AI-assisted interpretation of arterial blood gases using a hybrid Stewart and standard base excess model

Arterial blood gas (ABG) interpretation is essential in critical care, but conventional approaches may not fully characterize complex acid-base disturbances at the bedside. The Partitioned Standard Base Excess (SBE) model provides a structured framework for decomposing metabolic acid-base components, but its routine use requires repeated calculations. Large language models may support structured reporting and visualization when combined with predefined physiologic rules and supplied clinical context.

This retrospective, single-center study evaluated the clinical concordance of a ChatGPT-based arterial blood gas interpretation framework using the Partitioned SBE model. A cohort of 1234 ICU patients was stratified according to primary admission diagnosis. The model received ABG values, selected laboratory parameters, and the primary ICU admission diagnosis as predefined input variables, calculated BECl, BEAlb, BELac, and BEOther using explicit equations, generated graphical visualizations, and produced structured clinical comments and management suggestions. Two independent anesthesiologists reviewed the model outputs for clinical concordance.

The model correctly reproduced all predefined partitioned SBE calculations and reference-range classifications for ABG and electrolyte parameters. Clinical comments and management suggestions were rated concordant in 93.1% of cases (95% CI: 91.6-94.4). Concordance rates exceeded 96% in renal, metabolic, and sepsis subgroups, whereas lower concordance was observed in hematologic and malignancy subgroups. The graphical decomposition of acid-base components provided a visual representation of mixed metabolic disturbances.

A formula-guided ChatGPT-based framework using the Partitioned SBE model generated interpretable ABG reports, graphical visualizations, and contextual clinical comments with high concordance to expert assessment. These findings support its potential role as an adjunctive educational and reporting tool, although prospective validation, external testing, and outcome-based studies are required before clinical implementation.

A graphical summary of the model's stepwise interpretation process is provided separately as the Visual Abstract Figure.