Novel Predictive Model for Pediatric Bag Mask Ventilation Breath Quality Using Capnography Waveforms

Tuesday, February 25, 2025

8:55am – 9:55am Eastern Time

Location: Connections Central Theater 1

Elizabeth Hyde, MD, MSE (she/her/hers)

Children's Hospital of Philadelphia

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

Sanjiv D. Mehta, MD, MBE (he/him/his)

Children's Hospital of Philadelphia
University of Pennsylvania

Disclosure information not submitted.
RT

Rich Tsui, MS, PhD, FAMIA

Disclosure information not submitted.
LS

Luiz Eduardo V Silva, PhD, MS

Research Associate Scientist II
Children's Hospital of Philadelphia

Disclosure information not submitted.
Vinay Nadkarni, MD, MS, FCCM (he/him/his)

Dr.
University of Pennsylvania Perelman School of Medicine, CHOP

Disclosure information not submitted.
NN

Natalie Napolitano, PhD, MPH, RRT (she/her/hers)

The Children's Hospital of Philadelphia

Disclosure(s): Actuated Medical: Clinical Trial Investigator (Ongoing); Drager: Grant/Research Support (Ongoing), Scientific Advisory Board (Ongoing); Timpel: Grant/Research Support (Ongoing); VERO-Biotech: Consultant (Ongoing)
Lindsay N. Shepard, MD (she/her/hers)

Physician-Scientist Interim
Childrens Hospital of Philadelphia

Disclosure information not submitted.
BB

Benjamin B. Bruins, MD

Associate Professor of Anesthesiology, Critical Care, and Pediatrics
The Children's Hospital of Philadelphia

Disclosure information not submitted.
Matthew Kirschen, MD, PhD

Assoicate Professor of Anesthesiology, Critical Care, and Pediatrics
Childrens Hospital of Philadelphia

Disclosure information not submitted.
AT

Alexis Topjian, MD, MSCE, FCCM

Director, Neurocritical Care; Professor of Anesthesiology, Critical Care, and Pediatrics
Childrens Hospital of Philadelphia

Disclosure information not submitted.
EL

Elizabeth L. Laverriere McGovern, MD, MPH,

Assistant Professor of Anesthesiology, Critical Care, and Pediatrics
Children's Hospital of Philadelphia

Disclosure information not submitted.
Amanda J. Nickel, MS

Respiratory Research Clinical Expert
The Children's Hospital of Philadelphia

Disclosure information not submitted.
Cheryl Dominick, MA

Clinical Respiratory Specialist
Children's Hospital of Philadelphia

Disclosure information not submitted.
Akira Nishisaki, MD, MSCE (he/him/his)

Division of Critical Care Medicine, Department of Anesthesia and Critical Care
The Children's Hospital of Philadelphia

Disclosure information not submitted.

Disclosure(s):

Elizabeth Hyde, MD, MSE: No relevant financial relationship(s) to disclose.

Natalie Napolitano, PhD, MPH, RRT: Actuated Medical: Clinical Trial Investigator (Ongoing); Drager: Grant/Research Support (Ongoing), Scientific Advisory Board (Ongoing); Timpel: Grant/Research Support (Ongoing); VERO-Biotech: Consultant (Ongoing)

Introduction: Bag mask ventilation (BMV) is an essential ICU skill. Low quality BMV is associated with adverse airway outcomes. Current methods to assess BMV quality (e.g. observed chest rise) are unreliable. Capnography is widely used in intubated patients, but its utility in BMV is uncertain. We aimed to develop an innovative Machine Learning (ML) model to classify BMV breaths as high/low quality using CO2 waveforms among pediatric patients. We hypothesized that a model using CO2 waveforms could classify BMV breaths by quality with an area under the receiver operating curve (AUROC) > 0.75.

Methods: We performed a classification model development and validation study. We used a Respiratory Function Monitor to collect flow and pressure data at a 62.5 Hz sampling rate for 8,446 BMV breaths in 85 critically ill children prior to intubation in an academic ICU. Capnography data was collected simultaneously. Breaths were segmented based on flow data. Low quality breaths met at least 1 of the following criteria: 1) exhaled tidal volume < 4 mL/kg or > 12 mL/kg; 2) peak inspiratory pressure > 25 cmH20; 3) leak > 40%; or 4) presence of upper airway obstruction. Breaths longer than the 97.5th percentile were excluded. We used ML with 4 neural network models (ResNet, LSTM, FC-CNN, Attention) and 5-fold cross validation to identify low-quality breaths using capnography waveforms. Models were trained with a batch size of 128, learning rate of 0.001 - 0.0001, early stopping if loss did not reduce, a maximum of 10 epochs, and a 15% validation split. Performance was evaluated using AUROC, sensitivity, positive predictive value (PPV), and F1 score.

Results: 8,247 out of 8,446 total breaths were included, with 78% meeting low quality criteria. The Attention model performed best in identifying low-quality breaths using CO2 waveform data, with an average AUROC of 0.70 (95% CI, 0.67-0.73). At a threshold of 0.5, average sensitivity, PPV, and F1 score were 0.97 (95% CI, 0.96-0.98), 0.81 (95% CI, 0.79-0.83), and 0.88 (95% CI, 0.85-0.91) respectively.

Conclusions: The model accurately discriminated between high/low quality BMV breaths using capnography waveform data. Future work will optimize model accuracy through improved breath definitions and implementation at the bedside for prospective validation.