Development and Prospective Validation of the AutoEIT Foundation Model for ARDS

Condition(s) studied

Investigational drug(s) / intervention(s)

AutoEIT Model Training: Adult ARDS patients admitted to participating ICUs between January 2020 and December 2025 who received invasive mechanical ventilation and EIT monitoring. Data from this cohort will be used exclusively to train the shared spatiotemporal encoder and the five task specific prediction heads of the AutoEIT foundation model.

AutoEIT Blinded Validation: Adult ARDS patients prospectively enrolled between January 2026 and December 2026. For this cohort, strict millisecond-level synchronization between the EIT system, mechanical ventilator, and clinical events (e.g., ABG sampling) will be enforced. The AutoEIT model will run in the background (blinded to treating clinicians) to generate the 5-task predictions, which will later be compared against actual clinical ground truths.

Study summary

Acute Respiratory Distress Syndrome (ARDS) is a highly heterogeneous condition, making precision mechanical ventilation challenging. Electrical Impedance Tomography (EIT) provides real-time monitoring of regional lung ventilation, but interpreting its complex spatiotemporal data relies on subjective clinical experience. This multicenter ambispective cohort study aims to develop and validate an all-in-one Artificial Intelligence Foundation Model (AutoEIT). The study consists of two phases: (1) A Retrospective Phase (data from 2020 to 2025, N=1000) to train the AutoEIT model using continuous EIT images and synchronous high-frequency ventilator waveforms; and (2) A Prospective Temporal Phase (January 2026 to December 2026, N=200) to rigorously test the model's performance in a real-world clinical setting. The AutoEIT model is a deep multi-task learning framework designed to simultaneously perform five core clinical tasks: (1) classify ARDS subphenotypes, (2) recommend individualized positive end-expiratory pressure (PEEP), (3) detect occult patient-ventilator asynchrony, (4) predict ventilation/perfusion (V/Q) ratio, and (5) continuously estimate the non-invasive PaO₂/FiO₂ ratio. Ultimately, this system aims to reduce ventilator-induced lung injury (VILI) and improve precision respiratory care.

Eligibility

Primary outcome measure(s)

• Diagnostic Performance of ARDS Subphenotype Classification — At baseline, within 24 hours of ARDS diagnosis or Intensive Care Unit (ICU) admission.
  Area Under the Receiver Operating Characteristic curve (AUC-ROC) of the AutoEIT model for classifying ARDS subphenotypes into hyperinflammatory versus hypoinflammatory phenotypes, using latent class analysis (LCA) based on clinical and biomarker data as the reference ground truth. AUC-ROC will be calculated with 95% confidence intervals using DeLong's method. A predefined threshold of AUC-ROC ≥ 0.80 will be considered clinically acceptable.Unit of Measure: AUC-ROC (dimensionless, 0 to 1)
• Accuracy of Optimal PEEP Recommendation — At the time of the first standardized decremental PEEP trial, within 72 hours of enrollment.
  Root Mean Square Error (RMSE) between the AutoEIT-recommended positive end-expiratory pressure (PEEP) and the clinically optimal PEEP derived from the intersection point of lung collapse and overdistension curves obtained during a standardized decremental PEEP trial. RMSE will be calculated in cmH2O; lower values indicate better agreement. A clinically acceptable RMSE threshold is predefined as ≤ 2.0 cmH2O.Unit of Measure: cmH2O
• Detection Accuracy of Patient-Ventilator Asynchrony — Throughout the EIT monitoring period, within 72 hours of enrollment.
  Area Under the Receiver Operating Characteristic curve (AUC-ROC) of the AutoEIT model for detecting occult patient-ventilator asynchrony events, including pendelluft, breath-stacking, and ineffective triggering, validated independently against blinded expert review of synchronized video, EIT images, and ventilator waveforms by two intensivists. AUC-ROC will be calculated with 95% confidence intervals. A predefined threshold of AUC-ROC ≥ 0.85 will be considered clinically acceptable.Unit of Measure: AUC-ROC (dimensionless, 0 to 1)
• Accuracy of V/Q Ratio Prediction — Throughout the EIT monitoring period, within 72 hours of enrollment.
  Pearson Correlation Coefficient (PCC) between the AutoEIT-predicted ventilation/perfusion (V/Q) ratio and the physiological ground truth measured by the reference multiple inert gas elimination technique (MIGET) or a surrogate method. PCC will be calculated with 95% confidence intervals using Fisher's z-transformation. A predefined threshold of PCC ≥ 0.70 will be considered clinically acceptable.Unit of Measure: PCC (dimensionless, range -1 to 1)
• Accuracy of Non-invasive PaO2/FiO2 Ratio Prediction — Throughout the EIT monitoring period, within 72 hours of enrollment.
  Root Mean Square Error (RMSE) between the AutoEIT-predicted non-invasive PaO2/FiO2 ratio and the simultaneously measured arterial blood gas (ABG) PaO2/FiO2 ratio. RMSE will be calculated in mmHg; lower values indicate better agreement. A clinically acceptable RMSE threshold is predefined as ≤ 20 mmHg.Unit of Measure: mmHg

Trial sites (1)

More Ruijin Hospital trials in China

Other trials for the same condition