ROTEM Interpretation AI vs Experts

This prospective multicenter observational study is designed to evaluate the agreement between artificial intelligence (AI)-based interpretation and expert interpretation of rotational thromboelastometry (ROTEM) findings, with a focus on clinical decision-making in critically ill patients.

ROTEM is a point-of-care viscoelastic method providing real-time information on coagulation, including clot formation, strength, and fibrinolysis. It is widely used to guide targeted hemostatic therapy in trauma, major surgery, and critical care. However, interpretation of ROTEM findings is complex and requires clinical expertise. Interobserver variability among clinicians may lead to inconsistent therapeutic decisions. Although algorithm-based approaches have been introduced, their implementation remains variable.

Artificial intelligence (AI) has the potential to standardize interpretation by integrating multiple ROTEM parameters and generating consistent recommendations. Previous studies have shown that machine learning models can predict clinical outcomes or transfusion requirements based on viscoelastic data. However, evidence on agreement between AI-based interpretation and expert interpretation, particularly in real-world clinical decision-making, remains limited.

The primary objective of this study is to determine whether AI-based interpretation achieves a level of agreement comparable to inter-expert variability in ROTEM interpretation. This study does not assume a single gold standard; instead, it evaluates agreement between methods, reflecting real-world clinical practice.

Secondary objectives include:

* comparison of interpretation time between AI and expert clinicians, * assessment of consistency (intra-method variability) of AI compared to inter-expert variability, * evaluation of potential impact on clinical decision-making, including identification of coagulation abnormalities and proposed treatment strategies.

ROTEM measurements will be collected and presented in a standardized format, including graphical and numerical outputs. Each dataset will be independently evaluated by multiple expert anesthesiologists. The same datasets will be interpreted repeatedly by an AI-based large language model using a predefined standardized prompt, with multiple independent runs to assess intra-model variability.

For each ROTEM dataset, both experts and AI will assess:

* presence of a coagulation disorder, * dominant underlying abnormality, * appropriate therapeutic intervention, * and recommended treatment dose.

Agreement between experts and AI, as well as inter-expert agreement, will be analyzed using appropriate statistical methods for categorical and continuous variables (e.g., kappa statistics and intraclass correlation coefficients). Time required for interpretation will also be recorded and compared.

This study is not designed to determine the absolute correctness of interpretation, but to quantify agreement and variability between human experts and AI. By identifying clinically relevant discrepancies, the study aims to evaluate whether AI-assisted interpretation may serve as a reliable decision-support tool and reduce variability in ROTEM-guided hemostatic management.

• Interpretation time [Time frame: Up to 24 hours after ROTEM measurement.]
• Consistency of interpretation (intra-method variability) [Time frame: Up to 24 hours after ROTEM measurement]
• Proportion of clinically discordant decisions [Time frame: Up to 24 hours after ROTEM measurement]
• Inter-expert agreement [Time frame: Up to 24 hours after ROTEM measurement]