Barwise, A

Abstract Rationale Clear and accurate communication is essential in critical care, yet patients with non-English language preferences (NELP) often face limited access to professional interpreters. Although machine translation has been explored in healthcare, most studies have focused on written translation. Their use in real-time, two-way interpretation specifically for high-stakes ICU conversations remains largely unexamined. To bridge this gap, we developed AI-TransLATE (AI-enhanced Transition to Language-Agnostic Transcultural Engagement), a voice-to-voice multilingual interpretation tool designed to support real-time interactions in critical care. Following successful initial validation, we evaluated clinicians’ perceptions of its usability in simulated ICU encounters and compared its performance with commercially available AI tools, including Gemini and ChatGPT. Methods We developed three standardized ICU role-play scenarios focused on shared decision-making regarding ECMO, tracheostomy, and comfort care, designed to encourage natural, unscripted dialogue. Clinicians were asked to use the AI translation tools during these encounters in Spanish, Chinese, or Arabic, while a bilingual research team member acted as the patient’s family member. After each simulation, participants completed a post-session survey evaluating communication quality, satisfaction (using a 5-point Likert scale), and usability through the System Usability Scale (SUS). Results Eighteen clinicians participated, including residents, fellows, and attending physicians with an average of 3.6 ± 1.2 years of postgraduate experience. Fifty-four post-simulation surveys were completed across Arabic, Chinese, and Spanish languages comparing AI-TransLATE, ChatGPT, and Google Gemini. Participants rated the simulation as realistic across all cases (ECMO = 3.94 ± 0.66; Tracheostomy = 4.22 ± 0.43; Comfort Care = 4.06 ± 0.42; p = 0.32). Satisfaction was highest for AI-TransLATE (4.11 ± 0.68) and ChatGPT (4.00 ± 0.69) but lowest for Gemini (2.06 ± 1.00; p < 0.001). Willingness to use in real life followed the same trend. (AI-TransLATE = 4.00 ± 0.69, ChatGPT = 3.89 ± 0.96, Gemini = 2.00 ± 1.03 (p < 0.001)). SUS scores supported these findings (AI-TransLATE = 71.25 ± 13.35; ChatGPT = 72.22 ± 15.57; Gemini = 40.28 ± 21.30; p < 0.001), corresponding to good usability for AI-TransLATE and ChatGPT and poor usability for Gemini. Clinicians mentioned interruptions, uneven translation flow, and difficulty judging accuracy, but appreciated it when the AI-TransLATE conveyed tone and intonation, enabling smoother, more natural communication. Conclusion Overall, AI-TransLATE and ChatGPT demonstrated strong usability and satisfaction, showing promise as clinical communication tools; however, further validation in real-world ICU settings is warranted, and data privacy and confidentiality remain important considerations. This abstract is funded by: Critical care research committee mayo clinic

Abstract Introduction Large Language Models (LLMs) have proven effective at analyzing unstructured and structured data from electronic health records (EHR). Their reliance on static EHR data means they miss the dynamic, high-resolution physiological signals that are the first indicators of patient decline, creating a fundamental data lag as clinical documentation of an event may occur hours after its onset. This represents a critical gap in the ability to develop truly effective triage tools. Furthermore, while LLMs can synthesize multi-source data for acuity assessment, they exhibit known limitations, including hallucinations that constrain autonomous use in safety-critical settings. In contrast, a Vision-Language Model (VLM) with strong image-grounded clinical reasoning can interpret vital sign trend images, surpassing the limitations of current EHR-based models. This approach provides immediate triage evaluations that lead to earlier detection of patient deterioration. Methods This mixed-methods study deployed a VLM (Gemini 2.5) integrated within the eICU dashboard. We used near-real-time two-hour trends in heart rate, blood pressure, respiratory rate, and oxygen saturation from consecutive patient cases during alerts in hospital floor units monitored by eICU that admit acute medical and surgical patients. To increase the diversity of the sample, we sought routine cases from patients without alerts and emergent cases from patients admitted to the medical, surgical, trauma, and mixed ICUs. A structured context prompt was used to generate a standardized summary including clinical interpretation, data quality and artifact assessment, triage recommendation, and the model’s chain-of-thought reasoning. Results We included patients from Mayo Clinic Rochester, MN, using an 800-patient cohort for prompt development and a 100-patient cohort for validation against clinically adjudicated results. In the validation cohort the VLM achieved 80% accuracy (80/100). An analysis of the 20 incorrect predictions revealed a clear pattern of conservative over-triage: 80% (16/20) of errors were clinically acceptable adjacent-class discrepancies (routine↔urgent or urgent↔emergent). Only 4/20 discrepancies represented a substantive misclassification. Most errors therefore reflected conservative prioritization rather than incorrect assessment. The model demonstrated a perfect safety profile, underestimating patient acuity in 0% of cases, and thus avoiding the most clinically dangerous error. These results confirm the feasibility of VLM-based approach and provide a strong foundation for the proposed work. Conclusion This VLM performs cross-modal reasoning by fusing deep visual perception with broad clinical context. The result is a fully contextualized and explainable triage prioritization, which provides direct decision support and integrates into the clinical workflow, allowing for faster and more confident acute care assessments. This abstract is funded by: none