Automated Detection of Patient Ventilator Asynchrony Using Pes Signal

Automated Detection of Patient Ventilator Asynchrony Using Pes Signal A Feasibility Study Towards a Detection Algorithm

Rationale: Patient-ventilator asynchrony (PVA) in mechanical ventilation is associated with adverse patient outcome such as a prolonged stay in the ICU and even mortality. The prevalence of asynchronies is, however, difficult to quantify. It is common to use only the pressure and flow signal of the ventilator to detect asynchronies. The detection method is often based on definitions. The investigators will use new techniques (esophageal pressure signal and machine learning (ML)) to improve detection and quantification of patient-ventilator asynchronies. The hypothesis is that an algorithm which uses the Pes signal and ML to detect and quantify asynchronies is superior to previous techniques.

Objective: 1. To develop an asynchrony detection algorithm based on pressure, flow and Pes signal using ML. 2. To develop a second algorithm with the same ML technique based on pressure an flow signal only. 3. To compare the performance of these models in comparison with an expert team and with each other.

Study design: The investigators will collect internal data from the ventilator connected to patients on mechanical ventilation (population described below). First, the investigators will, with a dedicated expert team, identify and annotate the asynchronies based on visual inspection of the pressure, flow and Pes signal. Second, the investigators will develop an ML algorithm which will be trained with the annotated data from the visual inspection. Third, the performance of the AI algorithm will be compared with the performance of the expert panel using newly obtained data. Fourth, the performance of the AI algorithm will be compared with the second algorithm which uses the pressure and flow signal only.

Study population: All patients admitted to the adult ICU of the LUMC on mechanical ventilation who are ventilated > 24 hours and are equipped with an esophageal balloon catheter.

Intervention (if applicable): None.

Main study parameters/endpoints: The performance of the detection algorithm.

Study Overview

• Status: Recruiting
• Conditions: Mechanical Ventilation Complication
• Intervention / Treatment: Other: No intervention

Detailed Description

1. INTRODUCTION AND RATIONALE Mechanical ventilation should unload the respiratory muscles, provide adequate gas exchange and should be safe, i.e., harm due to mechanical ventilation should be reduced to a minimum. To achieve this the interaction between the ventilator and the patient is preferentially synchronous. Ventilator settings not being synchronized with patient respiratory drive or activity is a phenomenon known as patient-ventilator asynchrony (PVA). PVA may induce several deleterious effects.1 Studies have shown asynchronies to be associated with patient discomfort, increased work of breathing, prolonged weaning, and in one study, even increased mortality.

   Monitoring PVA however is difficult. Clinicians often have to rely on physical examination of the patient as well as visual inspection of pressure, flow and volume waveforms to identify an asynchrony.6 The sensitivity and positive predictive value of analyzing breath-to-breath waveforms are very low (22% and 32%, respectively).1 Artifacts such as cardiac oscillation may mimic asynchronies, and there are times when clinicians standing at the bedside are unable to distinguish between asynchronies and artifacts with certainty.6 Furthermore, detection of PVA is dependent of bedside examination. This challenge leads to the desire of developing effective automated PVA recognition algorithms.7 Various automated algorithms have been developed, however with a variable performance.1 For a correct analysis of asynchronies, the use of an esophageal balloon catheter, which measures the esophageal pressure (Pes), or a catheter which measures the electrical activity of the diaphragm, is necessary.1 Since the use of Pes catheters, it is possible to describe other forms of PVA, such as reverse triggering, which is an asynchrony in which the ventilator triggers the patient.8 Until recently, however, the esophageal catheter has not been routinely used in daily practice but more as a research tool. Since the introduction of personalized medicine, clinicians have gained interest in esophageal manometry to better titrate care to the unique physiology of a patient.9 There are in the current literature no reports of PAV detection algorithms that use the Pes signal for detection.

   In the LUMC the investigators use the esophageal catheter in all patients admitted with acute respiratory failure and in patients ventilated for more than 48 hours per protocol. The esophageal signal gives the opportunity to detect asynchronies more easily than without. The investigators therefore hypothesize that an algorithm based on the esophageal signal will perform better than an algorithm that only uses other ventilator waveforms.

2. OBJECTIVES 2.1 Primary Objective The primary objective of this study is to develop an asynchrony detection algorithm based on pressure, flow and Pes signals of patient data using ML.

   2.2 Secondary Objectives Secondary objectives are to validate the detection algorithm by comparing its performance with the assessment of the expert panel and to compare its performance with the performance of a second algorithm which is based on pressure- and flow signals only.

3. STUDY DESIGN This study will take place at the Intensive Care Unit of the LUMC. First, internal data of adult ICU patients on mechanical ventilation because of acute respiratory failure or with a ventilation duration of at least 24 hours and that are equipped with an esophageal balloon catheter will be collected from the ventilator. The data of interest include pressure, flow and Pes signals of the ventilation. It is necessary to collect as much data as possible as this is required for the development of the algorithm. A minimum of 50 patients will be included with from each a ventilation recording between 4 and 8 hours, which amounts to 200 - 400 hours of mechanical ventilation recording.

The following labels will be assigned to the data:

• Trigger asynchrony (early (reversed), false, failed)
• Cycle (early, late )
• Double trigger (combination of trigger and cycle problem)

The data will be checked for regions of interest. Regions of interest are regions in which a lot of asynchronies can be visually detected. These regions of interest will be annotated and labeled by three independent experts with a lot of experience in the field. If two of the three experts agree on a asynchrony label than the breath cycle will be included for construction of the model. An equivalent amount of normal breaths will also be used to train the algorithm.

The data will be separated in two datasets. The first set will contain the first half of the patients, i.e. the patients with an odd research number. The second set will contain the second half of the patients (patients with an even research number). Research numbers are allocated consecutively to patients after consent.

The machine learning (ML) technique that will be used is a Convolutional Neural Network which will be developed in cooperation with the Technical University of Delft (Technical Medicine and Dr. D.M.J. Tax). The algorithm will be developed based on this training data and the purpose is that the algorithm learns to generalize from the training set.

The algorithm will be validated on the test data for the detection of PAV. The test data will also be annotated by the expert panel. The algorithm's detection performance will be compared with the annotation of the expert panel on the same data.

Together with the development of the main algorithm, which is based on three signals (flow, pressure and esophageal pressure (Pes)), another algorithm based on two signals (flow and pressure) will be developed with the same AI technique. This second algorithm has most in common with previous described algorithms in the current literature. The performance of the first algorithm will be compared with the second algorithm using the test data (second part of the data) in order to investigate if addition of the Pes signal is superior in the detection of asynchronies.

• Study Type: Observational
• Enrollment (Estimated): 50

Contacts and Locations

• Study Contact: Name: Abraham Schoe, MD, PhD; Phone Number: +31-715265018; Email: a.schoe@lumc.nl

Study Locations

• Netherlands
  • Zuid - Holland
    • Leiden, Zuid - Holland, Netherlands, 2333 ZA
      • Recruiting
      • Leiden University Medical Centre
      • Contact: Abraham Schoe, MD, PhD; Phone Number: +32-715265018; Email: a.schoe@lumc.nl
      • Principal Investigator: Abraham Schoe, MD, PhD

Participation Criteria

Eligibility Criteria

• Ages Eligible for Study: Adult; Older Adult
• Accepts Healthy Volunteers: No

• Sampling Method: Non-Probability Sample

Study Population

This study will recruit as much patients as possible, but at least 50 patients, during the study duration. After every 25 patients the algorithm will be tested for improvement. The study population consists of ICU patients on mechanical ventilation because of acute respiratory failure or with a ventilation duration of at least 24 hours that are equipped with an esophageal balloon catheter. Patients are recruited in the ICU of the LUMC.

Description

Inclusion Criteria:

• admission to the ICU of the LUMC;
• age of 18 years or older;
• intubated and receiving mechanical ventilation because of acute respiratory failure or with a ventilation duration of at least 24 hours; and
• equipped with an esophageal balloon catheter

Exclusion Criteria:

• after recent pneumectomy or lobectomy;
• no informed consent

Study Plan

How is the study designed?

What is the study measuring?

Primary Outcome Measures

Outcome Measure	Measure Description	Time Frame
Performance of detection algorithm	Model evaluation: The first part of the dataset will be used to construct/train the model. The second part of the dataset will be used to evaluate the performance of the model. The labels attained by the experts are considered the ground truth. The labeling of the algorithm will be compared with the labels of the experts to assess the performance of the algorithm. The performance of the primary algorithm will be compared with the performance of the second algorithm, which is based only on pressure and flow signals. The performance of the second algorithm will be assessed as described above. The agreement between the experts will be assessed using Fleiss's kappa, which evaluates the agreement between more than two raters.	8 hours

Collaborators and Investigators

• Sponsor: Leiden University Medical Center
• Investigators: Principal Investigator: Abraham Schoe, MD, PhD, Leiden University Medical Centre

Publications and helpful links

General Publications

• Gilstrap D, MacIntyre N. Patient-ventilator interactions. Implications for clinical management. Am J Respir Crit Care Med. 2013 Nov 1;188(9):1058-68. doi: 10.1164/rccm.201212-2214CI.
• Blanch L, Villagra A, Sales B, Montanya J, Lucangelo U, Lujan M, Garcia-Esquirol O, Chacon E, Estruga A, Oliva JC, Hernandez-Abadia A, Albaiceta GM, Fernandez-Mondejar E, Fernandez R, Lopez-Aguilar J, Villar J, Murias G, Kacmarek RM. Asynchronies during mechanical ventilation are associated with mortality. Intensive Care Med. 2015 Apr;41(4):633-41. doi: 10.1007/s00134-015-3692-6. Epub 2015 Feb 19.
• Shi ZH, Jonkman A, de Vries H, Jansen D, Ottenheijm C, Girbes A, Spoelstra-de Man A, Zhou JX, Brochard L, Heunks L. Expiratory muscle dysfunction in critically ill patients: towards improved understanding. Intensive Care Med. 2019 Aug;45(8):1061-1071. doi: 10.1007/s00134-019-05664-4. Epub 2019 Jun 24.
• Esperanza JA, Sarlabous L, de Haro C, Magrans R, Lopez-Aguilar J, Blanch L. Monitoring Asynchrony During Invasive Mechanical Ventilation. Respir Care. 2020 Jun;65(6):847-869. doi: 10.4187/respcare.07404.
• Doorduin J, Roesthuis LH, Jansen D, van der Hoeven JG, van Hees HWH, Heunks LMA. Respiratory Muscle Effort during Expiration in Successful and Failed Weaning from Mechanical Ventilation. Anesthesiology. 2018 Sep;129(3):490-501. doi: 10.1097/ALN.0000000000002256.
• Rehm GB, Han J, Kuhn BT, Delplanque JP, Anderson NR, Adams JY, Chuah CN. Creation of a Robust and Generalizable Machine Learning Classifier for Patient Ventilator Asynchrony. Methods Inf Med. 2018 Sep;57(4):208-219. doi: 10.3414/ME17-02-0012. Epub 2018 Sep 24.
• Akoumianaki E, Lyazidi A, Rey N, Matamis D, Perez-Martinez N, Giraud R, Mancebo J, Brochard L, Richard JM. Mechanical ventilation-induced reverse-triggered breaths: a frequently unrecognized form of neuromechanical coupling. Chest. 2013 Apr;143(4):927-938. doi: 10.1378/chest.12-1817.

Study record dates

Study Major Dates

• Study Start (Actual): February 1, 2023
• Primary Completion (Estimated): December 31, 2024
• Study Completion (Estimated): April 30, 2025

Study Registration Dates

• First Submitted: March 1, 2023
• First Submitted That Met QC Criteria: December 15, 2023
• First Posted (Actual): January 2, 2024

Study Record Updates

• Last Update Posted (Actual): March 13, 2024
• Last Update Submitted That Met QC Criteria: March 12, 2024
• Last Verified: March 1, 2024

More Information

Terms related to this study

• Keywords: mechanical ventilation; patient-ventilator interaction; detection algorithm; convolutional neural network
• Other Study ID Numbers: 2202-061

Plan for Individual participant data (IPD)

• Plan to Share Individual Participant Data (IPD)?: NO

Drug and device information, study documents

• Studies a U.S. FDA-regulated drug product: No
• Studies a U.S. FDA-regulated device product: No