The UT-MOST team's latest artificial intelligence project is exploring the use of AI for the identification altered mental status (AMS) described in unstructured clinical notes.
A change in a patient’s mental status can be an important warning sign of sepsis. Yet those observations may sit in different parts of the electronic health record. Signs of altered mental status (AMS) may be documented in physician notes, nursing assessments, emergency department narratives, or progress notes, making it difficult to bring them together when assessing a patient’s risk.
That is the challenge behind UT-MOST's latest artificial intelligence project shared at the September collaborative meeting. The team is exploring the use of AI for the identification altered mental status (AMS) described in unstructured clinical notes, so this information can be combined with the structured data already collected through nursing and other clinical workflows.
The work began with a proof of concept using a publicly available EHR dataset. Using retrospective discharge summaries from patients with sepsis or septic shock, the team evaluated three large language models against clinician review. While the results suggest that AI can identify documented AMS in narrative text, they do not yet show that the models can improve bedside screening or recognize sepsis earlier.
The next phase will examine how findings from notes compare with structured information in the electronic health record. This matters for the National Early Warning Score 2 (NEWS2), which includes new confusion as one of its components (Royal College of Physicians, 2017). The 2026 Surviving Sepsis Campaign guidelines include NEWS2 among the recommended options when a single tool is used to screen acutely ill hospital patients for sepsis (Society of Critical Care Medicine, 2026).
UT-MOST aims to test whether combining these sources can produce more accurate NEWS2 scores. The longer-term goal is to help clinicians recognize deterioration and possible sepsis sooner. Reaching that goal will require further validation using documentation available during care and careful assessment of how AI findings fit into clinical workflows.
The opportunity is clear: important clinical insight may already be in the patient's record. Connecting the pieces could help make that insight more useful.
The team will present its AMS proof of concept at the UT Real Health AI Symposium, October 7th and 8th.
References
Royal College of Physicians. (2017). National Early Warning Score (NEWS) 2: Standardizing the assessment of acute-illness severity in the NHS. https://www.rcp.ac.uk/resources/national-early-warning-score-news-2/
Society of Critical Care Medicine. (2026). Surviving Sepsis Campaign: International guidelines for management of sepsis and septic shock 2026. https://www.sccm.org/clinical-resources/guidelines/guidelines/surviving…;
