Introduction

Knowledge synthesis, the process of integrating and summarizing relevant studies in the literature to gain an improved understanding of a topic, is a key component in identifying knowledge gaps and informing future research endeavors on a topic of interest [,]. Systematic and scoping reviews are among the most commonly used and rigorous forms of knowledge synthesis across multiple disciplines [,]. Given that the results from systematic and scoping reviews can inform guidelines, protocols, and decision-making processes, particularly for stakeholders in the realms of health care, the quality of the evidence presented by such reviews can significantly impact generated recommendations [].

The quality of systematic and scoping reviews is highly dependent on the comprehensiveness of the database searches and the subsequent article screening processes. Overlooking relevant articles during these critical steps can lead to bias [], while including discrepant studies can yield misleading conclusions and increase discordant heterogeneity []. Thus, guidelines surrounding the conduct of clinical reviews, such as the Cochrane Handbook [], recommend that article screening be completed in duplicate by at least 2 independent reviewers.

However, duplicate screening effectively doubles the financial and human resources needed to complete systematic reviews compared to single screening. This is especially problematic for small research groups, review projects with broad inclusion criteria (such as network meta-analyses), or time-constrained review projects (such as reviews relating to COVID-19 during the early stages of the pandemic) [,]. Additionally, there is often substantial interrater variability in screening decisions, leading to additional time spent on discussions to resolve disagreements []. Due to the time constraints and wasted resources that are often features of duplicate screening, research studies may also include a more tailored, sensitive search strategy that can lead to missing several articles during the retrieval process []. Furthermore, although the nuances of each study differ, many systematic reviews may contain thousands of retrieved articles, only to exclude the majority (ie, up to 90%) from the title and abstract screening [,].

Recent developments in artificial intelligence and machine learning have made it possible to semiautomate or fully automate repetitive steps within the systematic review workflow [-]. Prominent examples of such applications include RobotReviewer [], TrialStreamer [], Research Screener [], DistillerSR [], and Abstrackr [], which are artificial intelligence models developed to extract information from scientific articles or abstracts to judge study quality and infer treatment effects. More specifically, RobotReviewer (2016) was shown to have similar capabilities to assess the risk of bias assessment as a human reviewer, only differing by around 7% in accuracy []. Similarly, TrialStreamer was a system developed to extract key elements of information from full texts, such as inferring which interventions in a clinical paper worked best, along with comparisons in study outcomes between all relevant extracted full texts of a topic indexed on MEDLINE [].

While there have been previous attempts at automating the title and abstract screening process, they often involved labor- or computationally-intensive labeling, pretraining, or vectorizations []. For instance, Rayyan and Abstrackr are 2 free web tools that provide a semiautomated approach to article filtering by using natural language processing algorithms to learn when and where a reviewer includes or excludes an article and subsequently mimics a similar approach [,]. Rayyan also demonstrated high specificity, wherein 98% of all relevant articles were included after the tool had screened 75% of all articles to be analyzed in a study []. While automation using these tools was found to save time, there was still minimal to substantive risk that there would be missing studies if the tool were fully independent or automated [,]. Furthermore, current programs may use previously standard methods, including n-grams, in comparison to more updated techniques, such as the generative pretrained transformer (GPT) model, which is trained with data from a general domain and does not require additional training to learn embeddings that can represent the semantics and contexts of words in relation to other words [,].

In this paper, we introduce a novel workflow to screen titles and abstracts for clinical reviews by providing plain language prompts to the publicly available OpenAI GPT application programming interface (API). We aimed to assess GPT models’ ability to accurately and efficiently identify relevant titles and abstracts from real-world clinical review data sets, as well as their ability to explain their decisions and reflect on incorrect classifications. We compare the performance of our model against ground truth labeling by 2 independent human reviewers across 6 review papers in the screening of over 24,000 titles and abstracts.

Methods

Overview

In our study, we obtained a corpus of title and abstract data sets that have already been filtered by a minimum of 2 human reviewers to train our model (). Subsequently, we created a Python script that provides the screening criteria for each paper to the OpenAI Chat GPT or GPT-4 API, depending on the input token length. We then passed each paper to the API using a consistent instruction prompt to determine whether a paper should be included or excluded based on the contents of its title and abstract. The overall accuracy (computed by dividing papers selected by both GPT and human reviewers by the total number of papers), sensitivity of both included and excluded papers, and interrater reliability through Cohen κ and prevalence-adjusted and bias-adjusted κ (PABAK) were computed against the human-reviewed papers:

Where k is the number of categories and p_obs is the proportion of included papers. All data and code are available in Mendeley data sets [].

‎

Data Collection

To validate our proposed inclusion and exclusion methodology, we obtained 6 title and abstract screening data sets from different systematic and scoping reviews previously published by the authors of this study, each screened by 2 independent reviewers with conflicts resolved through consensus. These projects cover various medical science topics and vary in size, methodology, and complexity of screening criteria ( and Table S1 in [-]). We obtained the inclusion and exclusion decision from expert reviewers for each title and abstract entry, as well as the criteria provided to the expert reviewers during the screening process. A summary of the review characteristics is presented in .

App Creation

Given a data set, df_info, containing information about inclusion and exclusion criteria of the data sets containing titles and abstracts to be reviewed, the app calls the OpenAI GPT API to classify each paper to be screened as either included or excluded. The app was coded in Python. The prompt given to the GPT API is provided in .

Assessment and Data Analysis

After the app was run on all data sets included in our analysis, the following metrics were computed: accuracy, macro F₁-score, sensitivity for decision tags, κ, and PABAK. A subset of the results was selected for the GPT models to explain their reasoning. The following prompt was appended to the beginning of the original prompt given to the API: “Explain your reasoning for the decision given with the information below.” The human and GPT decisions were appended to the end of the prompt. A subset of incorrect results was selected for GPT to reflect on its incorrect answers. The following prompt was appended to the beginning of the original prompt given to the API: “Explain your reasoning for why the decision given was incorrect with the information below.” The human and GPT decisions were appended to the end of the prompt.

Results

The overall accuracy of the GPT models was 0.91, the sensitivity of included papers was 0.76, and the sensitivity of excluded papers was 0.91 ( and ). On the nonopioid analgesia (NOA) data set (354/14,771 included abstracts), the model ran for 643 minutes and 50.8 seconds, with an approximate cost of US $25. The data set characteristics are detailed in , the model performance is in and visualized in , and the reasoning from GPT is tabulated in .

‎

Discussion

Overview

In this study, we assessed the performance of the OpenAI GPT API in the context of clinical review paper inclusion and exclusion criteria selection. We report an overall accuracy of 0.91 and a PABAK of 0.96, indicating a high level of agreement between the app’s decisions and the reference standard. However, the κ was low, ranging from 0.21 to 0.26, largely due to the imbalanced nature of the data sets in this study. The sensitivity of the included papers was 0.76, suggesting that the app needs improvement to correctly identify relevant papers ( and ). The sensitivity of excluded papers was 0.91, showing promise in excluding irrelevant papers. These results highlight the potential of large language models (LLMs) to support the clinical review process.

Implications of GPT API’s Performance in the Review Process

GPT’s performance has several implications for the efficiency and consistency of clinical review paper inclusion and exclusion criteria selection. By prioritizing the workflow and acting as an aid rather than a replacement for researchers and reviewers, the GPT and other large language models have the potential to streamline the review process. This enhanced efficiency could save valuable time and effort for researchers and clinicians, allowing them to focus on more complex tasks and in-depth analysis. Further, the API does not require pretraining or seed articles and can provide reasoning for its decision to either include or exclude papers, an aspect traditional natural language processing algorithms lack in automated or semiautomated paper screening (). Interestingly, upon being asked to explain its reasoning for a subset of incorrect classifications, GPT corrected its initial decision. Ultimately, this increased efficiency, paired with reasoning capabilities, could contribute to the overall quality of clinical reviews, leading to more accurate and reliable conclusions in medical research.

The use of LLMs in the review process could also promote consistency in the selection of relevant papers. By automating certain aspects of the process and acting as an aid to researchers and clinicians, the model can streamline the review process and help reduce the potential for human error and bias, leading to more objective and reliable results []. This increased consistency could, in turn, improve the overall quality of the evidence synthesized in clinical reviews, providing a more robust foundation for medical decision-making and the development of clinical guidelines.

The potential of LLMs as a decision tool becomes particularly valuable when resources are limited. In such situations, LLMs can be used as a first-pass decision aid, streamlining the review process, and allowing human screeners to focus on a smaller, more relevant subset of papers. By automating the initial screening process, LLMs can help reduce the workload for researchers and clinicians, enabling them to allocate their time and effort more efficiently.

In particular, using the GPT API as a first-pass decision aid can also help mitigate the risk of human error and bias in the initial screening phase, promoting a more objective and consistent selection of papers. While the API’s sensitivity for including relevant papers may not be perfect, its high specificity for excluding irrelevant papers can still provide valuable support in narrowing down the pool of potentially relevant studies []. This can be particularly beneficial in situations where a large number of papers need to be screened and human resources are scarce [].

Comparison to Other Tools

The comparison of our proposed machine learning method to other tools, such as Abstrackr [], DistillerSR [], and RobotAnalyst [], provides evidence of its efficacy and reliability in the context of systematic review processes. On a data set of 24,307 abstracts and titles, our model achieved an accuracy of 0.91 and comparable sensitivity of 0.91 and 0.76 for excluded and included papers, respectively. The significant interrater agreement (κ=0.96) between our proposed method and consensus-based human decisions, juxtaposed to the lower interrater variability between 2 independent human screeners (κ=0.46), emphasizes the model’s robustness. In comparison, Abstrackr reported overall sensitivities of 0.96, 0.79, 0.92, and 0.82 on data sets ranging from 5243 to 47,385 records. When comparing the proportion of missed records across Abstrackr, DistillerSR, and RobotAnalyst on nonpublic medical title and abstract screening data sets, Abstrackr exhibited the lowest proportions of missed records, namely 28%, 5%, and 0%, respectively []. Conversely, DistillerSR showed a high proportion of missed records, reaching up to 100% in the last data set. RobotAnalyst’s performance fell between the 2, with missed proportions of 70%, 23%, and 100%, respectively. Future work will explore comparative analyses in greater depth and on a broader array of data sets to compare state-of-the-art screening tools.

Limitations and Challenges in Implementing GPT API in the Review Process

While the GPT API shows promise in streamlining the review process, it is important to acknowledge its limitations and challenges. One notable limitation is the disparity between the high specificity of 0.91 for excluding papers and the lower sensitivity of 0.76 for including papers. This discrepancy suggests that while the API effectively excludes irrelevant papers, it may not be as proficient in identifying relevant papers for inclusion. This could lead to the omission of important studies in the review process, potentially affecting the comprehensiveness and quality of the final review. Therefore, the GPT API should not be considered a replacement for human expertise. Instead, it should be viewed as a complementary tool that can enhance the efficiency and consistency of the review process. Human screeners should still be involved in the final decision-making process, particularly in cases where the API’s sensitivity for including relevant papers may be insufficient []. Another limitation arises in the selection of data sets for screening; 3 of the 6 data sets focused on the efficacy of various drugs for COVID-19, potentially limiting the generalizability of the results from other types of studies. Further work will assess GPT on a greater diversity of studies. By combining the strengths of the GPT API with human expertise, researchers can optimize the review process and ensure the accuracy and comprehensiveness of the final review.

Future Research and Development

Several avenues for future research and development include refining the GPT API’s performance in the clinical review paper context, incorporating metadata such as study type and year, and exploring few-shot learning approaches. Additionally, training a generator-discriminator model through fine-tuning could improve the API’s performance []. Expanding the application of the GPT API to other areas of medical research or literature review could also be explored. This would involve large language models for tasks such as identifying and extracting study design information, patient characteristics, and adverse events. As the maximum token length increases with future iterations of the GPT model, screening entire papers may become feasible []. Furthermore, exploring the use of LLMs to generate clinical review papers could be a promising research direction.

Conclusions

The GPT API shows potential as a valuable tool for improving the efficiency and consistency of clinical review paper inclusion and exclusion criteria selection. While there are limitations and challenges to its implementation, its performance in this study suggests that it could have a broader impact on clinical review paper writing and medical research. Future research and development should focus on refining the API’s performance, expanding its applications, and exploring its potential in other aspects of clinical research.