Elsevier

Computers in Biology and Medicine

Volume 97, 1 June 2018, Pages 137-144
Computers in Biology and Medicine

Replicating human expertise of mechanical ventilation waveform analysis in detecting patient-ventilator cycling asynchrony using machine learning

https://doi.org/10.1016/j.compbiomed.2018.04.016Get rights and content

Highlights

  • A mismatch between mechanical ventilator delivery and patient demand is referred to as patient-ventilator asynchrony.

  • Addressing patient-ventilator asynchrony requires a reliable framework for continuously monitoring the patient.

  • Machine learning may be used to replicate human expertise of mechanical ventilation waveform analysis.

Abstract

Background

Acute respiratory failure is one of the most common problems encountered in intensive care units (ICU) and mechanical ventilation is the mainstay of supportive therapy for such patients. A mismatch between ventilator delivery and patient demand is referred to as patient-ventilator asynchrony (PVA). An important hurdle in addressing PVA is the lack of a reliable framework for continuously and automatically monitoring the patient and detecting various types of PVA.

Methods

— The problem of replicating human expertise of waveform analysis for detecting cycling asynchrony (i.e., delayed termination, premature termination, or none) was investigated in a pilot study involving 11 patients in the ICU under invasive mechanical ventilation. A machine learning framework is used to detect cycling asynchrony based on waveform analysis.

Results

— A panel of five experts with experience in PVA evaluated a total of 1377 breath cycles from 11 mechanically ventilated critical care patients. The majority vote was used to label each breath cycle according to cycling asynchrony type. The proposed framework accurately detected the presence or absence of cycling asynchrony with sensitivity (specificity) of 89% (99%), 94% (98%), and 97% (93%) for delayed termination, premature termination, and no cycling asynchrony, respectively. The system showed strong agreement with human experts as reflected by the kappa coefficients of 0.90, 0.91, and 0.90 for delayed termination, premature termination, and no cycling asynchrony, respectively.

Conclusions

— The pilot study establishes the feasibility of using a machine learning framework to provide waveform analysis equivalent to an expert human.

Introduction

Clinical decision support systems are taking an ever-increasing role in the practice of healthcare. While providing clinical decision support using structured information has been widely investigated in the literature and is being translated to clinical practice [1], the next frontier is to address challenges in providing clinical decision support based on unstructured data (e.g., images and waveforms). Recent advances in machine learning have provided the opportunity to assist clinicians analyze unstructured data such as free text [2] and images [3]. One area that would benefit from automated analysis of waveforms is respiratory management for patients in the intensive care unit (ICU).

Acute respiratory failure due to infection, trauma, and major surgery is one of the most common problems encountered in the ICU and mechanical ventilation is the mainstay of supportive therapy for such patients. Patient-ventilator interaction is a critical component of the mechanical ventilation process. Specifically, the ventilator has to respond to a patient's respiratory demand including triggering (i.e., initiation of inspiration) and cycling of breaths (i.e., transition from inspiration to expiration). A mismatch between ventilator delivery and patient demand is referred to as patient-ventilator asynchrony (PVA) [4].

Mechanical ventilation may be a distressing and noxious, albeit necessary, procedure for the critically ill patient. This is often manifested by PVA, which in extreme cases is referred to as “fighting the ventilator.” The incidence of PVA is high, with estimates ranging from 12 to 43%, and has been associated with failure to wean from ventilation, longer duration of ventilation, and longer length of stay in critical care units [[5], [6], [7]]. While the association between PVA and poor patient outcome has been recognized, causality has not been established. However, it seems self-evident that the clinical provider should attempt to reduce PVA due to the ethical imperative of reducing patient distress.

The most common approach to correcting patient-ventilator asynchrony is to increase sedation. The authors in Ref. [8] observed that to address severe breath-stacking, clinical staff increase sedation and/or analgesia in half the cases. However, increased sedation is associated with increased duration of ventilation, increased length of stay, increased incidence of delirium, and increased mortality. Specifically, based on the results of a randomized controlled trial, reduction in sedation use resulted in a reduction of mechanical ventilation and ICU length of stay [9].

Larger multi-center studies confirmed that daily interruption of sedation in conjunction with spontaneous breathing trials reduced length of stay and significantly reduced mortality [10]. In addition, a randomized clinical trial indicated that interruption of sedation in conjunction with physical and occupational therapy reduces the duration of delirium [11]. See Ref. [12] for a comprehensive discussion and review of the literature related to ICU sedation and delirium.

An alternative approach to reducing PVA is to adjust the ventilator mode or settings to match patient demand and respiratory cycles. The utility of this approach has been demonstrated in Ref. [8]. Specifically, in this study the authors observed that changes to ventilation setting was a more effective intervention for breath stacking asynchrony as compared to increase in sedation/analgesia.

While there is variability in nomenclature, PVA may be generally classified as trigger asynchrony, flow asynchrony, and cycling asynchrony [13]. In some cases, the recognition of PVA is obvious; however, in other cases both recognition and classification may be subtle, and in the clinical setting, recognition of asynchrony may be delayed. More accurate methods to detect asynchrony involves the addition of invasive components to measure esophageal pressure or diaphragm electrical activity [4]. The additional complexity has prohibited their widespread adoption, and hence, there is a need for a reliable non-invasive method to detect asynchrony.

Most prior studies of PVA have been conducted with off-line post hoc analysis of pressure and flow versus time recordings. Clinicians with knowledge of asynchrony rely on waveforms to detect subtle cases of PVA, which generally cannot be detected by only observing the patient without considering the ventilator waveforms. However, very few institutions can support one-on-one respiratory therapist deployment, and often nursing staff are not familiar with the analysis of pressure and flow versus time recordings needed to recognize asynchrony in the absence of frank physical manifestations of a patient “fighting the ventilator.”

If PVA can be reliably detected via a non-invasive and automated algorithm, then this information can be used within a clinical decision support system for respiration management. This system can show detected asynchrony events and their trends, and assist respiratory therapists in timely detection of asynchrony. It is expected that timely detection of asynchrony and its type will assist clinicians to identify the source of asynchrony and address it through adjustments of the ventilator settings and mode before increasing the patient's sedation-analgesia.

Given these limitations for real-time monitoring by dedicated clinical staff, in this paper we investigate the feasibility of an automated waveform analysis algorithm that can replicate human expertise in detecting patient-ventilator cycling asynchrony. Our hypothesis is that the expertise of a human expert (or a panel of human experts) in analyzing mechanical ventilation waveforms can be replicated using a machine learning algorithm.

Rule-based as well as statistical asynchrony detection techniques have been developed for detecting ineffective triggering and double triggering [[14], [15], [16], [17]]. However, the feasibility of machine learning-based systems to detect cycling asynchrony has not been investigated. Discussions on cycling asynchrony are generally qualitative and involve multiple features that may be found on pressure and flow versus time recordings. Hence, detection of cycling asynchrony provides an opportunity to “capture” human expertise that is presented qualitatively in the clinical literature. A summary of the proposed approach to replicate human expertise in detecting patient-ventilator cycling asynchrony using machine learning and evaluating its performance is shown in Fig. 1.

Section snippets

Subjects

We analyzed mechanical ventilation waveform data from 11 patients admitted to the intensive care unit at the Northeast Georgia Medical Center, Gainesville, GA. This was a retrospective study, where collected waveform data did not include any patient identifiers, and hence, was exempt from IRB review. Our inclusion criteria for collection of data was that the patients were undergoing invasive ventilation in the pressure controlled-volume guaranteed (PCV-VG) mode of the GE Engstrom Carestation

Results

A breakdown of the classification provided by the machine learning classifier compared to human expert labels is given by the confusion matrix in Table 2. Of the 1204 breath cycles with labels determined by majority vote of the human expert reviewers, 802 breath cycles (67%) were majority-voted to have neither of the cycling asynchronies considered, 232 (19%) were voted as delayed termination, and 170 (14%) were voted as premature termination. The performance and the agreement between the

Discussion

Computer algorithms to detect patient-ventilator asynchrony have been mainly focused on the detection of ineffective efforts based on a series of features extracted from the flow and pressure waveforms [[14], [15], [16], [17]]. Specifically, in Ref. [15] the ability of the proposed algorithm to detect ineffective efforts was compared to assessments provided by a panel of five human experts. Furthermore, the agreement of the algorithm to detect ineffective triggering was compared to a detection

Ethical Approval

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. For this type of study formal consent is not required.

Summary

Acute respiratory failure due to infection, trauma, and major surgery is one of the most common problems encountered in the ICU and mechanical ventilation is the mainstay of supportive therapy for such patients. Patient-ventilator interaction is a critical component of the mechanical ventilation process. Specifically, the ventilator has to respond to a patient's respiratory demand including triggering (i.e., initiation of inspiration) and cycling of breaths (i.e., transition from inspiration to

Funding

This work was supported by the National Science Foundation grant IIP-1456404 awarded to Autonomous Healthcare, Inc.

Conflict of Interest

B.G. has stock ownership in Autonomous Healthcare Inc. T.S.P. has stock options in Autonomous Healthcare. B.G. and T.S.P. are inventors on a provisional patent application (assigned to Autonomous Healthcare) for a patient-ventilator asynchrony detection system. W.M.H. has stock ownership in Autonomous Healthcare. J.M.B. has stock options in Autonomous Healthcare and serves as its Chief Medical Officer. A.C., J.M., L.P. have received support from Autonomous Healthcare. A.C. is a part-time

Acknowledgements

The assistance of Candace Cox, Joseph Briggs, and Leonard T. Barrett in data collection is acknowledged.

Behnood Gholami received the B.Sc. degree in mechanical engineering from the University of Tehran, Tehran, Iran, in 2003, the M.A.Sc. degree in mechanical engineering from Concordia University, Montreal, Canada, in 2005, and the M.S. degrees in mathematics and aerospace engineering, and the Ph.D. degree in aerospace engineering from the Georgia Institute of Technology (Georgia Tech), Atlanta, GA, in 2009 and 2010, respectively. He was a Postdoctoral Fellow at the School of Electrical and

References (27)

  • J.P. Kress et al.

    Daily interruption of sedative infusions in critically ill patients undergoing mechanical ventilation

    N. Engl. J. Med.

    (2000)
  • M.C. Reade et al.

    Sedation and delirium in the intensive care unit

    N. Engl. J. Med.

    (2014)
  • M. De Wit

    Monitoring of patient-ventilator interaction at the bedside

    Respir. Care

    (2011)
  • Cited by (54)

    • Artificial intelligence in critical care

      2023, Artificial Intelligence in Clinical Practice: How AI Technologies Impact Medical Research and Clinics
    • Attention-based convolutional long short-term memory neural network for detection of patient-ventilator asynchrony from mechanical ventilation

      2022, Biomedical Signal Processing and Control
      Citation Excerpt :

      In addition, the choice of the classifier will also affect the performance of the model. The application of classical classifiers such as Random Forests, Adaptive Boosting, and Support Vector Machines to classify PVAs has been discussed in the literature [17–19]. We compared our model with two machine learning models.

    View all citing articles on Scopus

    Behnood Gholami received the B.Sc. degree in mechanical engineering from the University of Tehran, Tehran, Iran, in 2003, the M.A.Sc. degree in mechanical engineering from Concordia University, Montreal, Canada, in 2005, and the M.S. degrees in mathematics and aerospace engineering, and the Ph.D. degree in aerospace engineering from the Georgia Institute of Technology (Georgia Tech), Atlanta, GA, in 2009 and 2010, respectively. He was a Postdoctoral Fellow at the School of Electrical and Computer Engineering, Georgia Tech, from 2010 to 2011 and the Brigham and Women's Hospital, Harvard Medical School from 2011 to 2012. He is currently the Cofounder and CEO of Autonomous Healthcare, Inc., Hoboken, NJ. His current research interests include machine learning and application of dynamical systems and control theory to biomedicine.

    Timothy S. Phan received the B.S. degree in Electrical and Computer Engineering, B.S. degree in Biomedical Engineering, and M.S. degree in Electrical Engineering from Rutgers University, New Brunswick, NJ. He is currently the Director of Engineering at Autonomous Healthcare, Inc. His research interests include cardiovascular system dynamics, mathematical modeling of physiological systems, and noninvasive phenotyping of cardiopulmonary-vascular system coupling with machine learning and medical device design.

    Wassim M. Haddad received the B.S., M.S., and Ph.D. degrees in mechanical engineering from Florida Tech in 1983, 1984, and 1987. Since 1994 he has been with the School of Aerospace Engineering at Georgia Tech, where he holds the rank of Professor, the David Lewis Chair in Dynamical Systems and Control, and Chair of the Flight Mechanics and Control Discipline. He also holds a joint Professor appointment with the School of Electrical and Computer Engineering at Georgia Tech. Dr. Haddad has made numerous contributions to the development of nonlinear control theory and its application to aerospace, electrical, and biomedical engineering. His transdisciplinary research in systems and control is documented in over 630 archival journal and conference publications, and 7 books in the areas of science, mathematics, medicine, and engineering. Dr. Haddad is an NSF Presidential Faculty Fellow; a member of the Academy of Nonlinear Sciences; an IEEE Fellow; and the recipient of the 2014 AIAA Pendray Aerospace Literature Award.

    Andrew Cason received the B.S. degree in political science from the University of Georgia, Athens, GA in 1994. In 1998, he received the associate degree in respiratory therapy from the Gwinnett Technical College, Lawrenceville, GA. He has been a respiratory therapist since 1998 and has 16 years of critical care experience. His research interests include optimization of patient-ventilator interaction and its impact on decreasing ICU length of stay, morbidity and mortality.

    Jerry Mullis received the associate degree in respiratory therapy from the Heart of Georgia Technical College in 2006. He has been a respiratory therapist since 2006 and has over 11 years of critical care experience. Currently, he serves as the dayshift respiratory supervisor at the Northeast Georgia Medical Center, Gainesville, GA. His research interests include optimization of patient-ventilator interaction, airway pressure release ventilation (APRV) for trauma patients, and extracorporeal membrane oxygenation (ECMO) criteria and their correlation with outcome.

    Levi Price attended the North Georgia Technical College in Clarkesville, GA and became a Nationally Registered Emergency Medical Technician in 2003. He then joined the Northeast Georgia Medical Center, Gainesville, GA as an emergency medical technician in 2004. After developing an interest in respiratory therapy, he received the associate degree in respiratory therapy from Gwinnett Technical College, Lawrenceville, GA in 2011. He currently serves as a critical care respiratory therapist in the Trauma ICU at the Northeast Georgia Medical Center. His research interests include optimization of patient-ventilator interaction and improving patient compliance while being mechanically ventilated.

    James M. Bailey received his B.S. degree from Davidson College in 1969, a Ph.D. in Chemistry (Physical) from the University of North Carolina at Chapel Hill in 1973, and the M.D. degree from Southern Illinois University School of Medicine in 1982. After receiving his M.D. degree he completed a residency in anesthesiology and then a fellowship in cardiac anesthesiology at the Emory University School of Medicine affiliated hospitals. From 1986 to 2002 he was an Assistant Professor of Anesthesiology and then Associate Professor of Anesthesiology at Emory, where he also served as director of the critical care service. He is currently Medical Director of Critical Care Medicine at the Northeast Georgia Physicians Group.

    View full text