Rationalisation of a thrombophilia panel using laboratory medicine Delphi-like consensus evaluation and secondary artificial intelligence-based simulation assessment

Abstract Objectives It is important to review laboratory test panels regularly and omit unnecessary tests. This avoids overdiagnosis and makes laboratory work more targeted. Artificial intelligence is increasingly being discussed as a possible aid in such decisions. The aim of the study was to revise an existing thrombophilia panel with the help of a modified Delphi consensus of laboratory physicians and to examine whether large language models (LLMs) can mimic such decision-making processes and serve as a support tool. Methods The study was conducted in two steps. First, six experts evaluated various thrombophilia parameters in three Delphi rounds, assessing technical reliability and clinical significance. Selected LLMs (Elicit, Consensus, and STORM) were then tested with questions. Their results were compared with the Delphi consensus. Agreement was calculated using percentage concordance and Cohen’s κ. Results PAI-1 genotyping, MTHFR genotyping, homocysteine and APC resistance were removed from the standard panel, and anti-annexin-V antibodies and anti-phosphatidylserine/prothrombin antibodies were completely eliminated. The reduced panel was incorporated into routine practice, with facultative parameters remaining available as second-line tests. Agreement between the LLMs and the experts was slight when using open prompts (κ ≈ 0.25), although, with specific questions, the agreement was higher (κ 0.50–0.52). However, the LLMs did not take into account analytical and technical aspects. Conclusions Thrombophilia panels should be reviewed regularly to avoid the application of unnecessary tests and ensure high diagnostic quality. The Delphi process is a suitable tool for this. LLMs can provide supporting information, but are currently no substitute for the experience and consensus of medical experts.