Self-Supervised Representation Learning from Ventilator Waveforms: A Pilot Feasibility Study
Abstract
Rationale: Modern ventilators generate high-frequency pressure and flow waveforms that contain respiratory mechanics, patient respiratory effort, and patient–ventilator interactions. However, these waveform data remain underutilized for representation learning. Objective: We conducted a pilot feasibility study to evaluate whether self-supervised representation learning can extract clinically relevant respiratory information from ventilator waveforms. Methods: We developed a self-supervised representation learning framework for mechanical ventilation waveforms using airway pressure and flow signals. The model was pretrained on unlabeled waveform segments over 255,355 breaths spanning 197.1 hours of mechanical ventilation across 115 patients and learned latent respiratory representations. Frozen embeddings were subsequently evaluated on held-out patients using downstream tasks, including ventilator mode classification, prediction of respiratory physiological variables, and reconstruction of waveforms. Measurements and Main results: Pretrained embeddings enabled recovery of ventilator modes, achieving a bootstrap-derived median balanced accuracy of 0.791 (95% CI: 0.715, 0.858) and macro-F1 score of 0.765 (95% CI: 0.682, 0.839). The learned representations also supported the prediction of respiratory physiological variables, including respiratory rate and positive end-expiratory pressure. Reconstruction of unseen waveforms achieved a median Pearson correlation of 0.801 (95% CI: 0.727, 0.870) for pressure waveforms and 0.783 (95% CI: 0.709, 0.838) for flow waveforms between the reconstructed and original waveforms. Conclusions: This pilot study demonstrates the feasibility of self-supervised representation learning from ventilator waveforms. Our framework introduces ventilator waveforms as a potentially useful data modality in critical care and establishes a foundation for future investigations of patient–ventilator asynchrony, respiratory state monitoring, and optimization of liberation from mechanical ventilation.
Related
- JEPA-Mamba: a self-supervised joint-embedding architecture with bidirectional Mamba for ECG representation learning
- Enhancing self-supervised representation learning through lightweight learnable data augmentation
- Dual-domain self-supervised feature alignment via spectral–spatial representation learning for deepfake anomaly detection
- Semi-supervised clustering with knowledge-guided representation learning in cryo-electron tomography
- Deep learning–based CT lesion analysis and radiological burden stratification in clinically diagnosed ventilator-associated pneumonia
BibTeX
@misc{kobara2026selfsupervised,
title = {{Self-Supervised Representation Learning from Ventilator Waveforms: A Pilot Feasibility Study}},
author = {Seibi Kobara and Md Hassanuzzaman and Vijay Krishnamoorthy and Tetsu Onuma and Mihai V. Podgoreanu and Krista L. Haines and Suresh M. Agarwal and Rishikesan Kamaleswaran},
howpublished = {Research Square},
year = {2026},
month = oct,
doi = {10.21203/rs.3.rs-11325296/v1},
url = {https://doi.org/10.21203/rs.3.rs-11325296/v1},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Self-Supervised Representation Learning from Ventilator Waveforms: A Pilot Feasibility Study
AU - Seibi Kobara
AU - Md Hassanuzzaman
AU - Vijay Krishnamoorthy
AU - Tetsu Onuma
AU - Mihai V. Podgoreanu
AU - Krista L. Haines
AU - Suresh M. Agarwal
AU - Rishikesan Kamaleswaran
PY - 2026
DA - 2026/10/14
PB - Research Square
DO - 10.21203/rs.3.rs-11325296/v1
UR - https://doi.org/10.21203/rs.3.rs-11325296/v1
AB - Rationale: Modern ventilators generate high-frequency pressure and flow waveforms that contain respiratory mechanics, patient respiratory effort, and patient–ventilator interactions. However, these waveform data remain underutilized for representation learning. Objective: We conducted a pilot feasibility study to evaluate whether self-supervised representation learning can extract clinically relevant respiratory information from ventilator waveforms. Methods: We developed a self-supervised representation learning framework for mechanical ventilation waveforms using airway pressure and flow signals. The model was pretrained on unlabeled waveform segments over 255,355 breaths spanning 197.1 hours of mechanical ventilation across 115 patients and learned latent respiratory representations. Frozen embeddings were subsequently evaluated on held-out patients using downstream tasks, including ventilator mode classification, prediction of respiratory physiological variables, and reconstruction of waveforms. Measurements and Main results: Pretrained embeddings enabled recovery of ventilator modes, achieving a bootstrap-derived median balanced accuracy of 0.791 (95% CI: 0.715, 0.858) and macro-F1 score of 0.765 (95% CI: 0.682, 0.839). The learned representations also supported the prediction of respiratory physiological variables, including respiratory rate and positive end-expiratory pressure. Reconstruction of unseen waveforms achieved a median Pearson correlation of 0.801 (95% CI: 0.727, 0.870) for pressure waveforms and 0.783 (95% CI: 0.709, 0.838) for flow waveforms between the reconstructed and original waveforms. Conclusions: This pilot study demonstrates the feasibility of self-supervised representation learning from ventilator waveforms. Our framework introduces ventilator waveforms as a potentially useful data modality in critical care and establishes a foundation for future investigations of patient–ventilator asynchrony, respiratory state monitoring, and optimization of liberation from mechanical ventilation.
SN - 2693-5015
ER -