AI-ECHO Hemodynamic Phenotyping in Septic Shock

Background

Septic shock is the most severe stage of sepsis and is associated with mortality as high as 40%-50%. Its pathophysiological core involves profound hemodynamic disturbances resulting from complex interactions among abnormal vascular tone, relative or absolute intravascular volume deficiency, myocardial depression, and microcirculatory dysfunction.

Accumulating evidence has demonstrated substantial heterogeneity in hemodynamic profiles among patients with septic shock, including hyperdynamic states, left ventricular systolic or diastolic dysfunction, right ventricular dysfunction, and various combined phenotypes. The traditional binary classification of "warm shock" versus "cold shock" is too simplistic to capture this complexity. Such heterogeneity may partly explain why uniform treatment strategies, such as early goal-directed therapy, have yielded inconsistent results and may cause iatrogenic harm in certain subgroups.

Echocardiography has become an important tool for hemodynamic assessment in the intensive care unit (ICU). It enables noninvasive, real-time evaluation of cardiac structure and function, including left ventricular ejection fraction (LVEF), global longitudinal strain (GLS), right ventricular function, diastolic function, cardiac output, and volume status. However, its clinical utility is limited by:

• High operator dependency and interobserver variability;
• Time-consuming post-processing and interpretation;
• Difficulty in performing frequent dynamic assessments in busy ICU settings; and
• Limited integrated analysis of multiple parameters over time.

Recent advances in artificial intelligence (AI), particularly deep learning, have shown considerable promise in medical image analysis. In echocardiography, AI has demonstrated promising performance in view classification, cardiac structure segmentation, and automated measurement of functional parameters. Meanwhile, unsupervised machine learning methods provide potential approaches for identifying disease subtypes from high-dimensional clinical data.

Building on the established "Guangzhou Sepsis Cohort" and prior research experience in AI-assisted critical care medicine, this study will integrate AI and echocardiography to systematically characterize hemodynamic phenotypes and their dynamic evolution in patients with septic shock.

Study Design

This is a prospective, single-center, observational cohort study. A total of 360 adults with septic shock admitted to the ICU of The Second Affiliated Hospital of Guangzhou Medical University are anticipated to be consecutively enrolled.

Echocardiographic examinations will be performed on Day 1, Day 3, and Day 7 after enrollment. Clinical data, laboratory measurements, treatment information, and clinical outcomes will be collected according to the study protocol.

At each echocardiographic time point, standardized point-of-care echocardiography will be performed using the same model of ultrasound system and a phased-array transducer. Cine loops of the apical four-chamber, two-chamber, and three-chamber views will be acquired for at least three cardiac cycles. Additional Doppler, M-mode, right ventricular-focused, left ventricular outflow tract, and inferior vena cava recordings will be acquired as required for the prespecified echocardiographic measurements and as detailed in the imaging manual. All images will be stored in de-identified DICOM format using a unique study identification code.

AI-derived measurements will be used for research purposes only and will not guide clinical management. Clinical treatment will be determined by the treating clinical team and will not be assigned by the study.

Clinical data collected at or near the echocardiographic examinations will include vital signs, laboratory measurements, fluid balance, vasopressor dosage, and other treatment information. Blood samples will be collected on Day 1, Day 3, and Day 7 for targeted proteomic and metabolomic analyses, as specified in the study protocol.

Study Objectives

Primary Objective:

To identify and characterize distinct hemodynamic subtypes of septic shock using an unsupervised Gaussian mixture model based on prespecified multidimensional echocardiographic parameters.

Secondary Objectives:

• Longitudinal changes and dynamic evolution of hemodynamic status within 7 days after enrollment;

• Associations of hemodynamic phenotypes with clinical severity and clinical outcomes;
• Agreement between AI-derived and blinded expert echocardiographic measurements;
• Exploratory associations of targeted proteomic and metabolomic profiles with hemodynamic phenotypes; and
• Exploratory development and internal validation of an early-warning model for 28-day all-cause mortality.

Study Population

Adults (≥18 years) with septic shock, diagnosed according to the Sepsis-3 criteria, who meet the eligibility criteria listed in the Eligibility section will be consecutively enrolled from the ICU of The Second Affiliated Hospital of Guangzhou Medical University. Eligible participants must have an expected ICU stay of at least 72 hours, and written informed consent must be obtained from the patient or a legally authorized representative.

Outcome Measures

Primary Outcome:

Hemodynamic subtypes identified using an unsupervised Gaussian mixture model based on prespecified multidimensional echocardiographic parameters collected within 7 days after enrollment.

Secondary Outcomes:

• Agreement between AI-derived echocardiographic measurements and blinded expert reference measurements, evaluated using intraclass correlation coefficients and Bland-Altman analyses;
• Associations between hemodynamic subtypes and 28-day mortality;
• Associations between hemodynamic subtypes and organ failure severity, as assessed by the Sequential Organ Failure Assessment (SOFA) score;
• Duration of vasopressor use; and
• ICU and hospital lengths of stay.

Statistical Analysis

Sample Size Considerations: The planned enrollment of 240 participants is based on the anticipated number of eligible septic shock admissions during the recruitment period, expected consent and follow-up completion, and study feasibility. The target sample size is intended to provide an adequate dataset for exploratory hemodynamic phenotyping and longitudinal echocardiographic analyses. As the primary objective is exploratory hemodynamic phenotyping using unsupervised analysis, no formal hypothesis-based power calculation was performed. Analyses of clinical outcomes, multi-omics profiles, and early-warning models will be considered exploratory, and model complexity will be limited according to the effective sample size and number of observed events.

Continuous variables will be summarized as mean ± standard deviation or median with interquartile range, as appropriate. Categorical variables will be summarized as frequencies and percentages.

An unsupervised Gaussian mixture model will be applied to identify hemodynamic subtypes based on prespecified echocardiographic parameters. Clinical characteristics, treatment information, and outcomes will be compared among the identified subtypes.

Agreement between AI-derived measurements and blinded expert reference measurements will be evaluated using intraclass correlation coefficients and Bland-Altman analyses. Targeted proteomic and metabolomic analyses will evaluate associations between protein or metabolite abundances and hemodynamic phenotype membership using methods appropriate for high-dimensional data. Where applicable, the false discovery rate will be used to account for multiple comparisons. These analyses will be considered exploratory.

An exploratory early-warning model for 28-day all-cause mortality will be developed using prespecified candidate predictors. Model discrimination will be assessed using the area under the receiver operating characteristic curve, and calibration will be evaluated using calibration plots and calibration statistics. Internal validation will be performed using bootstrap resampling or cross-validation. The number of candidate predictors and model complexity will be limited according to the number of observed events to reduce the risk of overfitting.

Multivariable logistic regression or Cox proportional hazards models will be constructed, as appropriate, to explore factors associated with adverse clinical outcomes. Repeated measurements and missing data will be addressed according to the prespecified statistical analysis plan.

Statistical analyses will be performed using SPSS, R, and Python. A two-sided P value <0.05 will be considered statistically significant, with exploratory analyses interpreted accordingly.