Artificial intelligence (AI)-assisted systematic review and meta-analysis in medical sciences; How close are we to automated synthesis of evidence? A perspective

Background and aims: SR/MA is still considered somewhat new. Meta-analysis is less thana century old. Yet, the methods and guidelines on conducting them is ever changing. Artificialintelligence (AI) is often discussed as an accelerator of complex tasks in various health-relatedfields. This paper discusses the feasibility, reliability, and validity of synthesized evidence usingAI-generated SR/MA. We also discuss if AI is, or in the future could be, a smart enough means toprovide reasonable clinical decision-making based on such analyses.Method: We searched PubMed, Google Scholar, and Scopus for studies utilizing AI in health-relatedresearch. The studies were analyzed in the format of narrative review in terms of feasibilityand validity of using AI for qualitative and quantitative synthesis of evidence, field of study, addressingrisk of bias, and other criteria. No limit was considered for search data.Results: In a recent work, AI synthesized evidence for toxicity of hydroxychloroquine by combinationaluse of AI with human analysis. Study produced a clinical guidance in ۳۰ minutes. Authorsidentified ۱۱ articles evaluation ocular toxicity as an adverse event of hydroxychloroquineand approximated the rate to be ۳.۴%. Heterogeneity among individual articles was excessive.Although the use of AI was able to significantly accelerate the process, the time for preparation ofthe AI algorithm for each use case should also be considered. This was still not fully automated. Itis difficult to tweak AI to replace an expert panel that suggest how clinically relevant each resultmay be. Another advanced form of SR/MA is network meta-analysis (NMA) which summarizesmore complex study designs. For instance, NMA is able to rank more than two treatments, giventhe original studies are of proper design and quality. The model writing and choosing parametersare time-consuming and difficult. Often, NMAs are conducted via Markov chain Monte Carlotools like WinBUGS, demanding a model and information to be determined using a specific syntax.Automation is useful for simulations in which the significant number of NMAs that have to beestimated may prevent manual model specification and analysis. The authors present a method forgeneration Bayesian homogeneous variance random effects consistency models, such as selectingbase settings and trial baselines, priors, and initial values for the Markov chains. In ۲۰۲۲, a SRshowed that significant human confirmation still seems mandatory currently in implementing AImethods for health SR/MA. The application of AI in systematic analyses of evidence has a lot ofroom for improvement. An AI with the ability to fully-automate meta-analysis from protocol todesign to providing clinical advice may be the ultimate goal that the field is currently far from,but is slowly approaching.Conclusion: As the evidence stands, AI accelerated conducting a SR/MA, making what is normallypossible in months possible in minutes/hours. Decision-makers should consider the timeand effort put into developing AI technology. Optimal AI should be tweakable for different usecases of meta-analysis. AI-based studies on preclinical research are lacking, possibly due to diversedesigns.

کلیدواژه ها:

Artificial Intelligence ، systematic review and meta-analysis ، health sciences ، research methodology