Large Language Models for Viscoelastic Tests: A Head-to-Head Evaluation of ChatGPT, Gemini, and DeepSeek for Thromboelastography and ROTEM Interpretation

Introduction Hemostatic decisions are frequently time critical and must often be made under substantial diagnostic uncertainty, with management initiated based on incomplete clinical information and preliminary laboratory or point-of-care tests. Thromboelastography (TEG) and rotational thromboelastometry (ROTEM) are viscoelastic hemostatic assays (VHAs) that provide an immediate assessment of a patient's hemostatic status and mechanism of coagulopathy, enabling earlier, goal-directed therapy compared with conventional coagulation tests.[ 1 ] [ 2 ] [ 3 ] These point-of-care tests generate dynamic graphical outputs representing clot initiation, propagation, strength, and lysis in whole blood, guiding rapid transfusion management in settings such as trauma, cardiac surgery, liver transplantation, and obstetrics.[ 4 ] [ 5 ] [ 6 ] However, the clinical utility of VHAs is dependent on accurate and timely interpretation of complex multiparametric outputs, which suffers from inter-laboratory and clinician variability, making its interpretation poorly reproducible. Convolutional neural networks and other deep learning architectures have rapidly advanced the automated interpretation of complex biomedical signals, including electrocardiograms, electroencephalograms, and spirometric readings, which requires the analysis of complex graphical outputs, including lines, curves, and waveform shapes.[ 7 ] [ 8 ] These analytical requirements are also seen in VHAs, which involve the interpretation of dynamic graphical outputs representing clot formation, strength, and lysis over time. The graphical nature and need for nuanced pattern recognition in VHA interpretation make them well suited to task-specific deep learning models, which have demonstrated proficiency in extracting and classifying waveform features in other biomedical signal domains.[ 7 ] [ 9 ] More recently, large language models (LLMs) are increasingly adopted in clinical decision-making, with DeepSeek utilized in hospitals in China, assisting in interpreting clinical data and formulating diagnosis and treatment plans.[ 9 ] [ 10 ] [ 11 ] Various studies have also demonstrated LLMs' high overall accuracy in solving clinical cases, occasionally outperforming clinicians in generating differential diagnoses for complex cases.[ 12 ] [ 13 ] This has led to the rapid adoption of LLMs in clinical settings, despite limitations and lack of validation.[ 14 ] [ 15 ] Although LLMs are a class of generative deep learning systems that have demonstrated strong capabilities for multimodal reasoning and visual pattern recognition, they are not specifically trained for biomedical interpretation.[ 15 ] Amid its growing application in data-driven diagnosis and risk stratification, challenges of trust, interpretability, and validation remain.[ 16 ] This study evaluated the comparative performance of three widely used LLM platforms in interpreting ROTEM and TEG outputs from published clinical cases, focusing on pattern recognition accuracy and management appropriateness. Supplementary Material Supplementary Material (PDF) (opens in new window) Publication History Received: 09 February 2026 Accepted after revision: 17 July 2026 Article published online: 28 July 2026 © 2026. Thieme. All rights reserved. Thieme Medical Publishers, Inc. 333 Seventh Avenue, 18th Floor, New York, NY 10001, USA