature search and review tools. Although developed in the context of a realist review, the underlying process—systematically extending searches to surface evidence that conventional strategies may overlook—has broader applicability. Any type of literature review that seeks to identify material not easily captured through standard indexing or keyword-based searching could benefit from this approach. By clearly outlining the steps taken to retrieve and select relevant studies, the authors seek to enhance the transparency and replicability, making it easier for others to follow. In doing so, the case study provides a detailed, replicable account of how these advanced tools were integrated into the realist review process. The goal is not only to demonstrate how these innovative methods can be used effectively but also to encourage other researchers to adopt and apply these approaches in their own realist reviews, contributing to the ongoing evolution of the field. Methods Realist review is a flexible approach to evaluating the literature, allowing for variation in how the process is undertaken[13]. While authors may exceed or fall short of the number of steps as outlined by Pawson, his framework consists of six key stages: identifying the review question, searching for empirical evidence, quality appraisal, data extraction, data synthesis, and dissemination of findings[2]. This study focuses specifically on steps two and three—searching for empirical evidence and conducting quality appraisal—within the context of a realist review evaluating prehabilitation programmes for individuals awaiting spinal surgery for neurogenic claudication (NC)[14]. We describe the use of AI-powered tools, Scite and Undermind (among others available), to identify primary studies for testing CMOCs developed as part of a programme theory. The following sections outline the search process and provide a transparent and rigorous approach to screening results. We present this process, not as a prescriptive method or a model of "best practice," but rather as a resource for other researchers to adapt to their own studies. Searching for empirical evidence Our search focused on identifying empirical studies to test, refine, and potentially refute the CMOCs developed within an initial programme theory on prehabilitation for NC surgical candidates. This objective shaped our search strategy and influenced the selection of tools used. Guided by an experienced information specialist with expertise in realist methodology (AB), the research team decided to use Scite and Undermind. These AI-powered tools were chosen for their potential to expedite the retrieval of relevant studies by identifying research that conceptually aligns with the CMOCs under investigation (see Box 1 for further details on Scite and Undermind). Importantly, the methods we describe are not dependent on these particular tools; the principles outlined here are broadly applicable and can be operationalised using other similar technologies. Search strategy Scite's claim-analysis capabilities, which classify citations as ‘supporting,’ ‘contrasting,’ or ‘mentioning’ the claims of the cited work[15], were used to explore the core concepts within each CMOC. Targeted search queries, combining 'prehabilitation' with relevant keywords, were constructed (see supplementary file 1 for an example). This approach enabled the research team to analyse how concepts were used and referenced within the literature—specifically, the 'citation context'—allowing the research team to identify supporting or contrasting evidence and assess the credibility and application of these concepts within the prehabilitation literature. To maintain precision in our search, we privileged retrieval of the most relevant studies rather than expanding the search with numerous synonyms or alternative terms. For example, to test CMOC1, the search strategy used was: Prehabilitation "waiting list" OR Prehabilitation "waiting period". We chose these terms deliberately, as we considered it unlikely that a relevant full-text document would fail to reference either of these terms if it were addressing prehabilitation in waiting lists or periods. Although additional synonyms could have broadened our results, this focused strategy helped ensure that the studies retrieved were directly relevant to our research question. Such an approach contrasts with systematic reviews where a review team works painstakingly through a list of terms of increasing futility. Undermind generated detailed reports on specific research topics aligned to each CMOC. For our search related to CMOC 1, we began by copying and pasting the CMOC into the initial query box of Undermind, which then guided us through a series of answer-and-response prompts to clarify the focus of our search. Through this iterative process, Undermind helped clarify the key concepts of interest, such as the type of surgery and the desired outcomes. After refining these details, Undermind generated a final search query for CMOC 1: ‘ The impact of prehabilitation on patient empowerment and stress reduction in individuals awaiting orthopaedic surgeries’ . Based on this, Undermind produced a comprehensive report for CMOC 1, summarising relevant literature related to key concepts such as patient empowerment, stress reduction, multimodal programmes, and orthopaedic prehabilitation. The report also included curated citations, helping to streamline the identification of contextually relevant studies. (see supplementary file 2 for an example of a report). Scite: Claim verification and citation analysis • Claim Verification : Assesses scientific claims by analysing how they have been supported, contradicted, or mentioned across the literature. • Citation Analysis : Tracks how individual papers have been cited, offering insight into their influence and relevance. • Research Tool Integration : Compatible with Google Scholar, PubMed, and Web of Science. • Output : Provides contextual citation insights to enhance the interpretation of evidence. Undermind : Advanced literature searching for complex or interdisciplinary topics • Iterative Search Process : Refines and expands search strategies in real time to uncover both central and peripheral literature. • Semantic Understanding : Uses natural language processing to grasp the meaning behind research questions and extract conceptually relevant studies. • Personalised Recommendations : Adapts search results to match a researcher's specific interests and goals. • Output : Yields a broader and more nuanced evidence base tailored to the review’s evolving needs. Box 1: Overview of Scite and Undermind - purpose and capabilities[16] Initial screening Having retrieved a set of potentially relevant papers from both Scite and Undermind, the next step involved systematically screening these results to assess their relevance to the CMOCs. To reduce the volume of papers and make the dataset manageable, preliminary screening of the literature results was conducted by a single researcher (RH). Although this stage was not formally quality assessed, it was guided by broad criteria for assessing relevance, richness, and rigour (RRR)[17]. The Undermind search engine assigned a percentage score to each paper, indicating how closely it matched the research query. Papers scoring below 50% were excluded from further consideration. Among those scoring above 50%, Undermind highlighted ‘top scoring’ papers it considered key. The research team reviewed these to verify alignment with the CMOC and then applied the RRR framework to determine which papers should proceed to full-text review. This initial screening process evokes what Bates described as ‘berrypicking’—a model of real-world information seeking where queries evolve over time and researchers gather relevant material iteratively[18]. Just as a forager scans multiple bushes, choosing only the most promising berries, this stage involved scanning AI-ranked results and selectively “picking” studies based on both algorithmic cues and the RRR framework. ‘Berrypicking’ aligns well with the interpretive, non-linear nature of realist reviews, where studies are selected not through exhaustive retrieval but through theoretically informed, context-sensitive filtering[5]. The research team concluded that the risk of exclusion errors at this stage was minimal, as papers not selected were deemed unlikely to affect the overall relevance and rigour of the final set of papers for review. Application Example: Initial screening of results for CMOC 1 In the case of CMOC 1, Scite produced seven results —considered a manageable number—which were taken forward directly without the need for initial screening. In contrast, Undermind presented 23 results, of which 13 scored below the 50% threshold. After a brief review of titles and summaries, these 13 papers were excluded. Of the remaining ten, Undermind identified five papers as key. While these five were prioritised, the researcher reviewed all ten to validate Undermind’s assessment and ensure no relevant studies were overlooked. The RRR framework was then applied to the five key papers resulting in two being selected for full-text relevance screening. These two papers were then combined with the seven identified by Scite, resulting in a total of nine papers to be taken forward to a full text relevance screen (See Fig. 1 for illustration). Seed and key Informant papers To support the iterative nature of the review and maintain a transparent record of potentially relevant evidence, studies were categorised as seed papers or key informant papers . A seed paper refers to a secondary study that was not included in the analysis itself but cited primary studies that may be relevant to one or more CMOCs. The use of this term was informed by the Terminology, Application, and Reporting of Citation Searching TARCiS statement, which defines “seed references” as relevant articles known beforehand and used as a starting point for citation searches[12]. Adopting this label allowed for systematic tracking of secondary studies that served as gateways to further primary research. Key informant papers , on the other hand, were primary studies that were directly relevant to more than one CMOC within the programme theory (See box 2 for a definition of papers). This coding enabled the research team to ensure that studies with the potential to inform multiple CMOCs were flagged for further analysis, ensuring a thorough and connected synthesis across the programme theory. Paper Type Definition/Decision Criteria Seed paper A secondary study not included in the review itself but which cites primary studies that may be relevant to the programme theory’s CMOCs Key informant paper A primary study that is relevant to more than one CMOC Grandparent Paper • Primary or secondary studies identified in the initial screen as being potentially relevant to the CMOC being tested. • If the grandparent paper is a primary study and meets the relevance criteria, it is included in the second formal screening phase for Relevance, Richness, and Rigour (RRR). • If the grandparent paper is a secondary study , it is not included in the review. • If the grandparent paper is a secondary study that references potentially relevant primary studies informing the CMOC being tested or other CMOCs, it is colour-coded as a seed paper and tagged with hashtags for each CMOC it may inform. • If the grandparent paper is a primary study that does not directly inform the CMOC being tested but is relevant to other CMOCs in the programme theory, it is also categorised as a key informant and similarly tagged. Parent paper • A primary study identified from a grandparent paper that is considered potentially relevant for the CMOC being tested or for other CMOCs within the programme theory. • If a parent study is relevant to more than one CMOC, it is coded as a key informant paper and assigned hashtags for each CMOC it is considered potentially relevant to Child paper • A primary study identified from an included parent paper that is considered potentially relevant for the current context-mechanism-outcome configuration (CMOC) being tested. Box 2: Definition of paper types Relevance Screen After the initial screen, papers identified from the search results of Scite and Undermind as potentially relevant were categorised as ‘grandparent papers’ and subjected to a full-text relevance screen. This process involved assessing whether the studies directly addressed key concepts and variables within the CMOC being tested, as well as evaluating their potential to inform other CMOCs. As outlined in the criteria below (see Box 3), studies were classified based on their relevance to the CMOC being tested, with additional considerations for those that could inform multiple CMOCs. Include • Primary studies (qualitative or quantitative) that are directly relevant to the CMOC being tested, meaning the study explicitly addresses concepts or variables within the CMOC. • Primary studies that are relevant to the CMOC being tested but also identified as relevant for other CMOCs. Studies with dual relevance will be included in the current CMOC analysis and coded as a "key informant" for future re-evaluation with other CMOCs. • Primary studies not relevant to the current CMOC, but relevant to other CMOCs: These studies, will be coded as key informants for other CMOCs and retained for future analysis. Exclude • Primary studies that are not related to the CMOC being tested or any other CMOC within the programme theory. • Secondary studies that are not relevant to the CMOC being tested or any other CMOC within the programme theory. Inform • Seed papers : Secondary studies (e.g., reviews, meta-analyses) that cite primary studies potentially relevant to the CMOC being tested or other CMOCs within the programme theory. These papers will be reviewed to identify relevant primary studies. • Key informant papers : Primary studies that are not directly relevant to the CMOC being tested but may be relevant to other CMOCs within the programme theory. These papers will be flagged for future re-evaluation. If a study becomes relevant to a CMOC already tested, it will be re-evaluated for inclusion. Box 3: Relevance Screening and Categorisation Criteria for CMOC Analysis: Include, Exclude, and Inform Grandparent papers Since our literature search was specifically designed for searching for empirical evidence for testing CMOCs[2], only primary studies were considered for inclusion at this stage. Relevant grandparent studies (studies retrieved from the AI systems) that were primary studies advanced to RRR screening. Grandparent studies classified as secondary studies, retrieved similarly via the AI systems, were not includable in the review in their own right but were mined for any potentially relevant primary studies. If relevant primary studies were identified, the grandparent study was colour coded to indicate its status as a ‘seed paper’. This process continued until all relevant grandparent studies for the CMOC under testing had been screened. Figure 2 illustrates the decision tree to guide the researcher in determining the appropriate course of action regarding grandparent papers. Figure 2: Decision tree for evaluating Grandparent paper Application example: Grandparent paper relevance screen CMOC 1 Using CMOC 1 as an example, nine papers were identified from the initial screen as potentially relevant to CMOC 1 (see Fig. 3). These were read in full and screened for relevance both for CMOC 1 and for any other CMOCs in the programme theory. Six of the nine papers were considered not relevant and were excluded (see Box 3 for exclusion criteria). One paper was a secondary study which referenced potentially relevant primary studies, and thus it was coded as a seed paper. Two papers were primary studies that were not relevant to the current CMOC being tested but were considered relevant to other CMOCs in the programme theory, and so they were coded as key informant papers. Figure 3: Overview of relevance screen for CMOC 1 Parent papers Relevant primary studies identified from the grandparent papers were then, in turn, termed parent papers . Each parent paper was read in full to assess its relevance to the CMOC being tested (See Box 3). If a parent paper was relevant to the current CMOC, it was advanced to the RRR screening phase. In some cases, a parent paper was also relevant to other CMOCs within the broader programme theory; in these instances, it was both progressed to RRR screening and marked as a key informant paper for future analysis. If a parent paper was not relevant to the current CMOC but held potential relevance for other CMOCs, it was designated solely as a key informant paper , to be revisited during the testing of those CMOCs. Finally, if a parent paper was found to be irrelevant to both the current CMOC and all other CMOCs, it was excluded from further analysis. Figure 4 provides a visual illustration of the decision tree used to guide the assessment of parent papers, outlining the pathways for inclusion, exclusion, or designation as key informant papers based on their relevance to the CMOC under review and/or the broader programme theory. Application example: Parent paper relevance screen CMOC 1 From the relevance screen of the nine grandparent studies identified for CMOC 1, one was categorised as a seed paper—a secondary study that cited potentially relevant primary studies. From this seed paper, six primary studies were identified and termed ‘parent papers’. These parent papers were read in full to assess their relevance using the same criteria applied during the grandparent screening stage. Of the six, four were excluded as not relevant to the current CMOC or any other CMOCs. The remaining two were found to be relevant both to CMOC 1 and to other CMOCs within the broader programme theory. These two parent papers were therefore included for RRR screening and simultaneously coded as key informant papers to be revisited in future CMOC analyses (see Fig. 5 ). When a primary study is identified as relevant to more than one CMOC, it may seem redundant to code it as a key informant paper. It could be argued that, for expediency, the paper should go straight through to the RRR screening phase, tagged with hashtags for each relevant CMOC—bypassing the "key informant" categorisation. However, revisiting the paper when focusing on each new CMOC offers an opportunity to reassess its relevance in light of the current focus and the evolving data being analysed. Re-reading at this point allows researchers to refine their understanding and verify whether the study’s initial relevance remains valid or whether new insights have emerged that might affect its inclusion within another CMOC. The act of revisiting a study is analogous to re-interviewing a participant in qualitative research after having spoken to additional participants. Insights gained from the other interviews may influence how you interpret the second conversation with the same person, allowing for a deeper understanding or a different perspective based on the new information[19]. Similarly, when revisiting a study in the context of a new CMOC, the interpretation of its relevance may evolve as it is considered alongside the findings from other studies. While this process may seem time-consuming, it tends to become increasingly efficient with each iteration. Since the study has already been reviewed, subsequent readings are generally lighter-touch providing an opportunity to add nuanced insights and annotations. Revisiting a study when focusing on a new CMOC not only provides an opportunity to refine interpretations but also helps sensitise researchers to the emergence of demi-regularities—context-dependent patterns that reveal how mechanisms operate across different settings[20]. However, while searching for these demi-regularities, it is important to remain cautious of confirmation bias —the tendency to focus on evidence that supports an emerging idea while overlooking information that might contradict or complicate it[21]. This can lead to a reinforcing feedback loop where each item of supporting evidence strengthens belief in the prevalent idea and heightens sensitivity to similar patterns. This mirrors the ‘streetlamp effect’: searching only where the light is shining, rather than where the most meaningful evidence may lie[22]. Realist researchers are actively engaged in looking for rivalry and counterfactuals, which are fundamental for theory development[23]. By deliberately seeking alternative explanations and examining how different CMOCs might play out, researchers can refine their theories and ensure that they aren’t simply confirming pre-existing assumptions. While the challenge of confirmation bias cannot be fully eliminated, being aware of the broader aim of revisiting studies—testing, challenging, and refining emerging theories—helps mitigate this risk. This approach ensures that the review remains rigorous, balanced, and open to new insights, ultimately leading to a robust understanding of the relationships between context, mechanisms, and outcomes. Child papers In some cases, primary studies referenced within included parent papers—referred to as child studies (second stage referrals)—may be potentially relevant to the CMOC under testing. The screening process for child papers follows the same procedure as parent papers. Decisions on whether to screen and include child studies can be made by the research team, typically guided by the extent of available data and the need for further evidence to test the CMOC in question. In our review, where the literature on nutrition in prehabilitation programmes for NC surgical candidates was relatively sparse, references from a few child studies were included. However, given time constraints and limited resources, a pragmatic decision was made not to extend the review to the child level for the remaining CMOCs, as the prevalent literature base was deemed sufficient. Authors thus face an ongoing decision on how many iterations (generations) are required, both in general, and for particular CMOCs. In doing so they need to decide whether to prioritise standardisation across CMOCs (following up the same number of levels for each CMOC) or to favour customisation to the specific needs of each CMOC (with different levels for each). Quality assessment To minimise the risk of excluding relevant studies at the parent study stage, we implemented a quality assurance check (see Fig. 6 and the box highlighted as ‘relatively high risk’). Since decisions about which papers to advance to RRR screening were based on judgments of relevance, ensuring a consistent and rigorous approach was important. Given that we were trialling this method, we adopted a cautious approach by having a second screener review 50% of the records excluded by the first reviewer for each CMOC. While this percentage may seem high, it proved feasible, as a total of 18 papers across all CMOCs were excluded at this stage, meaning the second screener only needed to assess nine papers. This additional check helped strengthen the reliability of our process. During this process, the second screener noted the importance of keeping a copy of all CMOCs readily available whilst screening to ensure that papers were assessed not only for relevance to the CMOC being tested but also for potential relevance to other CMOCs within the broader framework. Figure 6 uses CMOC 1 as an example to illustrate the flow of studies included in the review, starting with the grandparent studies retrieved through the AI-powered searches using Scite, Undermind, and key informant papers. The diagram traces the path to the parent studies, highlighting key points at which studies were excluded from the review. Two boxes with dashed outlines indicate these exclusion points. The second exclusion box labelled ‘relatively high risk,’ marks the stage where a second reviewer was introduced for quality appraisal, as described earlier in the methodology. Such an approach harmonises with recent approaches to quality assessment, particularly for rapid review methods, where previous arbitrary models of double checking are increasingly being replaced with strategies specifically targeting stages considered of higher risk[24]. Similar flow diagrams were created for all 14 CMOCs included in the programme theory, ensuring consistency and transparency throughout the study selection process. Relevance, Richness and Rigour appraisal The RRR screening process served as a transparency step rather than a strict eligibility filter. By this stage, all papers had already undergone an initial relevance screen, ensuring they were relevant to one or more CMOCs. As a result, there were no exclusions, and the RRR screen focused on assessing the confidence with which conclusions could be drawn from each paper. Each paper was evaluated against three independent criteria: relevance, richness, and rigour, with scores of high, moderate, or low assigned to each (see supplementary file 3). While papers could theoretically be excluded if they scored low across all three criteria, in practice, the RRR screen primarily aimed to clarify the strength of the evidence and highlight any limitations. For example, papers with high relevance but lower richness or rigour were still included, but their limitations were explicitly acknowledged to ensure transparent and nuanced interpretation of the findings. This approach aligns with the guidance provided by Dada et al[17], which emphasises transparency in the evaluation of evidence to ensure clarity regarding the confidence in findings and conclusions drawn from studies. Discussion This case study illustrates an approach for integrating AI-powered tools into literature searching to uncover evidence that may be missed by conventional methods. By integrating emerging technologies such as Scite and Undermind into our workflow, we enhanced the efficiency, transparency, and conceptual focus of our review. Our structured approach to screening and appraisal—drawing on seed papers, key informant papers, and a grandparent–parent–child classification system— provided a clear and replicable framework for navigating complex literature and identifying papers relevant to testing and refining theory. Although this case study was developed in the context of a realist review, the approach has broader relevance. By systematically extending searches to detect evidence that conventional strategies may overlook, the method can support any type of literature review where conceptually rich or otherwise hard-to-find evidence is important. Conventional keyword searches often rely on the researcher's vocabulary aligning with that used in titles, abstracts, or indexing terms. This limitation can hinder the discovery of relevant literature that uses alternative language or conceptual framing. In contrast, AI-powered tools like Undermind and Scite use natural language processing (NLP) to assess full-text similarities and conceptual relationships. This allowed us to retrieve literature aligned with the theoretical underpinnings of our CMOCs—even when key terms were absent—thereby improving conceptual retrieval beyond the capabilities of traditional search strategies. Importantly, these tools were deployed within tightly defined conceptual boundaries (i.e., specific CMOCs), rather than broader population–intervention–comparison–outcome (PICO) queries, making them particularly suited to realist inquiry. AI-powered tools enhanced our ability to identify relevant papers for testing our CMOCs. Conventional keyword searches rely on specific terms and phrases inputted by the researcher matching either those used by the author in their title or abstract or those used by an indexer to enhance retrieval. Not only does this limit the search to the researcher's knowledge and their associated vocabulary but it also misses greater retrieval opportunities offered by the richness of full text retrieval. As a result, these searches may miss conceptually related literature that uses different terminology or phrasing. In contrast, tools such as Undermind and Scite employ natural language processing techniques to identify papers that align conceptually with the search query, even when the exact terms or synonyms are not present. By drawing on a broader knowledge base, these tools offer a nuanced approach that extends beyond keyword searches. This capability, broadening retrieval but only within very tight search parameters (the language associated with a CMOC, not a broader PICO review question) was particularly valuable for our realist review, as we sought literature to test concepts, identify mechanisms, and understand important contextual factors. While AI tools like Scite and Undermind improved the efficiency and conceptual precision of our screening process, they did not replace the interpretive work required in a realist review. Their ability to identify semantically relevant papers and provide summarised insights helped surface data that may have been overlooked using traditional keyword searches. However, the task of making sense of these findings —through retroductive and abductive reasoning[25] remained firmly with the research team. This interpretive work was essential in testing, refining, and refuting our CMOCs. Our findings therefore support the view increasingly reflected in the literature: AI is not a substitute for human judgement but a valuable adjunct to it. Studies such as Ge et al.[26] and Gwon et al.[27] highlight similar dynamics in other evidence synthesis contexts, where AI accelerates routine tasks like screening but requires human oversight for rigour and contextual interpretation. Feng et al.[28] further emphasise this by demonstrating that AI tools often achieve high recall (i.e., successfully identifying most relevant papers), but can vary in precision—reinforcing the need for critical human appraisal. As such, the integration of AI tools can enhance the scope and sensitivity of realist reviews, but only when grounded in thoughtful, theory-driven human interpretation. Challenges and Limitations While this manuscript is based on a single case study, using a detailed, single-case application to introduce and demonstrate a new methodological approach is a common practice. For example, Booth et al.[5] used a single-case review to develop and illustrate their CLUSTER methodology for identifying contextually and theoretically relevant studies. Nonetheless, there are several challenges and limitations in this case study that are worth noting, and the approach requires further testing and evaluation in other contexts and with other research teams to assess its broader applicability and robustness. AI-powered tools operate within the constraints of their underlying algorithms, and the lack of transparency in the coding of these algorithms may introduce biases. For instance, the relevance scoring could reflect limitations or preferences embedded in the training data or algorithmic design potentially leading to certain types of evidence being emphasised over others. Additionally, these tools may have limited access to certain databases or publication types, such as grey literature or very recent publications, which could result in gaps in the literature covered. Human reviewers are encouraged to be reflexive about their influence on research processes, AI systems are not – although some researchers have begun experimenting by scripting AI systems to identify their own limitations! Furthermore, as these tools are relatively new, limited evidence exists for their performance against librarian-led search strategies, especially in the context of realist reviews. Realist reviews prioritise identifying conceptually relevant literature over exhaustive searches[5], so the effectiveness of AI-powered tools should be evaluated based on their ability to capture studies that meaningfully contribute to programme theory development, using criteria such as relevance, richness, and rigour[17]. While AI-powered tools like Scite and Undermind are useful for identifying studies, their ability to pinpoint in an almost forensic way literature that directly informs the development and testing of CMOCs requires further evaluation. Without direct comparison to conventional search strategies, we cannot conclusively determine whether AI tools are more or less effective in identifying the studies most aligned with the goals of realist review. Despite the promise of AI-powered tools, they still require a degree of methodological expertise to use effectively. Researchers with a solid understanding of realist methodology—or working within a team with such experience—may be better equipped to formulate appropriate search queries and critically assess the conceptual relevance of retrieved studies. AI-powered tools are a useful adjunct to traditional search methods, helping to streamline the process. However, we would argue that they cannot replace the nuanced judgement and interpretation essential to realist review. Specifically, the time overhead required to oversee and validate their otherwise credible-looking output should not be overlooked. While AI-powered tools undoubtedly save time in searching and assembling the literature, time spent categorising papers may offset these gains. The multiple screening stages—such as categorising grandparent, parent, and child papers—demand significant researcher time and effort. However, this challenge arose primarily because this approach was novel and developed iteratively; with clearer instructions in place, it is likely to be quicker and more efficient in future applications. Our classification system and decision trees will require ongoing refinement and validation to ensure their continued credibility across different contexts. A limitation regarding accessibility stems from our decision to use AI tools which require paid subscriptions for full functionality. However, the use of specialised, commercial software is a standard practice in modern academic research (e.g., licensed statistical or reference management packages). Furthermore, many features of Scite and Undermind could be used in the free or 'lite' versions of these applications, thereby mitigating the access barrier for resource-constrained researchers. Finally, a key limitation of this review was the absence of quality assurance checks during the initial and grandparent screening stages. However, at the parent screening stage—where the risk of excluding rich and relevant studies was considered greater—we introduced a quality check, with a second reviewer screening 50% of excluded studies. While arguably risk averse, this step reflects a commitment to maintaining rigour in the screening process. Other research teams may wish to consider introducing similar checks earlier, depending on the stage at which key decisions are made and the level of confidence in the tools used. Conclusion This case study offers a methodological resource for researchers seeking to integrate and make the most of AI-powered literature search tools in support of realist reviews. Searching for nuanced concepts to build explanatory theories is fundamental to realist synthesis, but it often involves navigating large volumes of literature in pursuit of elusive 'mechanisms' or key contextual factors. By demonstrating the application of tools such as Scite and Undermind, this case study shows how AI can support targeted, conceptually driven searches—enabling researchers to access rich, relevant data with greater efficiency. However, as with any tool, their value lies in how they are used. This is especially true for AI-powered searches, which are still relatively new. We are continuing to learn how best to use these tools effectively, and their role is to complement—not replace—the judgment and interpretive thinking of the researcher. The methods outlined in this case study offer a practical framework for selecting, organising, and appraising search results in a transparent and robust way, contributing to the ongoing development of realist review methodology. While not intended as a definitive blueprint, this case study offers a foundation for others to build upon, test, and refine as methodological practices continue to evolve. Abbreviations AI Artificial Intelligence CMOC Context-Mechanism-Outcome configuration RRR Relevance, Richness, Rigour Declarations Ethical approval Not applicable Consent for publication Not applicable Availability of data and materials The authors confirm that key materials supporting the findings of this study are provided within the article and its supplementary materials. Additional flow diagrams and related data are available from the authors upon reasonable request Competing interests The authors declare that they have no competing interests Funding This methodological case study is derived from a realist review conducted as part of a larger programme of work funded by the National Institute for Health and Care Research (NIHR) Programme Development Grant Scheme [NIHR205671] and supported by the NIHR Exeter Biomedical Research Centre (BRC). The views expressed are those of the authors and do not necessarily reflect those of the NIHR or the Department of Health and Social Care. The funders had no role in the study’s design, data collection, analysis, interpretation, or manuscript drafting. The corresponding authors had unrestricted access to all data and the final decision to submit the manuscript for publication. Author contributions LW conceptualised the original realist review that informed this study. RH led the literature review from which the current methodology was developed, with RH and AB jointly designing the methodological approach. RH also created the system for organising selection and appraisal flow diagrams in consultation with AB. AB designed the literature search, and RH interpreted the results in collaboration with AB and LW. RH and AB co-wrote the initial draft of the manuscript, which was subsequently reviewed and edited by AB and LW. All authors read and approved the final manuscript. Acknowledgements For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising from this submission. References Roth S, Wermer-Colan A. Machine learning methods for systematic reviews: a rapid scoping review. Delaware Journal of Public Health. 2023;9(4):40. Pawson R. Realist synthesis: new protocols for systematic review. Evidence-based policy. 2006:73-104. Wong G. Data gathering in realist reviews: looking for needles in haystacks. Doing realist research London: SAGE. 2018:131-45. Booth A, Briscoe S, Wright JM. The “realist search”: A systematic scoping review of current practice and reporting. Research Synthesis Methods. 2019;11(1):14-35. Booth A, Harris J, Croot E, Springett J, Campbell F, Wilkins E. Towards a methodology for cluster searching to provide conceptual and contextual "richness" for systematic reviews of complex interventions: case study (CLUSTER). BMC medical research methodology. 2013;13:118-. Booth A, Wright J, Briscoe S. Scoping and searching to support realist approaches. Doing realist research. 2018:147-66. Duddy C, Roberts N. Identifying evidence for five realist reviews in primary health care: A comparison of search methods. Research Synthesis Methods. 2022;13(2):190-203. Noyes J, Hendry M, Lewin S, Glenton C, Chandler J, Rashidian A. Qualitative “trial-sibling” studies and “unrelated” qualitative studies contributed to complex intervention reviews. Journal of Clinical Epidemiology. 2016;74:133-43. Wong G, Westhorp G, Pawson R, Greenhalgh T. Realist synthesis training materials. 2013. Greenhalgh T, Wong G, Westhorp G, Pawson R. Protocol-realist and meta-narrative evidence synthesis: evolving standards (RAMESES). BMC medical research methodology. 2011;11:1-10. Wohlin C, editor Guidelines for snowballing in systematic literature studies and a replication in software engineering. Proceedings of the 18th international conference on evaluation and assessment in software engineering; 2014. Hirt J, Nordhausen T, Fuerst T, Ewald H, Appenzeller-Herzog C. Guidance on terminology, application, and reporting of citation searching: the TARCiS statement. bmj. 2024;385. Hunter R, Gorely T, Beattie M, Harris K. Realist review. International Review of Sport and Exercise Psychology. 2022;15(1):242-65. Hunter R, Booth A, Lamb SE, Williamson E, Hendrick P, Sahota O, et al. Prehabilitation for lumbar spinal stenosis: understanding mechanisms and contexts for enhanced engagement-a realist review. Age and Ageing. In press . Brody S. Scite. Journal of the Medical Library Association: JMLA. 2021;109(4):707. Labcritics. Undermind.ai Review: AI-Driven Revolution in the Literature Review: Labcritics, Lab Tech Review & Marketplace; 2024 [Available from: https://labcritics.com/undermind-ai-review-ai-driven-revolution-in-the-literature-review/. Dada S, Dalkin S, Gilmore B, Hunter R, Mukumbang FC. Applying and reporting relevance, richness and rigour in realist evidence appraisals: advancing key concepts in realist reviews. Research synthesis methods. 2023;14(3):504-14. Bates MJ. The design of browsing and berrypicking techniques for the online search interface. Online review. 1989;13(5):407-24. Goyes DR, Sandberg S. Trust, nuance, and care: Advantages and challenges of repeat qualitative interviews. Qualitative Research. 2024:14687941241246159. Jagosh J, Macaulay AC, Pluye P, Salsberg J, Bush PL, Henderson J, et al. Uncovering the benefits of participatory research: implications of a realist review for health research and practice. Milbank Q. 2012;90(2):311-46. Nickerson RS. Confirmation bias: A ubiquitous phenomenon in many guises. Review of general psychology. 1998;2(2):175-220. Kaplan A. The conduct of inquiry: Methodology for behavioural science: Routledge; 2017. Jagosh J. Realist synthesis for public health: building an ontologically deep understanding of how programs work, for whom, and in which contexts. Annual review of public health. 2019;40(1):361-72. Booth A, Sommer I, Noyes J, Houghton C, Campbell F. Rapid reviews methods series: guidance on rapid qualitative evidence synthesis. BMJ evidence-based medicine. 2024;29(3):194-200. Jagosh J. Retroductive theorizing in Pawson and Tilley's applied scientific realism. Journal of Critical Realism. 2020;19(2):121-30. Ge L, Agrawal R, Singer M, Kannapiran P, De Castro Molina JA, Teow KL, et al. Leveraging artificial intelligence to enhance systematic reviews in health research: advanced tools and challenges. Syst Rev. 2024;13(1):269. Gwon YN, Kim JH, Chung HS, Jung EJ, Chun J, Lee S, Shim SR. The use of generative AI for scientific literature searches for systematic reviews: ChatGPT and Microsoft Bing AI performance evaluation. JMIR Medical Informatics. 2024;12:e51187. Feng Y, Liang S, Zhang Y, Chen S, Wang Q, Huang T, et al. Automated medical literature screening using artificial intelligence: a systematic review and meta-analysis. Journal of the American Medical Informatics Association. 2022;29(8):1425-32. Additional Declarations No competing interests reported. Supplementary Files Supplementaryfile1SciteliteratureresultsforCMOC3.docx Supplementaryfile2UndermindliteratureresultsforCMOC1.docx Supplementaryfile3RelevanceRichnessandRigourscreeningtool.docx Cite Share Download PDF Status: Published Journal Publication published 04 Mar, 2026 Read the published version in BMC Medical Research Methodology → Version 1 posted Editorial decision: Revision requested 03 Dec, 2025 Reviews received at journal 18 Nov, 2025 Reviews received at journal 14 Nov, 2025 Reviewers agreed at journal 07 Nov, 2025 Reviewers agreed at journal 06 Nov, 2025 Reviewers invited by journal 06 Nov, 2025 Editor assigned by journal 29 Oct, 2025 Submission checks completed at journal 29 Oct, 2025 First submitted to journal 22 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7924327","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":545725341,"identity":"c191d0de-197f-4e99-b482-3b84bf7267ea","order_by":0,"name":"Rebecca Hunter","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABRElEQVRIie2RP0vDQBTA3xGwy9muVyLtV7gQSJWWfBGXlIO4NIqjWPAkkC6xcz6Gk9ot4SAuia6FOgiuDhVBAoKatvafjeDocL/h7t2994P3eAASyb8Ezy7Ev4PK7AVAp+dKSbFiAVQnddFflFmQKzRcKBslUxq9i+gx65qg8O1YPc5aR/rIjV9fbsR+QxXnY8hTNAlXlZ3kjml+zPImyrYaWPbe1UPMgigRzqDfdgnkKZryVYWQjqEirkxmMVRsCWoMOzpEnnAuE8QJbIVA79caI/XnxjviZ3Plk+rBUnEz+NhUCDYUxMVcCSklS8UjyAs3GsOHrOrHtxi5WG9im1EytBmk3oEz8JG32+4zXF0fn5TSaJx1T2taz9dGuGXSSsAEnHhN5xqXnobjN7NWTiwoAGtu0fdkUT8XuaD+W0IikUgkX/2PdPiGR6bmAAAAAElFTkSuQmCC","orcid":"","institution":"University of Exeter","correspondingAuthor":true,"prefix":"","firstName":"Rebecca","middleName":"","lastName":"Hunter","suffix":""},{"id":545725342,"identity":"51cab767-514c-405a-a92c-a24b87474e9f","order_by":1,"name":"Andrew Booth","email":"","orcid":"","institution":"University of Sheffield","correspondingAuthor":false,"prefix":"","firstName":"Andrew","middleName":"","lastName":"Booth","suffix":""},{"id":545725343,"identity":"6870dcf4-8b65-4709-b2f0-391863ca1300","order_by":2,"name":"Lianne Wood","email":"","orcid":"","institution":"University of Exeter","correspondingAuthor":false,"prefix":"","firstName":"Lianne","middleName":"","lastName":"Wood","suffix":""}],"badges":[],"createdAt":"2025-10-22 14:08:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7924327/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7924327/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12874-026-02814-3","type":"published","date":"2026-03-04T15:57:57+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":96251890,"identity":"a2bd7087-fd64-4ac5-b809-19ea6b69c851","added_by":"auto","created_at":"2025-11-19 07:40:09","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1140175,"visible":true,"origin":"","legend":"","description":"","filename":"SearchingSmarterNotHarderV7FINAL.docx","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/106f50bf73114df3c5afdc60.docx"},{"id":96159380,"identity":"3727146a-3cc4-4231-b167-17da3af8b799","added_by":"auto","created_at":"2025-11-18 08:41:43","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6170,"visible":true,"origin":"","legend":"","description":"","filename":"6a588841cfe84f08ad799190eebfac00.json","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/222fac035c023d8e75ae5628.json"},{"id":96159389,"identity":"65d206ff-3ce3-42f1-b3e2-26e45712aea6","added_by":"auto","created_at":"2025-11-18 08:41:43","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28681,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryfile1SciteliteratureresultsforCMOC3.docx","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/d25317e69ab39ece579e1442.docx"},{"id":96159386,"identity":"c34abd80-049f-40cb-a172-8ec4f91130ec","added_by":"auto","created_at":"2025-11-18 08:41:43","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":34633,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryfile2UndermindliteratureresultsforCMOC1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/d494f5584fa689075597eb9b.docx"},{"id":96252140,"identity":"914365f2-1086-4ec6-aad9-93fd467db946","added_by":"auto","created_at":"2025-11-19 07:40:30","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17787,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryfile3RelevanceRichnessandRigourscreeningtool.docx","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/d66b6235e648fd455a5ce6c0.docx"},{"id":96159400,"identity":"2618c50a-ba43-42f1-80fa-3e2e42645fb9","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"xml","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":72182,"visible":true,"origin":"","legend":"","description":"","filename":"6a588841cfe84f08ad799190eebfac001enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/80abdbd8cce6dbb6098eb7d6.xml"},{"id":96250003,"identity":"96f598de-8648-429b-aa61-511bbf52e635","added_by":"auto","created_at":"2025-11-19 07:37:03","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":115394,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/0fbd81a27ebe972b14710412.png"},{"id":96249275,"identity":"6ba34fa0-054d-49c9-8642-d880835a8c27","added_by":"auto","created_at":"2025-11-19 07:32:50","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":183465,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/f9a358c5b6dfb4a17df12f54.png"},{"id":96159404,"identity":"86e816a1-3517-4bbb-a6c1-fc755ee14eeb","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":142695,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/5f4ba75be9a37c425e7d0858.png"},{"id":96251069,"identity":"fa859dd2-bad8-4e45-8fe1-5ba972c08a1b","added_by":"auto","created_at":"2025-11-19 07:39:17","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":120813,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/0b689af61a89ca12ce6eab9d.png"},{"id":96159405,"identity":"552ec9c6-f2fc-4206-a946-7ec31c3c8dd4","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":313233,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/74245092af328c02a50ca23f.png"},{"id":96159403,"identity":"2f67e759-4dfd-49fb-b288-0987105edcd3","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":184522,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/a4d405edf76a0e6670f3f0a0.png"},{"id":96251078,"identity":"46a5ecf8-354e-4525-bcb1-d0e6b9be9b52","added_by":"auto","created_at":"2025-11-19 07:39:17","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":38627,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/e35bcf61ab4cc46e80beea26.png"},{"id":96159392,"identity":"040caa55-4ced-4a30-ad41-3a586088a4c8","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":60579,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/ca47222bab57f197a08c88db.png"},{"id":96159401,"identity":"f6377cb7-6ae7-4e67-b757-2a85c229133a","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":49214,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/b79ac117ef75ed87df8bfe11.png"},{"id":96159393,"identity":"38763024-90f3-4af4-bfc3-b9049ed73cd1","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":41905,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/64f54b95b1ef1f5a1b2ae27c.png"},{"id":96251084,"identity":"0c9cfcbc-b9a7-4280-a1db-52578a2d0aaf","added_by":"auto","created_at":"2025-11-19 07:39:17","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":111397,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/8813621e15ca9cd917f62898.png"},{"id":96159406,"identity":"f9eb35a3-127e-4286-b760-0c7ce0e71a13","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61314,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/a880089edc7761ef6d2c8242.png"},{"id":96249148,"identity":"a2ebd565-2a0c-480d-abda-d185644b867d","added_by":"auto","created_at":"2025-11-19 07:30:26","extension":"xml","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":70324,"visible":true,"origin":"","legend":"","description":"","filename":"6a588841cfe84f08ad799190eebfac001structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/31b16402d1b1d0719bcfbd32.xml"},{"id":96159408,"identity":"62166739-2422-4ea9-a523-4c78261e194a","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"html","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":78971,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/905c4b3279ed932e78eef9a0.html"},{"id":96159379,"identity":"460775ce-b139-4247-b636-3b483a0bfca4","added_by":"auto","created_at":"2025-11-18 08:41:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":115394,"visible":true,"origin":"","legend":"\u003cp\u003eAn initial screen of search results for CMOC 1\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/cf39a7b8218f5989c4c36cb2.png"},{"id":96251150,"identity":"0ee699e7-3567-4439-951e-47399af2575f","added_by":"auto","created_at":"2025-11-19 07:39:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":183465,"visible":true,"origin":"","legend":"\u003cp\u003eDecision tree for evaluating Grandparent paper\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/53ed7b4fceac6bdc960bbd02.png"},{"id":96159382,"identity":"5dc4ea2a-094d-4c01-8b5f-c0e5bff1a56e","added_by":"auto","created_at":"2025-11-18 08:41:43","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":142695,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of relevance screen for CMOC 1\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/c2e6779e947f20acb578fab5.png"},{"id":96249181,"identity":"2166ab4f-d270-449b-b126-05acc17e0666","added_by":"auto","created_at":"2025-11-19 07:30:36","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":120813,"visible":true,"origin":"","legend":"\u003cp\u003eDecision tree for screening and categorising parent papers\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/a9c76f0e65448d704d0196e4.png"},{"id":96159399,"identity":"65be4b65-3f1e-4367-8853-795979140f28","added_by":"auto","created_at":"2025-11-18 08:41:44","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":313233,"visible":true,"origin":"","legend":"\u003cp\u003eParent relevance screen for CMOC 1\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/03d522d714cb39e9e8801e72.png"},{"id":96159391,"identity":"285ea6e3-297b-4ed8-adb7-d7a4f9ec0350","added_by":"auto","created_at":"2025-11-18 08:41:43","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":184522,"visible":true,"origin":"","legend":"\u003cp\u003eStudy selection flow for CMOC \u003cem\u003e1\u003c/em\u003e with exclusion points highlighted\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/8b5969e46fd2acb7fe69b5f3.png"},{"id":104252019,"identity":"8bcddf4f-d542-4834-9f6e-1f96d1690449","added_by":"auto","created_at":"2026-03-09 16:16:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2288801,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/dcc002a8-8894-4fcd-901d-9a0a375e9ca0.pdf"},{"id":96159383,"identity":"fc09776c-def1-4359-8c4b-96cf2d9acf21","added_by":"auto","created_at":"2025-11-18 08:41:43","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":28681,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryfile1SciteliteratureresultsforCMOC3.docx","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/9dc9e379363cf37348ca490c.docx"},{"id":96249211,"identity":"e6d3277e-d34d-4d68-ad62-070df5e78001","added_by":"auto","created_at":"2025-11-19 07:30:41","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":34633,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryfile2UndermindliteratureresultsforCMOC1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/dc1f5c56025ca9509a2aff66.docx"},{"id":96251949,"identity":"5576917a-a5c4-48cf-9d14-51082dc2ae9a","added_by":"auto","created_at":"2025-11-19 07:40:13","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":17787,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementaryfile3RelevanceRichnessandRigourscreeningtool.docx","url":"https://assets-eu.researchsquare.com/files/rs-7924327/v1/3254e300f5ec5f8f971a97b4.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eSearching Smarter, Not Harder: Leveraging AI to Enhance Literature Searches for Theory-Driven reviews. A methodological case study\u003c/p\u003e","fulltext":[{"header":"Background","content":"\u003cp\u003eThe advancement of artificial intelligence (AI) capabilities presents an opportunity to enhance research processes, particularly in literature searching[1]. Capturing nuanced concepts, identifying theoretically relevant papers, or tracking related studies can be time- and resource-intensive using conventional methods. AI-powered tools offer opportunities to streamline these processes while maintaining conceptual depth and rigour.\u003c/p\u003e\u003cp\u003eThis is particularly relevant for literature reviews with a theoretical or explanatory focus, such as realist reviews. These reviews aim to understand complex interventions by examining how context interacts with underlying mechanisms to produce outcomes[2]. To do this, realist reviews often require broad, inclusive searches to identify conceptually rich material to develop and test programme theories[3].\u003c/p\u003e\u003cp\u003eTo ensure the process is both thorough and credible, the literature search in a realist review must be systematic and transparent. While a realist review is not an exhaustive literature search, it is intended to be conducted in a way that ensures transparency and rigour in identifying relevant sources. Well-established guidance exists to support researchers to conduct these searches well[4\u0026ndash;7]. However, the rapid development of innovative literature search tools\u0026mdash;particularly those powered by artificial intelligence (AI)\u0026mdash;has driven an emerging need for contemporary methodological guidance to help researchers in effectively utilising these tools. This case study aims to address this need by exploring how AI-driven tools can enhance the literature search process within realist reviews ensuring that research methodologies evolve alongside technological advancements.\u003c/p\u003e\u003cp\u003eAs a specific type of interpretive mixed methods review of complex interventions, realist reviews share a need for the identification of sufficiently numerous and rich reports to build a complete understanding of both context and of the theory underpinning the development of the intervention[3]. Multiple reports of the same intervention provide crucial contextual details that may be absent from a single source[8], allowing reviewers to develop robust programme theories explaining which mechanisms operate in different contexts to produce outcomes[9]. This thoroughness in identifying related reports is especially critical in realist syntheses, where understanding the interplay between context, mechanisms, and outcomes forms the methodological foundation[10]. Without this comprehensive approach, reviewers risk developing incomplete or misleading theoretical explanations that fail to capture the complexity of interventions as they operate in real-world settings.\u003c/p\u003e\u003cp\u003eResearchers have long used techniques such as snowballing, a process of searching the references and/or citations of articles to identify additional relevant material[11]. Databases like Scopus and Web of Science facilitate this process by allowing researchers to track references and citations efficiently. After many years of variations in practice, published guidance now exists to steer the application and reporting of such methods[12]. AI-powered tools expedite and automate this process, making it faster and more efficient. In our own work, we found AI tools particularly useful for the second stage of \u0026ldquo;a mushroom search\u0026rdquo;\u0026mdash;a two-stage search strategy where a broad, general search (the \"cap\") is first conducted to gather a wide range of studies, followed by targeted, focused searches (the \"stalk\") to narrow down and identify studies relevant to key concepts[4]. In our case, we manually entered key concepts drawn from our Context-Mechanism-Outcome configurations (CMOCs) and used AI to conduct a search based on these concepts. However, current AI applications can also automate stages through the search process by generating search strategies (AI-powered search engines) and directly identifying relevant studies (AI reference finders), removing the need for manual query entry and individual review of results.\u003c/p\u003e\u003cp\u003eThis case study describes the innovative approach the authors employed in conducting a realist review using AI-powered literature search and review tools. Although developed in the context of a realist review, the underlying process\u0026mdash;systematically extending searches to surface evidence that conventional strategies may overlook\u0026mdash;has broader applicability. Any type of literature review that seeks to identify material not easily captured through standard indexing or keyword-based searching could benefit from this approach. By clearly outlining the steps taken to retrieve and select relevant studies, the authors seek to enhance the transparency and replicability, making it easier for others to follow. In doing so, the case study provides a detailed, replicable account of how these advanced tools were integrated into the realist review process. The goal is not only to demonstrate how these innovative methods can be used effectively but also to encourage other researchers to adopt and apply these approaches in their own realist reviews, contributing to the ongoing evolution of the field.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eRealist review is a flexible approach to evaluating the literature, allowing for variation in how the process is undertaken[13]. While authors may exceed or fall short of the number of steps as outlined by Pawson, his framework consists of six key stages: identifying the review question, searching for empirical evidence, quality appraisal, data extraction, data synthesis, and dissemination of findings[2]. This study focuses specifically on steps two and three\u0026mdash;searching for empirical evidence and conducting quality appraisal\u0026mdash;within the context of a realist review evaluating prehabilitation programmes for individuals awaiting spinal surgery for neurogenic claudication (NC)[14]. We describe the use of AI-powered tools, Scite and Undermind (among others available), to identify primary studies for testing CMOCs developed as part of a programme theory. The following sections outline the search process and provide a transparent and rigorous approach to screening results. We present this process, not as a prescriptive method or a model of \"best practice,\" but rather as a resource for other researchers to adapt to their own studies.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eSearching for empirical evidence\u003c/h2\u003e\u003cp\u003eOur search focused on identifying empirical studies to test, refine, and potentially refute the CMOCs developed within an initial programme theory on prehabilitation for NC surgical candidates. This objective shaped our search strategy and influenced the selection of tools used. Guided by an experienced information specialist with expertise in realist methodology (AB), the research team decided to use Scite and Undermind. These AI-powered tools were chosen for their potential to expedite the retrieval of relevant studies by identifying research that conceptually aligns with the CMOCs under investigation (see Box 1 for further details on Scite and Undermind). Importantly, the methods we describe are not dependent on these particular tools; the principles outlined here are broadly applicable and can be operationalised using other similar technologies.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eSearch strategy\u003c/h3\u003e\n\u003cp\u003eScite's claim-analysis capabilities, which classify citations as \u0026lsquo;supporting,\u0026rsquo; \u0026lsquo;contrasting,\u0026rsquo; or \u0026lsquo;mentioning\u0026rsquo; the claims of the cited work[15], were used to explore the core concepts within each CMOC. Targeted search queries, combining 'prehabilitation' with relevant keywords, were constructed (see supplementary file 1 for an example). This approach enabled the research team to analyse how concepts were used and referenced within the literature\u0026mdash;specifically, the 'citation context'\u0026mdash;allowing the research team to identify supporting or contrasting evidence and assess the credibility and application of these concepts within the prehabilitation literature.\u003c/p\u003e\u003cp\u003eTo maintain precision in our search, we privileged retrieval of the most relevant studies rather than expanding the search with numerous synonyms or alternative terms. For example, to test CMOC1, the search strategy used was: Prehabilitation \"waiting list\" OR Prehabilitation \"waiting period\". We chose these terms deliberately, as we considered it unlikely that a relevant full-text document would fail to reference either of these terms if it were addressing prehabilitation in waiting lists or periods. Although additional synonyms could have broadened our results, this focused strategy helped ensure that the studies retrieved were directly relevant to our research question. Such an approach contrasts with systematic reviews where a review team works painstakingly through a list of terms of increasing futility.\u003c/p\u003e\u003cp\u003eUndermind generated detailed reports on specific research topics aligned to each CMOC. For our search related to CMOC 1, we began by copying and pasting the CMOC into the initial query box of Undermind, which then guided us through a series of answer-and-response prompts to clarify the focus of our search. Through this iterative process, Undermind helped clarify the key concepts of interest, such as the type of surgery and the desired outcomes. After refining these details, Undermind generated a final search query for CMOC 1: \u0026lsquo;\u003cem\u003eThe impact of prehabilitation on patient empowerment and stress reduction in individuals awaiting orthopaedic surgeries\u0026rsquo;\u003c/em\u003e. Based on this, Undermind produced a comprehensive report for CMOC 1, summarising relevant literature related to key concepts such as patient empowerment, stress reduction, multimodal programmes, and orthopaedic prehabilitation. The report also included curated citations, helping to streamline the identification of contextually relevant studies. (see supplementary file 2 for an example of a report).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e\u003ccolgroup cols=\"1\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eScite: Claim verification and citation analysis\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026bull; \u003cb\u003eClaim Verification\u003c/b\u003e: Assesses scientific claims by analysing how they have been supported, contradicted, or mentioned across the literature.\u003c/p\u003e\u003cp\u003e\u0026bull; \u003cb\u003eCitation Analysis\u003c/b\u003e: Tracks how individual papers have been cited, offering insight into their influence and relevance.\u003c/p\u003e\u003cp\u003e\u0026bull; \u003cb\u003eResearch Tool Integration\u003c/b\u003e: Compatible with Google Scholar, PubMed, and Web of Science.\u003c/p\u003e\u003cp\u003e\u0026bull; \u003cb\u003eOutput\u003c/b\u003e: Provides \u003cem\u003econtextual citation insights\u003c/em\u003e to enhance the interpretation of evidence.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eUndermind\u003c/b\u003e: Advanced literature searching for complex or interdisciplinary topics\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u0026bull; \u003cb\u003eIterative Search Process\u003c/b\u003e: Refines and expands search strategies in real time to uncover both central and peripheral literature.\u003c/p\u003e\u003cp\u003e\u0026bull; \u003cb\u003eSemantic Understanding\u003c/b\u003e: Uses natural language processing to grasp the meaning behind research questions and extract conceptually relevant studies.\u003c/p\u003e\u003cp\u003e\u0026bull; \u003cb\u003ePersonalised Recommendations\u003c/b\u003e: Adapts search results to match a researcher's specific interests and goals.\u003c/p\u003e\u003cp\u003e\u0026bull; \u003cb\u003eOutput\u003c/b\u003e: Yields a broader and more nuanced evidence base tailored to the review\u0026rsquo;s evolving needs.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\n\u003ch3\u003eBox 1: Overview of Scite and Undermind - purpose and capabilities[16]\u003c/h3\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003eInitial screening\u003c/h2\u003e\u003cp\u003eHaving retrieved a set of potentially relevant papers from both Scite and Undermind, the next step involved systematically screening these results to assess their relevance to the CMOCs. To reduce the volume of papers and make the dataset manageable, preliminary screening of the literature results was conducted by a single researcher (RH). Although this stage was not formally quality assessed, it was guided by broad criteria for assessing relevance, richness, and rigour (RRR)[17]. The Undermind search engine assigned a percentage score to each paper, indicating how closely it matched the research query. Papers scoring below 50% were excluded from further consideration. Among those scoring above 50%, Undermind highlighted \u0026lsquo;top scoring\u0026rsquo; papers it considered key. The research team reviewed these to verify alignment with the CMOC and then applied the RRR framework to determine which papers should proceed to full-text review.\u003c/p\u003e\u003cp\u003eThis initial screening process evokes what Bates described as \u0026lsquo;berrypicking\u0026rsquo;\u0026mdash;a model of real-world information seeking where queries evolve over time and researchers gather relevant material iteratively[18]. Just as a forager scans multiple bushes, choosing only the most promising berries, this stage involved scanning AI-ranked results and selectively \u0026ldquo;picking\u0026rdquo; studies based on both algorithmic cues and the RRR framework. \u0026lsquo;Berrypicking\u0026rsquo; aligns well with the interpretive, non-linear nature of realist reviews, where studies are selected not through exhaustive retrieval but through theoretically informed, context-sensitive filtering[5]. The research team concluded that the risk of exclusion errors at this stage was minimal, as papers not selected were deemed unlikely to affect the overall relevance and rigour of the final set of papers for review.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eApplication Example: Initial screening of results for CMOC 1\u003c/h3\u003e\n\u003cp\u003eIn the case of CMOC 1, Scite produced seven results \u0026mdash;considered a manageable number\u0026mdash;which were taken forward directly without the need for initial screening. In contrast, Undermind presented 23 results, of which 13 scored below the 50% threshold. After a brief review of titles and summaries, these 13 papers were excluded. Of the remaining ten, Undermind identified five papers as key. While these five were prioritised, the researcher reviewed all ten to validate Undermind\u0026rsquo;s assessment and ensure no relevant studies were overlooked. The RRR framework was then applied to the five key papers resulting in two being selected for full-text relevance screening. These two papers were then combined with the seven identified by Scite, resulting in a total of nine papers to be taken forward to a full text relevance screen (See Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e for illustration).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eSeed and key Informant papers\u003c/h2\u003e\u003cp\u003eTo support the iterative nature of the review and maintain a transparent record of potentially relevant evidence, studies were categorised as \u003cem\u003eseed papers\u003c/em\u003e or \u003cem\u003ekey informant papers\u003c/em\u003e. A \u003cem\u003eseed paper\u003c/em\u003e refers to a secondary study that was not included in the analysis itself but cited primary studies that may be relevant to one or more CMOCs. The use of this term was informed by the Terminology, Application, and Reporting of Citation Searching TARCiS statement, which defines \u0026ldquo;seed references\u0026rdquo; as relevant articles known beforehand and used as a starting point for citation searches[12]. Adopting this label allowed for systematic tracking of secondary studies that served as gateways to further primary research. \u003cem\u003eKey informant papers\u003c/em\u003e, on the other hand, were primary studies that were directly relevant to more than one CMOC within the programme theory (See box 2 for a definition of papers). This coding enabled the research team to ensure that studies with the potential to inform multiple CMOCs were flagged for further analysis, ensuring a thorough and connected synthesis across the programme theory.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePaper Type\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDefinition/Decision Criteria\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSeed paper\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eA \u003cb\u003esecondary study\u003c/b\u003e not included in the review itself but which cites primary studies that may be relevant to the programme theory\u0026rsquo;s CMOCs\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eKey informant paper\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eA \u003cb\u003eprimary study\u003c/b\u003e that is relevant to more than one CMOC\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eGrandparent Paper\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026bull; \u003cb\u003ePrimary\u003c/b\u003e or \u003cb\u003esecondary\u003c/b\u003e studies identified in the initial screen as being potentially relevant to the CMOC being tested.\u003c/p\u003e\u003cp\u003e\u0026bull; If the grandparent paper is a \u003cb\u003eprimary study\u003c/b\u003e and meets the relevance criteria, it is included in the second formal screening phase for Relevance, Richness, and Rigour (RRR).\u003c/p\u003e\u003cp\u003e\u0026bull; If the grandparent paper is a \u003cb\u003esecondary study\u003c/b\u003e, it is not included in the review.\u003c/p\u003e\u003cp\u003e\u0026bull; If the grandparent paper is a \u003cb\u003esecondary study\u003c/b\u003e that references potentially relevant \u003cb\u003eprimary studies\u003c/b\u003e informing the CMOC being tested or other CMOCs, it is colour-coded as a \u003cem\u003eseed paper\u003c/em\u003e and tagged with hashtags for each CMOC it may inform.\u003c/p\u003e\u003cp\u003e\u0026bull; If the grandparent paper is a \u003cb\u003eprimary study\u003c/b\u003e that does not directly inform the CMOC being tested but is relevant to other CMOCs in the programme theory, it is also categorised as a \u003cem\u003ekey informant\u003c/em\u003e and similarly tagged.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eParent paper\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026bull; A \u003cb\u003eprimary study\u003c/b\u003e identified from a grandparent paper that is considered potentially relevant for the CMOC being tested or for other CMOCs within the programme theory.\u003c/p\u003e\u003cp\u003e\u0026bull; If a parent study is relevant to more than one CMOC, it is coded as a \u003cb\u003ekey informant paper\u003c/b\u003e and assigned hashtags for each CMOC it is considered potentially relevant to\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eChild paper\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026bull; A primary study identified from an included parent paper that is considered potentially relevant for the current context-mechanism-outcome configuration (CMOC) being tested.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eBox 2: Definition of paper types\u003c/h3\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003eRelevance Screen\u003c/h2\u003e\u003cp\u003eAfter the initial screen, papers identified from the search results of Scite and Undermind as potentially relevant were categorised as \u0026lsquo;grandparent papers\u0026rsquo; and subjected to a full-text relevance screen. This process involved assessing whether the studies directly addressed key concepts and variables within the CMOC being tested, as well as evaluating their potential to inform other CMOCs. As outlined in the criteria below (see Box 3), studies were classified based on their relevance to the CMOC being tested, with additional considerations for those that could inform multiple CMOCs.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eInclude\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026bull; Primary studies (qualitative or quantitative) that are directly relevant to the CMOC being tested, meaning the study explicitly addresses concepts or variables within the CMOC.\u003c/p\u003e\u003cp\u003e\u0026bull; Primary studies that are relevant to the CMOC being tested but also identified as relevant for other CMOCs. Studies with dual relevance will be included in the current CMOC analysis and coded as a \"key informant\" for future re-evaluation with other CMOCs.\u003c/p\u003e\u003cp\u003e\u0026bull; Primary studies not relevant to the current CMOC, but relevant to other CMOCs: These studies, will be coded as key informants for other CMOCs and retained for future analysis.\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eExclude\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026bull; \u003cb\u003ePrimary studies\u003c/b\u003e that are not related to the CMOC being tested or any other CMOC within the programme theory.\u003c/p\u003e\u003cp\u003e\u0026bull; \u003cb\u003eSecondary studies\u003c/b\u003e that are not relevant to the CMOC being tested or any other CMOC within the programme theory.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eInform\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u0026bull; \u003cb\u003eSeed papers\u003c/b\u003e: Secondary studies (e.g., reviews, meta-analyses) that cite primary studies potentially relevant to the CMOC being tested or other CMOCs within the programme theory. These papers will be reviewed to identify relevant primary studies.\u003c/p\u003e\u003cp\u003e\u0026bull; \u003cb\u003eKey informant papers\u003c/b\u003e: Primary studies that are not directly relevant to the CMOC being tested but may be relevant to other CMOCs within the programme theory. These papers will be flagged for future re-evaluation. If a study becomes relevant to a CMOC already tested, it will be re-evaluated for inclusion.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eBox 3: Relevance Screening and Categorisation Criteria for CMOC Analysis: Include, Exclude, and Inform\u003c/h2\u003e\u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\u003ch2\u003eGrandparent papers\u003c/h2\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSince our literature search was specifically designed for searching for empirical evidence for testing CMOCs[2], only primary studies were considered for inclusion at this stage. Relevant grandparent studies (studies retrieved from the AI systems) that were primary studies advanced to RRR screening. Grandparent studies classified as secondary studies, retrieved similarly via the AI systems, were not includable in the review in their own right but were mined for any potentially relevant primary studies. If relevant primary studies were identified, the grandparent study was colour coded to indicate its status as a \u0026lsquo;seed paper\u0026rsquo;. This process continued until all relevant grandparent studies for the CMOC under testing had been screened. Figure\u0026nbsp;2 illustrates the decision tree to guide the researcher in determining the appropriate course of action regarding grandparent papers.\u003c/p\u003e\u003cp\u003eFigure 2: Decision tree for evaluating Grandparent paper\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eApplication example: Grandparent paper relevance screen CMOC 1\u003c/h2\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eUsing CMOC 1 as an example, nine papers were identified from the initial screen as potentially relevant to CMOC 1 (see Fig.\u0026nbsp;3). These were read in full and screened for relevance both for CMOC 1 and for any other CMOCs in the programme theory. Six of the nine papers were considered not relevant and were excluded (see Box 3 for exclusion criteria). One paper was a secondary study which referenced potentially relevant primary studies, and thus it was coded as a seed paper. Two papers were primary studies that were not relevant to the current CMOC being tested but were considered relevant to other CMOCs in the programme theory, and so they were coded as key informant papers.\u003c/p\u003e\u003cp\u003eFigure 3: Overview of relevance screen for CMOC 1\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eParent papers\u003c/h2\u003e\u003cp\u003eRelevant primary studies identified from the grandparent papers were then, in turn, termed \u003cem\u003eparent papers\u003c/em\u003e. Each parent paper was read in full to assess its relevance to the CMOC being tested (See Box 3). If a parent paper was relevant to the current CMOC, it was advanced to the RRR screening phase. In some cases, a parent paper was also relevant to other CMOCs within the broader programme theory; in these instances, it was both progressed to RRR screening and marked as a \u003cem\u003ekey informant paper\u003c/em\u003e for future analysis. If a parent paper was not relevant to the current CMOC but held potential relevance for other CMOCs, it was designated solely as a \u003cem\u003ekey informant paper\u003c/em\u003e, to be revisited during the testing of those CMOCs. Finally, if a parent paper was found to be irrelevant to both the current CMOC and all other CMOCs, it was excluded from further analysis. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003e provides a visual illustration of the decision tree used to guide the assessment of parent papers, outlining the pathways for inclusion, exclusion, or designation as key informant papers based on their relevance to the CMOC under review and/or the broader programme theory.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eApplication example: Parent paper relevance screen CMOC 1\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFrom the relevance screen of the nine grandparent studies identified for CMOC 1, one was categorised as a seed paper\u0026mdash;a secondary study that cited potentially relevant primary studies. From this seed paper, six primary studies were identified and termed \u0026lsquo;parent papers\u0026rsquo;. These parent papers were read in full to assess their relevance using the same criteria applied during the grandparent screening stage. Of the six, four were excluded as not relevant to the current CMOC or any other CMOCs. The remaining two were found to be relevant both to CMOC 1 and to other CMOCs within the broader programme theory. These two parent papers were therefore included for RRR screening and simultaneously coded as key informant papers to be revisited in future CMOC analyses (see Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eWhen a primary study is identified as relevant to more than one CMOC, it may seem redundant to code it as a key informant paper. It could be argued that, for expediency, the paper should go straight through to the RRR screening phase, tagged with hashtags for each relevant CMOC\u0026mdash;bypassing the \"key informant\" categorisation. However, revisiting the paper when focusing on each new CMOC offers an opportunity to reassess its relevance in light of the current focus and the evolving data being analysed. Re-reading at this point allows researchers to refine their understanding and verify whether the study\u0026rsquo;s initial relevance remains valid or whether new insights have emerged that might affect its inclusion within another CMOC.\u003c/p\u003e\u003cp\u003eThe act of revisiting a study is analogous to re-interviewing a participant in qualitative research after having spoken to additional participants. Insights gained from the other interviews may influence how you interpret the second conversation with the same person, allowing for a deeper understanding or a different perspective based on the new information[19]. Similarly, when revisiting a study in the context of a new CMOC, the interpretation of its relevance may evolve as it is considered alongside the findings from other studies. While this process may seem time-consuming, it tends to become increasingly efficient with each iteration. Since the study has already been reviewed, subsequent readings are generally lighter-touch providing an opportunity to add nuanced insights and annotations.\u003c/p\u003e\u003cp\u003eRevisiting a study when focusing on a new CMOC not only provides an opportunity to refine interpretations but also helps sensitise researchers to the emergence of demi-regularities\u0026mdash;context-dependent patterns that reveal how mechanisms operate across different settings[20]. However, while searching for these demi-regularities, it is important to remain cautious of confirmation bias \u0026mdash;the tendency to focus on evidence that supports an emerging idea while overlooking information that might contradict or complicate it[21]. This can lead to a reinforcing feedback loop where each item of supporting evidence strengthens belief in the prevalent idea and heightens sensitivity to similar patterns. This mirrors the \u0026lsquo;streetlamp effect\u0026rsquo;: searching only where the light is shining, rather than where the most meaningful evidence may lie[22]. Realist researchers are actively engaged in looking for rivalry and counterfactuals, which are fundamental for theory development[23]. By deliberately seeking alternative explanations and examining how different CMOCs might play out, researchers can refine their theories and ensure that they aren\u0026rsquo;t simply confirming pre-existing assumptions. While the challenge of confirmation bias cannot be fully eliminated, being aware of the broader aim of revisiting studies\u0026mdash;testing, challenging, and refining emerging theories\u0026mdash;helps mitigate this risk. This approach ensures that the review remains rigorous, balanced, and open to new insights, ultimately leading to a robust understanding of the relationships between context, mechanisms, and outcomes.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eChild papers\u003c/h2\u003e\u003cp\u003eIn some cases, primary studies referenced within included parent papers\u0026mdash;referred to as child studies (second stage referrals)\u0026mdash;may be potentially relevant to the CMOC under testing. The screening process for child papers follows the same procedure as parent papers. Decisions on whether to screen and include child studies can be made by the research team, typically guided by the extent of available data and the need for further evidence to test the CMOC in question. In our review, where the literature on nutrition in prehabilitation programmes for NC surgical candidates was relatively sparse, references from a few child studies were included. However, given time constraints and limited resources, a pragmatic decision was made not to extend the review to the child level for the remaining CMOCs, as the prevalent literature base was deemed sufficient. Authors thus face an ongoing decision on how many iterations (generations) are required, both in general, and for particular CMOCs. In doing so they need to decide whether to prioritise standardisation across CMOCs (following up the same number of levels for each CMOC) or to favour customisation to the specific needs of each CMOC (with different levels for each).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003eQuality assessment\u003c/h2\u003e\u003cp\u003eTo minimise the risk of excluding relevant studies at the parent study stage, we implemented a quality assurance check (see Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e6\u003c/span\u003e and the box highlighted as \u0026lsquo;relatively high risk\u0026rsquo;). Since decisions about which papers to advance to RRR screening were based on judgments of relevance, ensuring a consistent and rigorous approach was important. Given that we were trialling this method, we adopted a cautious approach by having a second screener review 50% of the records excluded by the first reviewer for each CMOC. While this percentage may seem high, it proved feasible, as a total of 18 papers across all CMOCs were excluded at this stage, meaning the second screener only needed to assess nine papers. This additional check helped strengthen the reliability of our process. During this process, the second screener noted the importance of keeping a copy of all CMOCs readily available whilst screening to ensure that papers were assessed not only for relevance to the CMOC being tested but also for potential relevance to other CMOCs within the broader framework.\u003c/p\u003e\u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e6\u003c/span\u003e uses CMOC 1 as an example to illustrate the flow of studies included in the review, starting with the grandparent studies retrieved through the AI-powered searches using Scite, Undermind, and key informant papers. The diagram traces the path to the parent studies, highlighting key points at which studies were excluded from the review. Two boxes with dashed outlines indicate these exclusion points. The second exclusion box labelled \u0026lsquo;relatively high risk,\u0026rsquo; marks the stage where a second reviewer was introduced for quality appraisal, as described earlier in the methodology. Such an approach harmonises with recent approaches to quality assessment, particularly for rapid review methods, where previous arbitrary models of double checking are increasingly being replaced with strategies specifically targeting stages considered of higher risk[24]. Similar flow diagrams were created for all 14 CMOCs included in the programme theory, ensuring consistency and transparency throughout the study selection process.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003eRelevance, Richness and Rigour appraisal\u003c/h2\u003e\u003cp\u003eThe RRR screening process served as a transparency step rather than a strict eligibility filter. By this stage, all papers had already undergone an initial relevance screen, ensuring they were relevant to one or more CMOCs. As a result, there were no exclusions, and the RRR screen focused on assessing the confidence with which conclusions could be drawn from each paper. Each paper was evaluated against three independent criteria: relevance, richness, and rigour, with scores of high, moderate, or low assigned to each (see supplementary file 3). While papers could theoretically be excluded if they scored low across all three criteria, in practice, the RRR screen primarily aimed to clarify the strength of the evidence and highlight any limitations. For example, papers with high relevance but lower richness or rigour were still included, but their limitations were explicitly acknowledged to ensure transparent and nuanced interpretation of the findings. This approach aligns with the guidance provided by Dada et al[17], which emphasises transparency in the evaluation of evidence to ensure clarity regarding the confidence in findings and conclusions drawn from studies.\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis case study illustrates an approach for integrating AI-powered tools into literature searching to uncover evidence that may be missed by conventional methods. By integrating emerging technologies such as Scite and Undermind into our workflow, we enhanced the efficiency, transparency, and conceptual focus of our review. Our structured approach to screening and appraisal\u0026mdash;drawing on seed papers, key informant papers, and a grandparent\u0026ndash;parent\u0026ndash;child classification system\u0026mdash; provided a clear and replicable framework for navigating complex literature and identifying papers relevant to testing and refining theory. Although this case study was developed in the context of a realist review, the approach has broader relevance. By systematically extending searches to detect evidence that conventional strategies may overlook, the method can support any type of literature review where conceptually rich or otherwise hard-to-find evidence is important.\u003c/p\u003e\u003cp\u003eConventional keyword searches often rely on the researcher's vocabulary aligning with that used in titles, abstracts, or indexing terms. This limitation can hinder the discovery of relevant literature that uses alternative language or conceptual framing. In contrast, AI-powered tools like Undermind and Scite use natural language processing (NLP) to assess full-text similarities and conceptual relationships. This allowed us to retrieve literature aligned with the theoretical underpinnings of our CMOCs\u0026mdash;even when key terms were absent\u0026mdash;thereby improving conceptual retrieval beyond the capabilities of traditional search strategies. Importantly, these tools were deployed within tightly defined conceptual boundaries (i.e., specific CMOCs), rather than broader population\u0026ndash;intervention\u0026ndash;comparison\u0026ndash;outcome (PICO) queries, making them particularly suited to realist inquiry.\u003c/p\u003e\u003cp\u003eAI-powered tools enhanced our ability to identify relevant papers for testing our CMOCs. Conventional keyword searches rely on specific terms and phrases inputted by the researcher matching either those used by the author in their title or abstract or those used by an indexer to enhance retrieval. Not only does this limit the search to the researcher's knowledge and their associated vocabulary but it also misses greater retrieval opportunities offered by the richness of full text retrieval. As a result, these searches may miss conceptually related literature that uses different terminology or phrasing. In contrast, tools such as Undermind and Scite employ natural language processing techniques to identify papers that align conceptually with the search query, even when the exact terms or synonyms are not present. By drawing on a broader knowledge base, these tools offer a nuanced approach that extends beyond keyword searches. This capability, broadening retrieval but only within very tight search parameters (the language associated with a CMOC, not a broader PICO review question) was particularly valuable for our realist review, as we sought literature to test concepts, identify mechanisms, and understand important contextual factors.\u003c/p\u003e\u003cp\u003eWhile AI tools like Scite and Undermind improved the efficiency and conceptual precision of our screening process, they did not replace the interpretive work required in a realist review. Their ability to identify semantically relevant papers and provide summarised insights helped surface data that may have been overlooked using traditional keyword searches. However, the task of making sense of these findings \u0026mdash;through retroductive and abductive reasoning[25] remained firmly with the research team. This interpretive work was essential in testing, refining, and refuting our CMOCs. Our findings therefore support the view increasingly reflected in the literature: AI is not a substitute for human judgement but a valuable adjunct to it. Studies such as Ge et al.[26] and Gwon et al.[27] highlight similar dynamics in other evidence synthesis contexts, where AI accelerates routine tasks like screening but requires human oversight for rigour and contextual interpretation. Feng et al.[28] further emphasise this by demonstrating that AI tools often achieve high recall (i.e., successfully identifying most relevant papers), but can vary in precision\u0026mdash;reinforcing the need for critical human appraisal. As such, the integration of AI tools can enhance the scope and sensitivity of realist reviews, but only when grounded in thoughtful, theory-driven human interpretation.\u003c/p\u003e\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\u003ch2\u003eChallenges and Limitations\u003c/h2\u003e\u003cp\u003eWhile this manuscript is based on a single case study, using a detailed, single-case application to introduce and demonstrate a new methodological approach is a common practice. For example, Booth et al.[5] used a single-case review to develop and illustrate their CLUSTER methodology for identifying contextually and theoretically relevant studies. Nonetheless, there are several challenges and limitations in this case study that are worth noting, and the approach requires further testing and evaluation in other contexts and with other research teams to assess its broader applicability and robustness.\u003c/p\u003e\u003cp\u003eAI-powered tools operate within the constraints of their underlying algorithms, and the lack of transparency in the coding of these algorithms may introduce biases. For instance, the relevance scoring could reflect limitations or preferences embedded in the training data or algorithmic design potentially leading to certain types of evidence being emphasised over others. Additionally, these tools may have limited access to certain databases or publication types, such as grey literature or very recent publications, which could result in gaps in the literature covered. Human reviewers are encouraged to be reflexive about their influence on research processes, AI systems are not \u0026ndash; although some researchers have begun experimenting by scripting AI systems to identify their own limitations!\u003c/p\u003e\u003cp\u003eFurthermore, as these tools are relatively new, limited evidence exists for their performance against librarian-led search strategies, especially in the context of realist reviews. Realist reviews prioritise identifying conceptually relevant literature over exhaustive searches[5], so the effectiveness of AI-powered tools should be evaluated based on their ability to capture studies that meaningfully contribute to programme theory development, using criteria such as relevance, richness, and rigour[17]. While AI-powered tools like Scite and Undermind are useful for identifying studies, their ability to pinpoint in an almost forensic way literature that directly informs the development and testing of CMOCs requires further evaluation. Without direct comparison to conventional search strategies, we cannot conclusively determine whether AI tools are more or less effective in identifying the studies most aligned with the goals of realist review.\u003c/p\u003e\u003cp\u003eDespite the promise of AI-powered tools, they still require a degree of methodological expertise to use effectively. Researchers with a solid understanding of realist methodology\u0026mdash;or working within a team with such experience\u0026mdash;may be better equipped to formulate appropriate search queries and critically assess the conceptual relevance of retrieved studies. AI-powered tools are a useful adjunct to traditional search methods, helping to streamline the process. However, we would argue that they cannot replace the nuanced judgement and interpretation essential to realist review. Specifically, the time overhead required to oversee and validate their otherwise credible-looking output should not be overlooked.\u003c/p\u003e\u003cp\u003eWhile AI-powered tools undoubtedly save time in searching and assembling the literature, time spent categorising papers may offset these gains. The multiple screening stages\u0026mdash;such as categorising grandparent, parent, and child papers\u0026mdash;demand significant researcher time and effort. However, this challenge arose primarily because this approach was novel and developed iteratively; with clearer instructions in place, it is likely to be quicker and more efficient in future applications. Our classification system and decision trees will require ongoing refinement and validation to ensure their continued credibility across different contexts.\u003c/p\u003e\u003cp\u003eA limitation regarding accessibility stems from our decision to use AI tools which require paid subscriptions for full functionality. However, the use of specialised, commercial software is a standard practice in modern academic research (e.g., licensed statistical or reference management packages). Furthermore, many features of Scite and Undermind could be used in the free or 'lite' versions of these applications, thereby mitigating the access barrier for resource-constrained researchers.\u003c/p\u003e\u003cp\u003eFinally, a key limitation of this review was the absence of quality assurance checks during the initial and grandparent screening stages. However, at the parent screening stage\u0026mdash;where the risk of excluding rich and relevant studies was considered greater\u0026mdash;we introduced a quality check, with a second reviewer screening 50% of excluded studies. While arguably risk averse, this step reflects a commitment to maintaining rigour in the screening process. Other research teams may wish to consider introducing similar checks earlier, depending on the stage at which key decisions are made and the level of confidence in the tools used.\u003c/p\u003e\u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis case study offers a methodological resource for researchers seeking to integrate and make the most of AI-powered literature search tools in support of realist reviews. Searching for nuanced concepts to build explanatory theories is fundamental to realist synthesis, but it often involves navigating large volumes of literature in pursuit of elusive 'mechanisms' or key contextual factors.\u003c/p\u003e\u003cp\u003eBy demonstrating the application of tools such as Scite and Undermind, this case study shows how AI can support targeted, conceptually driven searches\u0026mdash;enabling researchers to access rich, relevant data with greater efficiency. However, as with any tool, their value lies in how they are used. This is especially true for AI-powered searches, which are still relatively new. We are continuing to learn how best to use these tools effectively, and their role is to complement\u0026mdash;not replace\u0026mdash;the judgment and interpretive thinking of the researcher.\u003c/p\u003e\u003cp\u003eThe methods outlined in this case study offer a practical framework for selecting, organising, and appraising search results in a transparent and robust way, contributing to the ongoing development of realist review methodology. While not intended as a definitive blueprint, this case study offers a foundation for others to build upon, test, and refine as methodological practices continue to evolve.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAI\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Artificial Intelligence\u003c/p\u003e\n\u003cp\u003eCMOC\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;Context-Mechanism-Outcome configuration\u003c/p\u003e\n\u003cp\u003eRRR \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; Relevance, Richness, Rigour\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthical approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors confirm that key materials supporting the findings of this study are provided within the article and its supplementary materials. Additional flow diagrams and related data are available from the authors upon reasonable request\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;The authors declare that they have no competing interests\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis methodological case study is derived from a realist review conducted as part of a larger programme of work funded by the National Institute for Health and Care Research (NIHR) Programme Development Grant Scheme [NIHR205671] and supported by the NIHR Exeter Biomedical Research Centre (BRC). The views expressed are those of the authors and do not necessarily reflect those of the NIHR or the Department of Health and Social Care. The funders had no role in the study\u0026rsquo;s design, data collection, analysis, interpretation, or manuscript drafting. The corresponding authors had unrestricted access to all data and the final decision to submit the manuscript for publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLW conceptualised the original realist review that informed this study. RH led the literature review from which the current methodology was developed, with RH and AB jointly designing the methodological approach. RH also created the system for organising selection and appraisal flow diagrams in consultation with AB. AB designed the literature search, and RH interpreted the results in collaboration with AB and LW. RH and AB co-wrote the initial draft of the manuscript, which was subsequently reviewed and edited by AB and LW. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising from this submission.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eRoth S, Wermer-Colan A. Machine learning methods for systematic reviews: a rapid scoping review. Delaware Journal of Public Health. 2023;9(4):40.\u003c/li\u003e\n\u003cli\u003ePawson R. Realist synthesis: new protocols for systematic review. Evidence-based policy. 2006:73-104.\u003c/li\u003e\n\u003cli\u003eWong G. Data gathering in realist reviews: looking for needles in haystacks. Doing realist research London: SAGE. 2018:131-45.\u003c/li\u003e\n\u003cli\u003eBooth A, Briscoe S, Wright JM. The \u0026ldquo;realist search\u0026rdquo;: A systematic scoping review of current practice and reporting. Research Synthesis Methods. 2019;11(1):14-35.\u003c/li\u003e\n\u003cli\u003eBooth A, Harris J, Croot E, Springett J, Campbell F, Wilkins E. Towards a methodology for cluster searching to provide conceptual and contextual \u0026quot;richness\u0026quot; for systematic reviews of complex interventions: case study (CLUSTER). BMC medical research methodology. 2013;13:118-.\u003c/li\u003e\n\u003cli\u003eBooth A, Wright J, Briscoe S. Scoping and searching to support realist approaches. Doing realist research. 2018:147-66.\u003c/li\u003e\n\u003cli\u003eDuddy C, Roberts N. Identifying evidence for five realist reviews in primary health care: A comparison of search methods. Research Synthesis Methods. 2022;13(2):190-203.\u003c/li\u003e\n\u003cli\u003eNoyes J, Hendry M, Lewin S, Glenton C, Chandler J, Rashidian A. Qualitative \u0026ldquo;trial-sibling\u0026rdquo; studies and \u0026ldquo;unrelated\u0026rdquo; qualitative studies contributed to complex intervention reviews. Journal of Clinical Epidemiology. 2016;74:133-43.\u003c/li\u003e\n\u003cli\u003eWong G, Westhorp G, Pawson R, Greenhalgh T. Realist synthesis training materials. 2013.\u003c/li\u003e\n\u003cli\u003eGreenhalgh T, Wong G, Westhorp G, Pawson R. Protocol-realist and meta-narrative evidence synthesis: evolving standards (RAMESES). BMC medical research methodology. 2011;11:1-10.\u003c/li\u003e\n\u003cli\u003eWohlin C, editor Guidelines for snowballing in systematic literature studies and a replication in software engineering. Proceedings of the 18th international conference on evaluation and assessment in software engineering; 2014.\u003c/li\u003e\n\u003cli\u003eHirt J, Nordhausen T, Fuerst T, Ewald H, Appenzeller-Herzog C. Guidance on terminology, application, and reporting of citation searching: the TARCiS statement. bmj. 2024;385.\u003c/li\u003e\n\u003cli\u003eHunter R, Gorely T, Beattie M, Harris K. Realist review. International Review of Sport and Exercise Psychology. 2022;15(1):242-65.\u003c/li\u003e\n\u003cli\u003eHunter R, Booth A, Lamb SE, Williamson E, Hendrick P, Sahota O, et al. Prehabilitation for lumbar spinal stenosis: understanding mechanisms and contexts for enhanced engagement-a realist review. Age and Ageing. \u003cem\u003eIn press\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eBrody S. Scite. Journal of the Medical Library Association: JMLA. 2021;109(4):707.\u003c/li\u003e\n\u003cli\u003eLabcritics. Undermind.ai Review: AI-Driven Revolution in the Literature Review: Labcritics, Lab Tech Review \u0026amp; Marketplace; 2024 [Available from: https://labcritics.com/undermind-ai-review-ai-driven-revolution-in-the-literature-review/.\u003c/li\u003e\n\u003cli\u003eDada S, Dalkin S, Gilmore B, Hunter R, Mukumbang FC. Applying and reporting relevance, richness and rigour in realist evidence appraisals: advancing key concepts in realist reviews. Research synthesis methods. 2023;14(3):504-14.\u003c/li\u003e\n\u003cli\u003eBates MJ. The design of browsing and berrypicking techniques for the online search interface. Online review. 1989;13(5):407-24.\u003c/li\u003e\n\u003cli\u003eGoyes DR, Sandberg S. Trust, nuance, and care: Advantages and challenges of repeat qualitative interviews. Qualitative Research. 2024:14687941241246159.\u003c/li\u003e\n\u003cli\u003eJagosh J, Macaulay AC, Pluye P, Salsberg J, Bush PL, Henderson J, et al. Uncovering the benefits of participatory research: implications of a realist review for health research and practice. Milbank Q. 2012;90(2):311-46.\u003c/li\u003e\n\u003cli\u003eNickerson RS. Confirmation bias: A ubiquitous phenomenon in many guises. Review of general psychology. 1998;2(2):175-220.\u003c/li\u003e\n\u003cli\u003eKaplan A. The conduct of inquiry: Methodology for behavioural science: Routledge; 2017.\u003c/li\u003e\n\u003cli\u003eJagosh J. Realist synthesis for public health: building an ontologically deep understanding of how programs work, for whom, and in which contexts. Annual review of public health. 2019;40(1):361-72.\u003c/li\u003e\n\u003cli\u003eBooth A, Sommer I, Noyes J, Houghton C, Campbell F. Rapid reviews methods series: guidance on rapid qualitative evidence synthesis. BMJ evidence-based medicine. 2024;29(3):194-200.\u003c/li\u003e\n\u003cli\u003eJagosh J. Retroductive theorizing in Pawson and Tilley\u0026apos;s applied scientific realism. Journal of Critical Realism. 2020;19(2):121-30.\u003c/li\u003e\n\u003cli\u003eGe L, Agrawal R, Singer M, Kannapiran P, De Castro Molina JA, Teow KL, et al. Leveraging artificial intelligence to enhance systematic reviews in health research: advanced tools and challenges. Syst Rev. 2024;13(1):269.\u003c/li\u003e\n\u003cli\u003eGwon YN, Kim JH, Chung HS, Jung EJ, Chun J, Lee S, Shim SR. The use of generative AI for scientific literature searches for systematic reviews: ChatGPT and Microsoft Bing AI performance evaluation. JMIR Medical Informatics. 2024;12:e51187.\u003c/li\u003e\n\u003cli\u003eFeng Y, Liang S, Zhang Y, Chen S, Wang Q, Huang T, et al. Automated medical literature screening using artificial intelligence: a systematic review and meta-analysis. Journal of the American Medical Informatics Association. 2022;29(8):1425-32.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-research-methodology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmrm","sideBox":"Learn more about [BMC Medical Research Methodology](http://bmcmedresmethodol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmrm/default.aspx","title":"BMC Medical Research Methodology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Literature searches, review methodology, realist reviews, evidence appraisal, literature screening, artificial intelligence","lastPublishedDoi":"10.21203/rs.3.rs-7924327/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7924327/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground:\u003c/h2\u003e\u003cp\u003eIntegrating artificial intelligence (AI) into literature searching has the potential to enhance research synthesis by improving the identification of conceptually rich or otherwise difficult-to-locate evidence. Theoretical or conceptual literature reviews, including realist reviews, often involve resource-intensive searches because they aim to trace nuanced ideas, mechanisms, or conceptual relationships across multiple sources. This case study illustrates the use of AI-powered tools to support and streamline such literature searching, using a realist review as an example.\u003c/p\u003e\u003ch2\u003eMethods:\u003c/h2\u003e\u003cp\u003eWe applied AI tools\u0026mdash;Scite and Undermind\u0026mdash;in the context of a realist review to facilitate the identification of relevant studies. Seed papers and key informant papers guided the search, and a novel classification system (grandparent, parent, and child papers) was used to systematically organise studies for developing and refining theoretical constructs. Transparent screening procedures and decision-making frameworks were employed to ensure methodological rigour and reproducibility.\u003c/p\u003e\u003ch2\u003eResults:\u003c/h2\u003e\u003cp\u003eThe integration of AI tools improved the feasibility of identifying conceptually relevant literature and helped manage complex datasets. The classification system enabled structured organisation of studies, supporting iterative testing and refinement of theoretical constructs. The workflow demonstrated flexibility and adaptability, suggesting potential applicability beyond realist review\u003c/p\u003e\u003ch2\u003eConclusions:\u003c/h2\u003e\u003cp\u003eOur findings suggest that AI-powered tools can support and streamline literature searching, improving efficiency in identifying conceptually relevant studies. However, these tools do not replace the critical interpretive work required by researchers. Human judgement remains essential to assess relevance, evaluate nuanced concepts, and make informed decisions throughout the search process, with AI serving as a valuable adjunct rather than a substitute.\u003c/p\u003e","manuscriptTitle":"Searching Smarter, Not Harder: Leveraging AI to Enhance Literature Searches for Theory-Driven reviews. A methodological case study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-18 08:41:38","doi":"10.21203/rs.3.rs-7924327/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-12-03T08:24:53+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-18T13:35:26+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-14T17:18:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"180353378522117111852415175220408879919","date":"2025-11-07T18:38:28+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"16021775335445857170924574747247166878","date":"2025-11-06T09:36:46+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-06T08:29:17+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-29T14:53:43+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-29T14:52:26+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Research Methodology","date":"2025-10-22T13:58:46+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-research-methodology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmrm","sideBox":"Learn more about [BMC Medical Research Methodology](http://bmcmedresmethodol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmrm/default.aspx","title":"BMC Medical Research Methodology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f7a44917-7237-400d-8bf1-3ee118389683","owner":[],"postedDate":"November 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-03-09T16:12:39+00:00","versionOfRecord":{"articleIdentity":"rs-7924327","link":"https://doi.org/10.1186/s12874-026-02814-3","journal":{"identity":"bmc-medical-research-methodology","isVorOnly":false,"title":"BMC Medical Research Methodology"},"publishedOn":"2026-03-04 15:57:57","publishedOnDateReadable":"March 4th, 2026"},"versionCreatedAt":"2025-11-18 08:41:38","video":"","vorDoi":"10.1186/s12874-026-02814-3","vorDoiUrl":"https://doi.org/10.1186/s12874-026-02814-3","workflowStages":[]},"version":"v1","identity":"rs-7924327","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7924327","identity":"rs-7924327","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}