Development and Prospective Validation of the AutoEIT Foundation Model for ARDS

This is a multicenter ambispective cohort study designed to develop and prospectively validate an all-in-one Artificial Intelligence Foundation Model (AutoEIT) for the precision management of Acute Respiratory Distress Syndrome (ARDS). The study comprises two distinct phases:

Phase 1 - Retrospective Training and Internal Validation (2020-2025): A total of 1000 adult ARDS patients who received invasive mechanical ventilation and continuous Electrical Impedance Tomography (EIT) monitoring across three participating tertiary hospitals will be enrolled. High-frequency ventilator waveforms (≥ 200 Hz), synchronized EIT image sequences (≥ 20 frames/second), and concurrent arterial blood gas (ABG) measurements will be extracted from electronic health records. This dataset will be used to train the shared spatiotemporal encoder and five task-specific prediction heads of the AutoEIT foundation model using a deep multi-task learning framework. Internal validation will be performed via 5-fold cross-validation, with a 70/30 training-to-validation split.

Phase 2 - Prospective Temporal and External Validation (January 2026 - December 2026): A total of 200 adult ARDS patients will be prospectively enrolled at the same centers. For this critical validation phase, millisecond-level synchronization between the EIT system, mechanical ventilator, and clinical events (e.g., ABG sampling timestamps) will be enforced prior to data recording. The AutoEIT model will run in the background in a fully blinded manner (i.e., clinicians will not receive any model-generated outputs), and its five-task predictions will be compared against the clinical ground truths at the end of the monitoring period.

The Five Core Clinical Tasks of the AutoEIT Model:

ARDS Subphenotype Classification: The model will classify patients into hyperinflammatory versus hypoinflammatory phenotypes using latent class analysis (LCA) as the reference standard. Performance will be evaluated by Area Under the Receiver Operating Characteristic curve (AUC-ROC), with a predefined clinically acceptable threshold of AUC-ROC ≥ 0.80.

Individualized Positive End-Expiratory Pressure (PEEP) Recommendation: The model will recommend an optimal PEEP level derived from the intersection of lung collapse and overdistension curves during a standardized decremental PEEP trial. Accuracy will be quantified by Root Mean Square Error (RMSE) in cmH₂O, with a predefined acceptable threshold of ≤ 2.0 cmH₂O.

Occult Patient-Ventilator Asynchrony Detection: The model will detect asynchrony events (including pendelluft, breath-stacking, and ineffective triggering) in real-time, validated against blinded expert review of synchronized video, EIT, and waveform data. Performance will be assessed by AUC-ROC, with a predefined threshold of AUC-ROC ≥ 0.85.

Ventilation/Perfusion (V/Q) Ratio Prediction: The model will continuously estimate the regional V/Q ratio, with reference values obtained from the multiple inert gas elimination technique (MIGET) or a validated surrogate. Performance will be measured by Pearson Correlation Coefficient (PCC), with a predefined acceptable threshold of PCC ≥ 0.70.

Non-invasive PaO₂/FiO₂ Ratio Estimation: The model will predict the PaO₂/FiO₂ ratio without invasive ABG sampling, compared against simultaneous ABG measurements. Accuracy will be evaluated by RMSE (dimensionless), with a predefined acceptable threshold of ≤ 20 units. Bland-Altman limits of agreement and mean bias will be reported as descriptive secondary analyses.

Statistical Analysis Plan:

For the retrospective phase, model training will be conducted using PyTorch with an Adam optimizer and a learning rate of 1e-4. Early stopping with a patience of 10 epochs will be applied to prevent overfitting. For the prospective validation phase, all five primary outcomes will be assessed using a per-patient analysis. AUC-ROC will be calculated with 95% confidence intervals using DeLong's method; RMSE will be reported with bootstrap 95% confidence intervals (1000 resamples); PCC will be computed with 95% confidence intervals using Fisher's z-transformation. Subgroup analyses will be performed according to ARDS severity (moderate vs. severe), age group (≥ 60 vs. < 60 years), and primary etiology (pulmonary vs. extrapulmonary). All statistical tests will be two-sided with a significance level of α = 0.05. Missing data will be handled via multiple imputation by chained equations (MICE) if the proportion of missingness is < 20%; otherwise, complete-case analysis will be performed.

Sample Size Justification: The retrospective cohort of 1000 patients is derived from the available data from 2020 to 2025 at the three participating centers and is considered adequate for deep learning model training based on prior similar EIT-AI studies. The prospective cohort of 200 patients was calculated to provide > 80% power to detect a moderate effect size (Cohen's d = 0.5) for the primary outcomes, assuming a two-sided α of 0.05 and a 15% dropout rate. This sample size also ensures sufficient event counts (≥ 10 events per predictor variable) for the Cox proportional hazards models used in secondary mortality analyses.

Study Governance: The study is sponsored by Ruijin Hospital, with collaborating centers including Peking Union Medical College Hospital and Fujian Provincial Hospital. An independent Data Monitoring Committee (DMC) will oversee patient safety and interim analyses. The study protocol has been approved by the Ruijin Hospital Ethics Committee (Approval No. 202647), and all patients or their legal surrogates will provide written informed consent prior to enrollment in the prospective phase.