Intravenous extravasation report generation using deep learning, generative artificial intelligence, and visual question answering techniques

Peripheral intravenous extravasation poses significant clinical challenges, demanding accurate assessment for timely intervention. This study advances extravasation management by integrating Visual Question Answering (VQA) technology within the Thammasat University eXtravasation Assessment Tool (TUXAT) framework, designed for generating structured clinical reports. The investigation evaluated two distinct VQA approaches-a single large model versus a mixture of models-assessing their capacity to produce reports detailing Findings (Skin Discoloration, Integrity, and Edema), Implications (Severity), and Treatment Plans based on established clinical guidelines. Inter-rater reliability analysis indicated moderate-to-substantial agreement for Discoloration assessment, although inherent subjectivity limited the precision of Integrity and Edema scoring. Statistical analysis confirmed the system's sensitivity to clinical severity (Severity effect, [Formula: see text]) and captured expected expert variability (Annotator effect, [Formula: see text]). Notably, while overall annotator assessments varied, the relative differences in their ratings across severity levels remained consistent (Severity×Annotator interaction, [Formula: see text]). Crucially, both VQA approaches yielded comparable performance profiles (Model effect, [Formula: see text]), demonstrating VQA's robustness for this application irrespective of architectural choice. These findings highlight VQA's significant potential to automate and standardize extravasation reporting, offering a promising pathway towards improved clinical decision support.