Search Filters to Identify Automation in HEOR: An Umbrella Review of Performance and Overlap

Introduction

Health Economics and Outcomes Research (HEOR) plays a pivotal role in ensuring patients have access to new technologies that have been evaluated for efficacy, safety, and cost-effectiveness, thereby optimizing resource allocation within the healthcare sector [¹]. With the escalating costs of new and innovative medicines and the constraints on public funding, the efficient allocation of healthcare resources is imperative to ensure that cost-effective treatments are prioritized and accessible to those in need. However, the health technology appraisal (HTA) landscape is also becoming more complex, with frequent changes to reimbursement legislations, and with more advanced and complex HEOR analytics (e.g., multi-level network meta regression [ML-NMR], synthetic control arms, quantitative bias analysis), it is not surprising that the consideration on the use of artificial intelligence (AI) in HEOR is being considered, to expedite timelines, to get new technology to patients on time [²].

HTA bodies have begun using AI in reimbursement submissions. Most recently, the National Institute for Health and Care Excellence (NICE) accepts the use of AI, however, its use must be declared, the methods transparently reported, should only be applied if there is a rationale, and its use adds value, and most critically, that the AI only augments human involvement, and does not replace it [³]. Additionally, CADTH outlines the use of AI in systematic literature reviews (SLRs) to automate data extraction and searching [⁴]. Additionally, the Institut National d’Excellence en Santé et Services Sociaux (INESSS) created a proof-of-concept demonstrating the application of a large language model (LLM) based on GPT-4 to aid title and abstract screening [⁵].

As AI continues to be integrated into HTA submissions, researchers, decision-makers, and HEOR experts must remain up to date with the latest developments in AI methodologies, applications, and software. The rapid expansion of research on AI and automation in HEOR has created a vast, continually growing body of literature, posing significant challenges for efficiently identifying the most relevant studies. Current search strategies often struggle to balance sensitivity (capturing all relevant studies) and specificity (excluding irrelevant studies), leading to either information overload or missed key evidence [⁶]. There is therefore an urgent need for robust, validated search filters that can accurately navigate the complex, abundant research landscape in this domain. This umbrella review aims to systematically identify, evaluate, and compare the performance and overlap of search filters used to identify evidence on automation in HEOR. To our knowledge, few studies have compared the overlap and performance of filters used to retrieve automation studies in HEOR-relevant contexts, highlighting a critical gap in the literature.

Methods

The umbrella review followed the reporting requirements of the Preferred Reporting Items for Systematic reviews and Meta-Analyses 2020 (PRISMA) [⁷] and the methodology outlined in the Cochrane Handbook for Overviews of Reviews [⁸]. The protocol for this systematic review is registered in the Open Science Framework (OSF) platform (https://osf.io/dzumv).

Studies were included if they were (a) SLRs, with reproducible search filters, or publications on search filters; (b) SLR or search filter was on any automation; (c) SLRs on automation was tested in HEOR, defined as SLRs, indirect treatment comparisons (ITCs) or health economic modeling; (d) full-text publication; (e) written in English; and (f) published in 2023 or later, to capture studies since the availability of ChatGPT.

Results

A total of 11,319 records were identified from EMBASE via OVID. No duplicates were identified after the records were imported into EndNote X9 software for deduplication. Out of these, 25 reports were retrieved for full-text review. Among these 25 reports, five SLRs were included. A flow of literature and reasons for exclusion are outlined in Figure 2.

An additional three reports were included: a search filter from the ISSG Search Filter Resource [¹⁷] and two SLRs identified through forward citation chasing in Google Scholar [¹⁰,¹⁵]. Among the eight reports [¹⁰,¹⁴,¹⁷,¹¹,¹³,¹²,¹⁵,¹⁶], seven were search filters used in an SLR [¹⁰,¹⁴,¹¹,¹³,¹²,¹⁵,¹⁶] and one was a standalone search filter on the ISSG website [¹⁷]. One SLR, combined searches from four previously identified reviews on SLR automation [¹⁵]. As a result, this umbrella review identified 11 individual search filters.

Among the seven SLRs [¹⁰,¹⁴,¹¹,¹³,¹²,¹⁵,¹⁶], two had objectives related to the automation and processing of biomedical literature, as well as to the preparation of SLRs [¹⁴,¹¹]. Two evaluated the automation of any SLR task [¹⁰,¹⁵]; one restricted to the automation of data extraction in preparation of an SLR [¹³]; another to the automation of data extraction from grey literature and soft data [¹²]. Lastly, one SLR was restricted to automation used in the screening of cancer topics [¹⁶]. The number of title and abstract records screened across the SLRs ranged from 3,947 [¹⁶] to 27,542 [¹³]; between 7 [¹⁶] and 721 [¹⁴] full-texts assessed for eligibility; and between 5 [¹⁶] and 273 [¹¹] reports included. A summary of the seven SLRs is provided in Table 1.

Across the seven SLRs [¹⁰,¹⁴,¹¹,¹³,¹²,¹⁵,¹⁶], 16 bibliographic databases were searched, with PubMed used most frequently (6 out of 7 SLRs) [¹⁴,¹¹,¹³,¹²,¹⁵,¹⁶]; followed by EMBASE [¹⁴,¹¹,¹⁶], Web of Science [¹⁰,¹¹,¹³] and Institute of Electrical and Electronics Engineers (IEEE) [¹⁰,¹¹,¹³], each searched by three SLRs. In addition to IEEE, three more bibliographic databases were identified and searched by two SLRs each, specializing in computing and technology; these databases were: the Association for Computing Machinery (ACM) Digital Library [¹⁰,¹¹]. ACL Anthology [¹³,¹²], and dblp [¹³,¹²]. For a full list of bibliographic databases searched by the SLRs, see Table 2. One of the reports identified was a search filter published by the University of Alberta, and available on the ISSG Search Filter Resource website [¹⁷]. The search filter, which is not validated, was translated for EMBASE, MEDLINE, PsycINFO, Cumulative Index to Nursing and Allied Health Literature (CINAHL), Scopus, and Cochrane Library [¹⁷].

Of the eight reports summarizing 11 search filters on automation, five used subject headings [¹⁴,¹⁷,¹¹,¹³,¹²]. Four search filters exploded the subject headings ‘artificial intelligence’ [¹⁷,¹¹,¹³,¹²], ‘deep learning’[¹⁷,¹¹,¹³,¹²], ‘machine learning’[¹⁷,¹¹,¹³,¹²], and ‘natural language processing’ [¹⁴,¹¹,¹³,¹²]. Followed by two search filters, each exploding subject headings on ‘data mining’ [¹³,¹²], ‘supervised machine learning’ [¹⁷,¹³], support vector machine’[¹⁴,¹³], and ‘unsupervised machine learning’ [¹⁷,¹³]. Subject headings such as ‘artificial intelligence’, ‘automation’, ‘data mining’, ‘machine learning’ (using both the exact term and the wildcard search ‘machine learn*’), and ‘natural language processing’ were also searched as major headings in PubMed and as descriptors in EMBASE. For a full list of subject headings searched, see Table 3.

Across the 11 search filters, 196 variations of free text terms were used to capture studies on automation. Less than a third of these free-text terms were used by more than one search, with only twelve used in three or more search filters. “Artificial intelligence” was the most frequently utilized in all search filters, followed by “machine learning” and “natural language processing”, each present in six search filters (see Figure 3. For the full list of search terms used, please see (Supplement Table 2).

When grouping subject headings and free text terms into common concepts, for example, the following search terms were grouped under machine learning (general terms): “machine learning”, “ml”, [machine NEXT/1 (intelligence OR learning)], exp machine learning/, machine learning/de); “machine learning (general terms)” and “automation” were the most common constructs searched across the search filters (8 search filters each). This was followed by “mining” and “natural language processing” (7 search filters each), and “text grouping”, “neural network”, “identification, interpretation and pattern recognition”, “support vector machine”, and “artificial intelligence” (6 search filters each). “Deep learning” and “machine learning approaches” were searched in 5 search filters, while “semi automation” and “decision support systems” were searched in 4. The remaining concepts were searched in 3 (“AI models and application”, “extraction”, and “text analysis”), and 2 search filters (“knowledge systems”). The cluster with the most variation in terms used was “AI models and applications”, followed by “machine learning approaches”. All concepts identified are shown in Figure 4, with a list of terms for each concept in (Supplement Table 3). Similar search filter characteristics were identified by the PRESS checklist (Supplement Table 4): only 5 of the included reports used subject headings, and the majority of the search filters reported a narrow list of keywords. Only five of the 11 search filters included truncation. No incorrect use of system syntax and spelling errors were identified. A mapping framework (Supplement Table 5) was developed to link each included automation review or search filter to HEOR-relevant evidence-generation tasks. Reviews were categorized by their core automation focus (e.g., AI/NLP, ML, decision systems) and cross-referenced with HEOR functions, including SLR automation for economic model inputs, ITCs, and HTA submissions. Each review/filter was coded for “HEOR applicability (Y/N)” with justification provided under the column “Basis.”

Among the seven SLRs [¹⁰,¹⁴,¹¹,¹³,¹²,¹⁵,¹⁶], one of which searched using 4 previously identified search filters, and combined using “or” [¹⁵]; none had a title and abstract inclusion rate (defined as the proportion of studies selected for full-text retrieval out of the total number of title and abstract records screened) above 8%, with two searches below 1% [¹⁰,¹⁶] (see Table 4). The full-text inclusion rate (defined as the proportion of studies included after full-text review compared to the total number of full-text articles retrieved for evaluation) varied from 27% [¹⁴,¹³] to 86% [¹¹]. The overall inclusion rate (defined as the proportion of studies included after the full review compared to the total number of title and abstract records initially screened) ranged from 0.13% [¹⁶] to 2.31% [¹⁵]. When evaluating the number of hours needed to include one study the SLR by Santos [¹¹], followed by Toth [¹⁵] had the best performance (0.6 hours and 0.8 hours, respectively), followed by Doneva (1.1 hours) [¹⁴], Schmidt(b) (1.2 hours) [¹²], De la Torre-Lopes (2.8 hours), Schmidt(a) (4.0 hours) [¹³], with Yao, which was only interested in AI used in HEOR among oncology studies, requiring 8 hours of screening to identify one included study [¹⁶] (see Table 4). Of note: performance metrics could not be calculated for the standalone search filter identified from the ISSG Search Filter resource, as it did not present a flow of literature as part of an SLR. These time metrics were directly extracted from the included reviews, not calculated by the authors.

For this umbrella review, a comprehensive search was conducted in EMBASE (via Ovid), the InterTASC Information Specialists’ Sub-Group (ISSG) Search Filter Resource, and Google Scholar. The search covered studies published between January 1, 2023, and July 3, 2024, to capture the post-ChatGPT era when large language models (LLMs) became widely adopted. Search terms combined concepts for automation and artificial intelligence (AI) with terms representing systematic evidence synthesis (systematic review, indirect treatment comparison, economic model). In contrast, the included SLRs and search filters used a variety of databases and approaches. Across the eight included publications, 16 databases were searched, with PubMed being the most frequently used (6/7 SLRs), followed by EMBASE (3/7). This distinction between our umbrella review’s search process and that of the included reviews is summarized in Table X (“Databases Searched: This Review vs. Included SLRs”). Among the eight included publications (seven SLRs and one standalone filter), three reviews (38%) explicitly evaluated automation methods within HEOR-relevant contexts, such as SLRs for cost-effectiveness models or indirect treatment comparisons.

Discussion

There is very little overlap or consensus among existing search filters for identifying papers on AI in HEOR. Of the 11 search filters identified, only five used subject headings, and among these, no subject heading was searched across all five. For example, only four out of 31 different subject headings (‘artificial intelligence’, ‘deep learning’, ‘machine learning’, and ‘natural language processing’) were searched across four of the five filters. There were also differences in how subject headings were searched: as major headings in PubMed and as descriptors in EMBASE. The PRESS checklist further confirmed the deficits in the search filters, identifying fewer reports using subject headings and the majority of the searches reporting a narrow list of keywords. Quantitative overlap analysis (e.g., Jaccard similarity) was not conducted because the terminology across filters was highly heterogeneous and inconsistently formatted, making reliable computation of similarity indices infeasible without extensive normalization.

A lack of consistency in search filters was identified among free-text terms, with less than a third used by more than one search, and only 12 by more than three searches. With the abundance of AI research in the HEOR field and no uniform, validated search filter available, it is not surprising that the SLRs ranged from just under 4,000 to over 27,000 records. This equated to a wide range of title and abstract screening time, from 0.6 hours to 8 hours, to identify one relevant study. This wide variation underscores the potential importance of a validated, optimized filter that can supplement rapid access to high-quality evidence supporting timely health technology assessment and reimbursement decisions.

To the best of the author’s knowledge, this umbrella review appears to be the first systematic review to explore the identification and performance of existing AI search filters in HEOR. One of the identified SLRs combined the search strategies of previous reviews to provide a comprehensive overview of automation studies in the SLR process, but did not evaluate the search filters themselves. Our research identified seven additional search filters that were not included in the previous publication. Our research provides a focused synthesis of the overlap among existing search filters for retrieving AI studies. It explores performance using metrics such as the title-abstract inclusion rate, full-text inclusion rate, and overall inclusion rate.

With key decision-making bodies such as CADTH, INESSS, and NICE exploring the use of AI in HTA submissions, as a community, we must stay ahead of new research to adopt the most reliable and proven automation methods in HEOR. To do this, we must develop a validated search filter that balances the sensitivity and specificity of AI records, specifically for HEOR, to make this manageable. We observed from our search that we had to screen 11,319 records, to full-text review 25 reports, and to include 5 reports from the electronic database search. Most research was excluded at the title and abstract stages because it focused on the direct application of AI in medical and healthcare settings, not HEOR. By creating a tested and validated search strategy, it is more viable that a living SLR on this topic can be maintained, with insights shared with HEOR experts and decision-makers.

Although this was a systematic, umbrella review of search filters used in AI research for HEOR, a validated search filter was not used, as none had yet been developed. In addition, only a single screening was conducted with quality checks, which may indicate some human error in missing studies; however, this was mitigated by reference checking and forward citation chasing. Only EMBASE was searched, while the field of AI may require automation-specific databases such as the ACM Digital Library and IEEE; this needs to be thoroughly investigated. A systematic search restricted to the EMBASE database represents a methodological limitation, especially in the context of AI and LLMs, where relevant work often appears in interdisciplinary and technical venues. Rapidly evolving AI terminology can shorten the longevity of search filters, as keywords and concepts may quickly become outdated or replaced. This can reduce search precision and recall over time, requiring regular updates or adaptive filtering to maintain relevance and effectiveness. As included SLRs were not quality-assessed using a validated tool, such as the AMSTAR2 tool, this may affect the strength of the conclusions. Lastly, restricting studies published in 2023 or later might introduce bias and limit the generalizability of the review’s findings. Broader multilingual and multi-database searches could yield a more accurate picture of true variability across automation-related reviews.

To ensure technical accuracy and reproducibility, all search filters extracted from the included reviews were verified against the native database syntax for both PubMed (MEDLINE) and EMBASE (via Ovid). Verification focused on identifying and correcting invalid or outdated field tags, truncation errors, Boolean logic inconsistencies, and incorrect use of subject headings. For PubMed, MeSH terms were verified against the current 2024 MeSH database. Only exact headings (e.g., “Artificial Intelligence”[Mesh]) were retained; wildcards and free-text modifiers were removed from MeSH field tags. Free-text terms were confirmed to use the [tiab] or [tw] fields as appropriate. For EMBASE, Emtree terms were verified using the 2024 Emtree browser. Exploded subject headings (e.g., ‘artificial intelligence’/exp) were checked for availability and correct hierarchy placement. Incorrect truncation (e.g., “machine learn”/de*) and misplaced slashes were corrected according to Ovid syntax.

Future studies must focus on using these initial findings to create a validated search filter on AI in HEOR. Filters will need to explore the right AI search terms, but the most appropriate combination of terms to identify research in HEOR specifically. Suppose a search can be highly specific while still sensitive enough to ensure relevant studies are not missed. In that case, we can start standardizing how the industry searches for research in a rapidly evolving field. However, any search filter for AI must acknowledge that periodic review will be required as new AI methods emerge that may require additional search concepts, and as older methods are retired.

This umbrella review identified significant variability in the search filters used for identifying literature on automation in HEOR, highlighting the lack of a standardized approach. The low inclusion rates across the SLRs indicate challenges in optimizing search filter performance. The variation in search strategies and the high yield of records on AI in HEOR call for more research to develop, validate, and standardize search filters in this field to optimize efficiency and keep up to date with the ongoing stream of new evidence.

Conclusion

This umbrella review identified significant variability in the search filters used for identifying literature on automation in HEOR, highlighting the lack of a standardized approach. The low inclusion rates across the SLRs indicate challenges in optimizing search filter performance. The variation in search strategies and the high yield of records on AI in HEOR call for the need for more research to create, validate, and standardize search filters in this field to optimize efficiency, and keep up to date, with ongoing stream of new evidence.