Early Detection of Respiratory Compromise to Prevent Harm of the Hospitalized Opioid Treated Patient

Overview

Imagine a hospital or ambulatory surgical work environment where clinicians could look at an electronic respiratory monitoring device and observe the patient's data over time, and be cued by the monitor before the patient exhibits dangerous opioid induced respiratory depression/respiratory compromise. Currently, clinicians use electronic monitoring data for real-time assessment of respiratory status. Alarms set at thresholds alert a clinician when the patient is currently experiencing respiratory compromise. Adverse events secondary to opioid induced respiratory compromise (OIRC) continue to occur in 0.5-4.2% of hospitalized patients receiving opioids for acute pain. Opioids continue to be a staple for acute pain management. In this environment of litigation around adequate pain management and the use of opioids, clinicians need a more sensitive and specific way to determine which patients are at risk of severe respiratory depression when using opioids for acute pain management in the hospital setting. This study proposes to evaluate algorithms preliminarily developed in the computer laboratory. This translational research will compare and test replication of our algorithms in a new sample of patients. Patients' electronic monitor data will be used to further develop our algorithms for identifying patients who exhibit OIRC and predicting OIRC events. Explicitly, we will monitor post-operative patients using pulse oximetry, capnography, minute ventilation, and transcutaneous PCO2 during recovery from anesthesia (in PACU), and on the general care floor for up to 72 hours. This data, along with covariates collected from the electronic medical record and environment will be used in machine learning models to develop our algorithms in an iterative process. Future studies will involve instituting these algorithms into a monitoring interface and testing in simulation and in real-time on patients. Please see AHRQ summary sheets from a submission that occurred earlier this year.

Setting: This research project will be performed at an inner-city hospital in Western New York, Buffalo General Medical Center (BGMC) is typical of an under-resourced facility providing care to a substantial proportion of the indigent, minority, immigrant and medically underserved population of a region. In 2017, 36% of BGMC patients identified themselves as Black or African American. Typical patients face economic, cultural or linguistic barriers to healthcare. By focusing on OIRC at BGMC, this study will help in informing how health disparities may impact the incidence of OIRC. At BGMC in 2017, 11,744 surgical procedures were performed and 2% of general surgical cases experienced an adverse event (code blue). Information on adverse events related to OIRC is not available. Aim 1. After recruiting and performing informed consent pre-operatively, we will monitor post-operative patients using pulse oximetry, capnography, TCpCO2, and minute ventilation during recovery from anesthesia (in PACU), and on the general care floor for up to 72 hours. An observational study of 50 surgical patients will be performed to record electronic respiratory monitoring data as well as patient characteristics. This information will be used for validation and iterative development of prediction models using machine learning techniques. In our preliminary work, we used data that was collected by the research assistants reading the data off the device display. During the proposal proposed study, we will record the data from each electronic device directly on USB memory sticks attached to the device. In our preliminary work, we had data from the PACU stay only. During this study, we will collect data prospectively throughout the hospital stay to further inform changes in respiratory compromise as the patient transitions away from the anesthetic and paralytic agents . On the machine learning side, we will explore long-short term memory networks (LSTM), which have become the state of art machine learning models to deal with sequence and time series data (24), including applications in the healthcare domain (25), including recent work by Co-I Chandola. The justification behind using these models over the support vector machine model used in our preliminary study is that they are able to explicitly model the temporal dependencies in the data, which is expected to provide significant improvements in the predictive performance of the model (26). Aim 2. To further understand factors related to OIRC and to assist in responding to the AHRQ reviewers' comments, we will perform a root cause analysis of all adverse events found in the patients we recruited for Aim 1, as well as all rapid response calls, naloxone deliveries, and code blue calls for 2018 at Buffalo General Medical Center (BGMC ). We will examine each case specifically for nursing assessment and monitoring procedures as well as all patient and environmental factors that may have contributed to the adverse event. The patient safety physician and quality assurance nurses from BGMC will be interviewed to perform root cause analysis of all opioid-related adverse events that have occurred over the past year at the facility. Each event will be broken down by who was involved, what they were doing, what technologies were used, where did the event take place, and what outside factors may have contributed to the event. This information will be used to group the potential causes and the progression toward the adverse event, which will allow for identification of the roles of staff workload and patient monitoring on OIRC occurrence. We have received a letter of support from the medical director of patient safety at BGMC and Kaleida Health chief nursing office for our intended projects.