EMBC Special Issue: Modeling Early Oxygenation Trajectory in PARDS from High-Frequency Mechanical Ventilation Signals...

Area of Science:

• Critical Care Medicine
• Pediatric Pulmonology
• Biomedical Engineering

Background:

• Pediatric acute respiratory distress syndrome (PARDS) is a significant cause of morbidity in pediatric intensive care units (PICUs).
• Predicting oxygenation trajectory is crucial for effective management of PARDS.
• Current methods may not fully leverage the rich data available from mechanical ventilators.

Purpose of the Study:

• To investigate if early high-frequency ventilator waveforms can predict subsequent oxygenation trajectory in PARDS patients.
• To assess the predictive power of engineered statistical features and learned representations from ventilator signals.
• To evaluate different sequence architectures for modeling oxygenation trends.

Main Methods:

• Analysis of breath-by-breath flow waveforms from mechanically ventilated pediatric patients.
• Extraction of engineered statistical features (oxygenation indices, time/frequency domain variables) and learned representations (1D-CNN embeddings).
• Application of RNN, LSTM, GRU, Transformer, and Mamba architectures for prediction tasks (12-hour oxygenation saturation index regression and classification).

Main Results:

• Achieved cross-validated AUROC of 0.819 ± 0.030 for classification, with consistent performance on independent test and validation cohorts (AUROC 0.787-0.832).
• Reported Brier scores between 0.12-0.15, indicating good calibration.
• Best cross-validated RMSE for regression was 2.45 ± 0.81 OSI units, with test/validation RMSEs of 2.60-3.48.

Conclusions:

• Early high-frequency ventilator waveforms contain significant predictive information about short-term oxygenation trajectory in PARDS.
• Feasible to model early oxygenation trends using routinely available ventilator data.
• Findings support further prospective, multi-institutional validation for clinical translation.