Modeling Early Oxygenation Trajectory in PARDS From High-Frequency Mechanical Ventilation Signals Using Deep Sequence Architectures.

OBJECTIVE

PARDS causes substantial morbidity in PICU. We investigated whether high-frequency ventilator waveforms from first hour of invasive ventilation contain predictive information for oxygenation trajectory.

METHODS

We analyzed mechanically ventilated pediatric patients using engineered statistical features and learned representations from high-frequency ventilator signals, including breath-by-breath flow waveforms. Engineered features included early oxygenation indices and ventilator variables derived from time- and frequency-domain analyses. Learned representations were obtained from 1D-CNN embeddings with PCA-based dimensionality reduction. We evaluated two tasks: 12-hour OSI regression and classification of mild vs moderate-or-higher impairment (OSI). Five sequence models (RNN, LSTM, GRU, Transformer, Mamba) and nine feature configurations were assessed using repeated cross-validation and validation on held-out and temporally separated cohorts.

RESULTS

For classification, the best cross-validated AUROC was 0.819 ± 0.030. Performance remained consistent across independent cohorts, with AUROC values of 0.787-0.823 on the test cohort and 0.808-0.832 on the temporal validation cohort, and Brier scores $\sim$0.12-0.15. For regression, the best cross-validated RMSE was 2.45 ± 0.81. Test RMSE ranged from 2.99 to 3.48 OSI units and validation RMSE from 2.60 to 2.85.

CONCLUSION

High-frequency ventilator waveforms from the first hour of mechanical ventilation contain measurable information about short-horizon oxygenation trajectory in PARDS.

SIGNIFICANCE: These findings support feasibility of modeling early oxygenation trajectory using routinely available ventilator waveform data and motivate prospective multi-institutional validation.