Herasevich, V

BACKGROUND. Survey research in critical care has dramatically increased over the past 25 years, and use of surveys within Delphi methods, which provide consensus definitions for heterogeneous disease states, has similarly expanded significantly. It is crucial for survey researchers to understand the nuances of survey development and expectations of information provided by surveys, to use them correctly. Few simple sources exist to guide development considerations to generate high-quality survey research. No studies to our knowledge exist that draw parallels between the principles of survey research as a guide within the Delphi process, as one potential conduit between a steering committee and expert panel. METHODS. We present a narrative review of and framework for rigorous design of survey research, highlighting the applicability to the design of surveys administered within Delphi methods. RESULTS AND CONCLUSIONS. The methodologic considerations presented for survey research and Delphi methods can inform high-quality critical care research.

Abstract Rationale Per the 2021 Sepsis Campaign, it is recommended that hospitals use a performance improvement program for sepsis screening, utilizing machine learning to improve specificity. Around 20% of sepsis at the Mayo Clinic, Rochester site is hospital acquired. As inpatient sepsis is usually detected by extreme physiologic derangement or clinical concern, we seek to leverage risk factors beyond traditional SIRS and sepsis criteria to create a machine and deep learning based risk stratification model to improve the precision of sepsis monitoring. Methods We designed a retrospective observational case control study, including patients diagnosed with sepsis as an inpatient at the Mayo Clinic, Rochester hospital and compared them to patients without a sepsis diagnosis over the last 7 years. Sepsis diagnosis was labeled by a combination of sepsis ICD codes and initial lactate elevation. We excluded transfers from outside hospitals, direct admissions for sepsis and patients diagnosed with sepsis in the emergency department. Data was extracted using Mayo Data Explorer, and data preparation and validation were performed using JMP Statistical Software and Python libraries. Feature selection was critical as we examined all data from previous contacts with healthcare, environmental factors, and dynamic trends of inpatient objective data; amongst 5,000 options of lab tests and vital measurements, we selected near 100 to reduce model complexity. We used XG Boost to train our model with a baseline cohort of 100 sepsis patients and 1000 patients without sepsis. Results We created a model that categorically classified patients into three sepsis risk level profiles (low-risk, medium-risk, high-risk) with performance metrics of a specificity of 97%, sensitivity of 65% and area under the receiver operating characteristic curve of 95%. Amongst laboratory data, lactate and creatinine had the highest feature importance (figure below). Our model also identified variables with the least importance to further refine our architecture. Conclusions Our model is one of the first sepsis risk stratification models whose algorithm is optimized for non-ICU inpatient wards, while the majority of similar studies were conducted on ICU and emergency department patients due to limited data availability. The model’s high specificity will allow for earlier sepsis recognition. By including our full cohort of 7,500 patients with sepsis and 35,000 patients without sepsis, we hope to improve class imbalance and increase sensitivity. By integrating transformer architecture, we seek to eliminate variables of low significance to create a model with more practical use, while maintaining high specificity. This abstract is funded by: None

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