Machine Learning Algorithms to Detect Patient-Ventilator Asynchrony (PDF)

**Machine Learning Algorithms to Detect Patient-Ventilator Asynchrony. A Feasibility Study.** **Guillermo Gutierrez, MD, PhD^1^** **Kendrew Wong, MD^2^** **Arun Jose, MD^3^** 1\. Emeritus Professor of Medicine, Anesthesiology and Engineering,\ The George Washington University 2\. Department of...

**Machine Learning Algorithms to Detect Patient-Ventilator Asynchrony. A Feasibility Study.** **Guillermo Gutierrez, MD, PhD^1^** **Kendrew Wong, MD^2^** **Arun Jose, MD^3^** 1\. Emeritus Professor of Medicine, Anesthesiology and Engineering,\ The George Washington University 2\. Department of Medicine, NYU Grossman School of Medicine 3\. Department of Medicine, University of Cincinnati **Correspondence:** Guillermo Gutierrez, MD, PhD The George Washington University 700 New Hampshire Ave, NW, Suite 1107 Washington, DC 20037 **ORCID 0000-0001-7754-5816** **Running Head: ML Algorithms to** Detect Patient-Ventilator Asynchrony **Manuscript word count: 3112** **Abstract word count: 250\ \ \ \ ** **ABSTRACT** **Background: Effective ventilatory support requires regular assessments of patient-ventilator interactions. Developing a dependable, automated method for this evaluation is crucial. We explored the feasibility of using machine-learning algorithms to replicate the assessment of breathing patterns by experienced clinicians, based on airway flow and pressure signals.** **Methods: We studied 44 adult patients on invasive mechanical ventilation for various reasons. Airway flow and pressure signals were digitally acquired from the ventilator and recorded as 2.2-minute epochs. Experienced clinicians analyzed and categorized 50,712 epochs, which included roughly 2.6 million breathing cycles. We developed four Random Forest algorithms to: 1) detect asynchronous breathing, 2) classify types of breathing asynchrony, 3) assess the extent of signal disruption, and 4) identify dynamic hyperinflation. The accuracy of these algorithms was evaluated based on their ability to correctly identify epochs, and their clinical reliability was assessed by comparing their predictions to those of clinicians with different levels of experience in asynchrony classification.** **Results: The algorithms achieved accuracies of 91%, 82%, 87%, and 93% in detecting asynchronous breathing, classifying asynchrony types, assessing disruption severity, and identifying dynamic hyperinflation, respectively. Algorithm classification was more consistent with expert clinicians (kappa = 0.46 and 0.59) than with non-experts (kappa = 0.25 and 0.38; p \< 0.05). Longer duration of asynchronous breathing was associated with increased 28-day mortality (p = 0.015).** **Conclusions: Machine-learning algorithms can effectively emulate expert clinician assessments of breathing patterns in mechanically ventilated patients. Enhancements in algorithm accuracy could be achieved through larger databases and further advancements in artificial intelligence.** **Keywords.** Patient ventilator asynchrony, ICU monitoring, artificial intelligence, respiratory rate variability. **Take-home message:** Artificial intelligence machine learning algorithms can be effectively trained to replicate the expertise of experienced clinicians in evaluating breathing patterns during mechanical ventilation, including the detection of patient-ventilator asynchrony. **Statements and Declarations.** **Ethical Approval and Consent to participate:** The database used in the present study was collected with approval from The George Washington University Institutional Review Board (IRB No. 101228) in compliance with the 1964 Helsinki Declaration. Informed consent was obtained from the patients or their designated surrogates for participation in these studies. **Availability of supporting data:** The dataset analyzed during the current study can be found in the Supplementary Material Section. Access to the database storing the raw data may be granted to qualified researchers upon reasonable request and adequate vetting. **Competing interests:** GG owns US Patent 8,573,207 B2, titled \"Method and System to Detect Respiratory Asynchrony.\" However, the method described in this patent is not related to the methodology detailed in the present study. The other authors declare no competing interests. **Authors contributions:** **GG -** Conception and design of the work; development of the machine-learning algorithms, acquisition, analysis, and interpretation of data; drafting the manuscript. All authors agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. **KW -** Acquisition, analysis, and interpretation of data; drafting the manuscript. **AJ** - Acquisition and interpretation of data; drafting the manuscript. **Source of funding:** Self-funded study. **Acknowledgments:** The authors wish to thank the nurses of The George Washington University Intensive Care Unit for their cooperation in the study. We also extend our thanks to Drs. Ana McLean, Lisa Glass, Ali Khalofa, and Nawaf Almeshal for their help in data collection. **Disclaimer:** The content is solely the authors' responsibility and does not necessarily represent the official views of The George Washington University. **\ ** **Introduction** Mechanically ventilated patients often develop asynchronous breathing, defined as a mismatch between the patient's neural drive and the set rhythm of the ventilator \[1\]. The incidence of patient-ventilator asynchrony (PVA) ranges from 25% to 100%, and it has been associated with prolonged ventilatory support \[2\] and increased mortality \[3\]. Despite its clinical significance, PVA identification can be challenging \[4\], as detection relies on the clinician\'s expertise and the intermittent nature of ventilator waveform monitoring. Therefore, it is desirable to develop a method to detect and classify PVA in ventilated patients. Ideally, this method should be autonomous, non-invasive, and capable of providing real-time analysis. Advances in computer software and data storage capacity have made it possible to utilize a subset of artificial intelligence known as machine learning for developing mathematical algorithms that can continuously analyze airway waveforms to detect patient-ventilator asynchrony (PVA). We present a study exploring the feasibility of developing machine learning algorithms that can detect asynchronous breathing patterns, emulating the approach of a clinician analyzing ventilator airway signals. These algorithms were created by applying machine learning techniques to a substantial database of airway signals collected from adult patients on mechanical ventilation. **MATERIALS AND METHODS** This was a prospective, observational study performed at The George Washington University Hospital and approved by The George Washington University IRB (\# 101228). Informed consent to record and analyze deidentified data was obtained from all patients or surrogates prior to data collection. We sequentially enrolled adult patients within 24 hours from the initiation of invasive mechanical ventilation. All patients were intubated via the nasotracheal or orotracheal route and received ventilatory support using Servo\_i or Servo\_s ventilators (Getinge, Solna, Sweden) with various modes of ventilation. Treatment decisions were independent of the study. Data were collected for a maximum of four days or until weaned from the ventilator. Patients treated with serial or continuous use of paralytic agents were excluded from enrollment. We recorded days on mechanical ventilation, ICU and hospital length of stay, and all-cause mortality at 28 days. Demographic data, SAPS II scores \[5\] and hourly use of intravenous propofol were extracted from the patient's chart and recorded in physical notebooks that were destroyed after transferring nonidentifiable details to the database using a study number. We sampled airway pressure and flow signals continuously at 31.25 Hz using a Raspberry Pi 3B microprocessor connected to the ventilator\'s RS232 data port via a null DB9 cable. The data were stored as sequential 2.2-minute epochs, each containing 4,096 samples per signal. Upon completing each epoch's signal sampling, the microprocessor collected other ventilator-related information, including the set mode of ventilation and fraction of inspired oxygen (F~I~O~2~), and calculated breathing data such as peak pressure, respiratory rate, and expired tidal volume. Data were transferred from the microprocessor to a digital computer for analysis once the patient completed the study. Software was written to display data pertaining to any given epoch, showing the airway signals, patient demographics, ventilator settings, and breathing data (Figure 1E of Supplementary Material Section). Selectable graphic radio buttons allowed for epoch classification by *clinicians with expertise in mechanical ventilation who visually evaluated each epoch's* breathing pattern. *Epochs exhibiting fewer than four asynchronies were classified as either \"synchronous\" or \"variable-timing.\" Synchronous epochs were characterized by near-perfect regularity in breath timing, while variable timing epochs displayed observable differences in time between breaths. Epochs with four or more identifiable asynchronies were designated as asynchronous and further subdivided into specific categories: "ineffective effort", "double trigger", "auto trigger", and "cycling" asynchrony. The latter classification included both delayed and premature cycling. Epochs demonstrating two types of asynchronies were classified according to the predominant type. Epochs with three or more asynchrony types were categorized as "complex". Figure 2E* (Supplementary Material Section) *illustrates examples of breathing patterns with corresponding asynchrony classifications.* Epochs were also evaluated for severity based on the extent of airway signal disruption, categorized as mild, moderate, or severe. This assessment was designed to mimic clinical decision-making and was thus subjective, relying on the clinical judgment of the classifier. Epochs where the end-expiratory flow did not return to the zero baseline in four or more breaths were classified as potentially exhibiting dynamic hyperinflation \[6\]. Given the large number of epochs generated during the study, each epoch was classified only once by one assessor. All information related to an epoch was recorded as a row, or instance, in the database. *Machine learning algorithms.* Following best practices for machine learning algorithm development \[7\], we randomly selected 70% of the database instances to create the training dataset for algorithm development. The remaining 30% of the database served as the testing dataset to evaluate the algorithms\' performance. Four Random Forest algorithms were developed (Figure 1). A binary classifier trained to detect the presence or absence of asynchronous breathing (Model 1). A multiclass classifier trained to recognize signal morphology, including synchrony, variable breathing and the five types of asynchronies previously described (Model 2). A multiclass classifier trained to predict epoch severity (Model 3). Lastly, a binary classifier trained to detect dynamic hyperinflation (Model 4). Algorithm assessment metrics included accuracy, defined as the fraction of correct predictions made by the algorithms using the testing dataset; precision (positive predictive value, PPV); sensitivity; specificity; and the F1-score, defined as the harmonic mean of precision and sensitivity. We also evaluated the clinical reliability of Models 2 and 3 by comparing the algorithms predictions to those made by two groups of clinicians with different levels of expertise in identifying asynchronous breathing. The \"Non-expert\" group comprised junior clinicians with limited experience in identifying asynchrony (n = 4), while the \"Expert\" group included attending physicians with self-assessed proficiency in asynchrony identification (n = 3). Members of both groups were tasked with classifying identical datasets consisting of 1,000 epochs randomly selected from the database. We hypothesized that the classifications made by the Expert group would more closely align with the predictions of the Random Forest algorithms. Cohen's kappa statistic \[8\] was used to compare the agreement between the predictions made by each reader and those of the models. The magnitude of agreement was categorized *a priori* \[9\] as poor (0 to 0.19), fair (0.20 to 0.39), moderate (0.40 to 0.59), and substantial (0.60 to 0.79). The Mann-Whitney and Kruskal-Wallis tests were used to assess differences in distributions between groups without assuming a normal distribution. Unless specified otherwise, data are shown as median \[IQR\]. Additional information on epoch scoring and machine learning algorithm development may be found in the online data supplement. **RESULTS** We enrolled 44 patients aged 62.5 \[57.0, 70.8\] years. All participants were medical patients, the majority were African-American, and half were women (Table 1E of Supplementary Material Section). Most patients were intubated due to pulmonary conditions (52.3%), including COPD, pneumonia, and ARDS. Patients remained on ventilatory support for a median of 3 \[2, 5\] days, with ICU and hospital stays of 6 \[3, 11\] days and 15 \[8, 26\] days, respectively. The modes of ventilation used were pressure-regulated volume control (PRVC; 61% of epochs), pressure control (PC; 31%), pressure support (PS; 6%), and volume control (VC; 2%). *Patients were* enrolled within 13 \[6, 17\] hours from the time they were placed on invasive ventilatory support. Patient were monitored for 46 \[24, 70\] hours for a total cohort monitoring time of 2,112 hours. We captured and classified 50,712 epochs, encompassing approximately 2.6 million breathing cycles. As shown in Table 1, variable breathing was the most frequently observed pattern, occurring in 36.5% of epochs, while fully synchronous breathing was present in 21.0%. Asynchronous breathing accounted for 42.5% of epochs. Patients experienced asynchrony for 39 \[20,59\] % of the monitored time, with all patients displaying asynchronies at some point during the time monitored. Among asynchronous epochs, 57.2% were cycling asynchronies, 20.2% were ineffective efforts, 12.9% were double triggering, and 1.2% were auto-triggering. 33.8% of epochs classified as asynchronous displayed more than one morphological type of asynchrony, with 8.5% exhibiting three or more types. The distribution of epoch severity varied. Only 7.4% of epochs classified as variable breathing were rated as severe. In contrast, nearly half of the asynchronous epochs were deemed severe, with auto-triggering and complex asynchronies having the highest percentages of severe classifications. Figure 2 illustrates the number of epochs classified as synchronous, variable-timing, and asynchronous. In this figure, black bars represent the percentage of patients in each group who were treated with intravenous propofol, while grey bars indicate the percentage who received a propofol infusion rate exceeding 30 ml·min·kg⁻¹. A higher percentage of synchronous epochs were from patients treated with intravenous propofol (p \ 30 ml·min·kg⁻¹ (p \