2026 Critical Care Congress Skip to main content
2026 Critical Care Congress Main banner
Final Program


Advance Program
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Session Icons
Thought Leader Session
Thought Leader Session
Guidelines Session
Guidelines Session
Aging Adult Track
Aging Adult Track
Additional Registration Fee
Additional Registration Fee

Research Snapshot Theater: Pulmonary, Adult VI

Breath Cycle Segmentation of Invasive Mechanical Ventilator Waveforms Using Deep Learning

Monday, March 23, 2026
10:15 AM - 11:15 AM Central Time
Location: Connections Central - RST 13
  • PG

    Preeti Gupta, MD, MPH

    Disclosure(s): No relevant financial relationship(s) to disclose.

  • AN

    Aditya Nemani, MS

    Researcher
    University of California, San Diego

    Disclosure information not submitted.

  • AP

    Alex K. Pearce, MD

    Associate Physician Diplomate
    University of California, San Diego

    Disclosure information not submitted.

  • SN

    Shamim Nemati, PhD

    Associate Professor
    University of California San Diego, School of Medicine

    Disclosure information not submitted.

  • Vd

    Virginia R. de Sa, PhD

    Professor
    University of California, San Diego

    Disclosure information not submitted.

  • AM

    Atul Malhotra, MD

    Professor
    University of California San Diego Thornton Hospital

    Disclosure information not submitted.

  • JA

    Jason Adams, MD, MS

    Professor
    University of California Davis Medical Center

    Disclosure information not submitted.

Introduction: Patient-ventilator asynchrony (PVA) arises when ventilator breath delivery is misaligned with a patient’s respiratory demand. PVA has been linked to adverse outcomes but remains understudied in part from a paucity of tools to detect inspiratory and expiratory phases within ventilator waveform data that often lack these labels. Existing rule-based methods may perform poorly in noisy, real-world data, limiting accurate classification of PVA. We hypothesized that a deep learning model for breath segmentation would outperform heuristic methods.


Methods: This was a secondary analysis of ventilator waveforms enriched for PVA, collected at UC Davis Health (Adams et al., 2017). Pressure and flow signals were sampled at 50 Hz (Puritan Bennett 840 ventilator). Inspiratory onset labels were ventilator generated, and a critical care physician manually labeled expiratory onsets. Data were split by patient into training (80%), validation (10%), and test (10%) sets. A 1-D CNN with residual connections (U-Net architecture) was applied to 3.5-second sliding windows to classify cycling events using a dual-head classifier and custom loss function. Due to class imbalance, we used F1 score, sensitivity (Se) and positive predictive value (PPV) to compare model performance to two heuristic methods: 1) zero-crossing of the flow signal, and 2) backtracking from a threshold along the flow derivative to detect slope changes (Blanch et al, 2015). A prediction within 0.1 seconds of the label was deemed a true positive.


Results: The dataset had 9,719 breaths from 35 patients with 1,199 breaths from 4 patients as the test set (380 (32%) were PVA). For inspiration onset detection, the model’s F1 score was 0.99 (Se 0.98, PPV 0.99), outperforming zero-crossing (F1 0.68, Se 0.88, PPV 0.55) and slope-based (F1 0.79, Se 0.91, PPV 0.70) methods. For expiratory onsets, the model (F1 0.98, Se 0.98, PPV 0.99) outperformed both heuristic methods (F1 0.59, Se 0.77, PPV 0.48 and F1 0.71, Se 0.71, PPV 0.71, respectively).


Conclusions: Our deep learning model substantially outperformed heuristic methods for detecting inspiratory and expiratory onsets in data with a high PVA prevalence. Improved breath segmentation may enable more accurate assessments of PVA, enhancing research into the effects of PVA on patient outcomes.