given from the resultant physiologic database of signal process- ... build a database of physiologic signals that .... A custom software system was devel- oped by ...
Physiologic data acquisition system and database for the study of disease dynamics in the intensive care unit* Brahm Goldstein, MD, FCCM; James McNames, PhD; Bruce A. McDonald, PhD; Miles Ellenby, MD; Susanna Lai, BS; Zhiyoung Sun, MS; Donald Krieger, PhD; Robert J. Sclabassi, MD, PhD
Objective: To describe a real-time, continuous physiologic data acquisition system for the study of disease dynamics in the intensive care unit. Design: Descriptive report. Setting: A 16-bed pediatric intensive care unit in a tertiary care children’s hospital. Patients: A total of 170 critically ill or injured pediatric patients. Interventions: None. Main Outcome Measures: None. Results: We describe a computerized data acquisition and analysis system for the study of critical illness and injury from the perspective of complex dynamic systems. Both parametric (1 Hz) and waveform (125–500 Hz) signals are recorded and analyzed. Waveform data include electrocardiogram, respiration, systemic arterial pressure (invasive and noninvasive), central venous pres-
C
urrent bedside monitoring techniques have remained relatively unchanged over the past 20 yrs and do not take advantage of advances in fields such as physiology, computer science, biomedical engineering, and mathematics. Standard bedside monitoring consists of tracking clinical variables, including the continuous noninvasive monitoring of the electrocardiogram, respirations, arterial oxygen saturation, temperature, and endtidal CO2 concentration. Insertion of specialized catheters enables continuous *See also p. 642. From Complex Systems Laboratory, Division of Pediatric Critical Care, Doernbecher Children’s Hospital, Oregon Health Sciences University, Portland, OR (BG, ME, SL); the Department of Electrical and Computer Engineering, Portland State University, Portland, OR (JM); the Departments of Electrical & Computer Engineering (BAM) and Computer Science (ZS), Northeastern University, Boston, MA; and the Laboratory for Computational Neuroscience, Department of Neurological Surgery, University of Pittsburgh Medical Center, Pittsburgh, PA (DK, RJS). Presented, in part, at the 29th Scientific and Educational Symposium of the Society of Critical Care Medicine, Orlando, FL, February 2000. Copyright © 2003 by Lippincott Williams & Wilkins DOI: 10.1097/01.CCM.0000050285.93097.52
Crit Care Med 2003 Vol. 31, No. 2
sure, pulmonary arterial pressure, left and right atrial pressures, intracranial pressure, body temperature, and oxygen saturation. Details of the system components are explained and examples are given from the resultant physiologic database of signal processing algorithms and signal analyses using linear and nonlinear metrics. Conclusions: We have successfully developed a real-time, continuous physiologic data acquisition system that can capture, store, and archive data from pediatric intensive care unit patients for subsequent time series analysis of dynamic changes in physiologic state. The physiologic signal database generated from this system is available for analysis of dynamic changes caused by critical illness and injury. (Crit Care Med 2003; 31:433–441) KEY WORDS: data acquisition; bedside monitoring; disease dynamics; waveform data
invasive monitoring of arterial blood pressure, central venous pressure, pulmonary arterial pressure, right and left atrial pressure, intracranial pressure, and intermittent calculations of stroke volume, systemic and pulmonary vascular resistance, and cardiac output. Arrhythmia detection algorithms are the first advances to take advantage of improved computer power and use real-time R wave and dysrhythmia detection algorithms to identify and detect cardiac dysrhythmias with a sufficient sensitivity and specificity (1). Despite the multitude of physiologic signals available for monitoring, we suggest that a wealth of potentially valuable information that may affect clinical care remains largely an untapped resource. Current methods of clinical monitoring rely on the analysis of discrete 3- to 5-sec time-averaged parametric data that are numerically displayed next to the waveforms on most bedside monitors. Many monitors can also display longer periods of organ system function in a trend mode of parametric data that spans minutes to hours. Previous studies have shown that additional “hidden” information is contained in commonly monitored
physiologic signals that can provide information about interconnections between organ systems, regulatory and feedback relationships, and the overall physiologic state of the organism (2). The paradigm shift that we are studying is that of dynamic (i.e., temporal changes) analysis of physiologic signals and their relationship to organ system interconnections rather than individual organ system function. The study of disease dynamics, or how disease states change with respect to time, is proving key to understanding abnormalities in underlying physiologic control mechanisms. New information regarding the predictability of clinical deterioration, improvements in diagnostic capabilities, and assessment of therapy have been reported in intensive care unit (ICU) diseases and clinical states such as sudden cardiac death after acute myocardial infarction, congestive heart failure, brain injury, sepsis, and the prediction of hypotension during dialysis (2–9). Collection of dynamic time series data in the ICU is the key first step in obtaining a system dynamics oriented perspective of critical illness and injury. Collection of long time series (hours-days) has proven 433
Figure 1. Schematic diagram of the design and communication links of the Pediatric Intensive Care Unit (PICU), the Philips patient data server (PDS), the Complex Systems Laboratory, the Laboratory for Computational Neuroscience, Department of Neurologic Surgery, at the University of Pittsburgh Medical Center (LCN @ UPMC), and the Biomedical Signal Processing Laboratory, Department of Computer and Electrical Engineering, Portland State University (BSP Lab @ PSU). OHSU, Oregon Health Sciences University; HP PC, Hewlett Packard personal computer. Reprinted with permission from Ellenby et al (30).
difficult (3) in the past due to insufficient computational power, inadequate specialized software, and incompatibility between monitoring, data collection, and data analysis systems. Recently, a number of investigators have described their systems for physiologic data acquisition, archive, and analysis (10 –12). We have developed a real-time, continuous physiologic data acquisition system for the study of disease dynamics in the pediatric ICU (PICU) that can capture data from multiple patients simultaneously on a continuous basis. This manuscript describes our physiologic data acquisition system in detail. The system is used in three ICUs, in the operating rooms at the University of Pittsburgh, and in the PICU at Doernbecher Children’s Hospital, Oregon Health and Science University. Because the systems used in the ICUs of both medical centers have the same underlying system architecture, we limit our discussion to the system at Doernbecher Children’s Hospital. We also briefly discuss some of the techniques for signal 434
processing and analysis that we have developed.
METHODS This study was approved by the Institutional Review Boards of Oregon Health and Science University and University of Pittsburgh. Acquisition of Physiologic Signals. We used the computer system in the Complex Systems Laboratory located within in the PICU at Doernbecher Children’s Hospital to build a database of physiologic signals that characterize dynamic disease processes in critically ill and injured children. The system has four main components (Fig. 1): ●
Patient monitors: A total of 16 Philips Merlin (Philips Medical Systems, Eindhoven, The Netherlands) patient monitors (formerly Hewlett Packard/Agilent), one located in each patient room in the PICU. These monitors apply anti-aliasing filters to the raw analog signals, perform analog-to-digital conversion, and sample the continuoustime signals to form discrete-time series. The patient monitors limit the sampling rates to three different values. Variable signals, which are slowly varying signal aver-
●
●
ages such as heart rate, are sampled at 0.98 Hz. All other signals are sampled at either 500 Hz or 125 Hz, depending on the known signal properties. We chose to sample electrocardiogram at 500 Hz to capture the high-frequency QRS component of this signal. The other signals (respiratory and pressure waveforms) do not contain these highfrequency components so we chose 125 Hz to conserve data storage space. Patient data server: The signals acquired by the patient monitors are sent to the patient data server (PDS) via a serial distribution network (Philips). The PDS is a personal computer located in a telecommunications room next to the PICU Data storage workstation: The PDS is connected through a local area network to a data storage workstation (HPUX, Hewlett Packard, Palo Alto, CA). The workstation is used to convert the signals to text files and to archive the files on CD-ROM. The data storage workstation is located in the Complex Systems Laboratory within the PICU.
Analysis workstation. A Hewlett Packard Vectra personal computer (Windows) is used for data analysis and is connected to the HPUX workstation through a local area network. The analysis workstation is Crit Care Med 2003 Vol. 31, No. 2
located in the Complex Systems Laboratory within the PICU. Each bedside Philips Merlin monitor is capable of state-of-the-art monitoring of all physiologic waveforms required in critically ill patients, including electrocardiogram, respiration, systemic arterial pressure (invasive and noninvasive), central venous pressure, pulmonary arterial pressure, left and right atrial pressures, intracranial pressure, body temperature, and oxygen saturation. Parametric data are defined as the time averaged mean data for each waveform displayed in parentheses next to the real-time waveforms on the monitor’s display screen. Parametric data are calculated over a 3to 5-sec moving window for each waveform displayed. Both waveform and variable data are captured and recorded. The Hewlett Packard serial distribution network PDS is a system composed of a Hewlett Packard Vectra personal computer configured with specialized hardware and software that provides a digital programmable interface to the Hewlett Packard Merlin bedside monitors connected to Hewlett Packard’s serial distribution network. Attached to both the serial distribution network and to the Complex Systems Laboratory’s and Laboratory for Computational Neuroscience’s local area network, the PDS effectively provides a gateway allowing local computer systems access to waveform and variable data produced by the Philips Merlin monitors for real-time display or acquisition. For Internet access, standard network communications protocols extend this capability to the Laboratory for Computational Neuroscience, Department of Neurologic Surgery, at the University of Pittsburgh Medical Center, subject to bandwidth limitations of the Internet connections. The PDS system provides the interface that allows third-party software to access Philips Merlin patient data. Using appropriate network controls, a program running on a remote computer system from the Complex Systems Laboratory communicates with the PDS to learn the status of each bedside monitor and dynamically identifies the physiologic data available on each channel of an active monitor. Using this information, the program requests the PDS to capture desired data and forwards it to the remote program for processing. Two types of physiologic data are available to remote programs through the PDS: parametric data that is sampled at a rate of 0.98 Hz and Crit Care Med 2003 Vol. 31, No. 2
Figure 2. Graphical user interface for data acquisition system in the pediatric intensive care unit. Bottom screen shows current data being recorded. Top, bed selection graphical user interface; bottom, subject information graphical user interface.
continuous wave data that is sampled at either 500 Hz (electrocardiogram) or 125 Hz (pressures, arterial saturation, and respiration). Data are transmitted by the PDS to a remote program at a rate of 0.98 Hz; therefore, each transmission consists of 1, 125, or 500 data points. In addition to physiologic data, we also capture alarm text and alert status information. Future upgrades of the bedside monitoring system may allow capture of additional patient identification information.
One of the biggest obstacles that we encountered during the development of this system was a periodic corruption of the continuous wave data. We determined that the corruption was caused by the clinical staff adding or removing a waveform module (e.g., arterial blood pressure, intracranial pressure) in the patient monitor after data acquisition had begun. We have recently developed software to detect when this occurs and automatically restart the data acquisition process. 435
Figure 3. Plot of the intracranial pressure vs. time during two heart beats. This figure also illustrates the time-domain signal metrics that can be used for further studies once the waveform components are detected for each beat. These metrics include the intervals and relative amplitudes of each pair of the PTND components. In this example, the tidal peak is absent. Reprinted with permission from McNames et al (7). A, amplitude; D, dichrotic peak; M, minimum; N, dichrotic notch; P, percussion peak; T, time interval. Reprinted with permission from Aboy et al (34). Table 1. Table listing disease categories, number of patients, average number of recorded physiologic signals, and range of recording times from the Complex Systems Laboratory physiologic signal database (as of 11/15/01) Diagnostic Group
n
Average No. of Signals
Recording Time Range, hrs
Asthma Brain injury Cardiac surgery Diabetic ketoacidosis Sepsis syndromes Spinal fusion Miscellaneousa
6 60 20 6 55 7 16
7 8 10 6 7 7 6
6–162 6–342 24–120 18–36 18–336 48–168 6–168
a Diagnoses include acute respiratory failure (n ⫽ 3), encephalitis (n ⫽ 2), hypertensive crisis (n ⫽ 2), cervical spine injury (n ⫽ 2), overdose (clonidine, n ⫽ 2; amitryptiline, n ⫽ 2), and terminal event (n ⫽ 3).
A custom software system was developed by the Laboratory for Computational Neuroscience at the University of Pittsburgh Medical Center that runs under HPUX and communicates with the PDS system using standard transmission control protocol/Internet protocol communications protocols. The system is currently running in the Complex Systems Laboratory and continuously captures data from any six of the 16 PICU bedside Philips monitors chosen by the investigator. A new graphical user interface has been developed to simplify operator initiation of the system (Fig. 2). The data are stored in an ASCII format. The system continuously captures data on a 24-hr by 7-day schedule. The data are archived on a CD-ROM each 436
weekday, from which it is accessible according to date, unit, and bed number. Up to 7 days of data per subject are kept on-line in the HPUX workstation at any time. The data for each subject/bed is maintained in separate files, which are 2– 6 hrs in duration. The data are archived to CD-ROM discs using a Yamaha CD-R recorder (Yamaha, San Jose, CA) with specialized software for the HPUX environment (Test and Measurement Systems, Denver, CO). Once data are recorded to CD-ROM, they are preprocessed, and the physiologic time series are analyzed using custom software written for MATLAB (Mathworks, http://www.mathworks.com). Our off-line analysis includes time and frequency domain methods and employs linear and nonlinear signal metrics.
Building a PICU Physiologic Signal Library. CD-ROM data are stored in the Complex Systems Laboratory and are available to our colleagues via the Internet. We are in the process of appending the physiologic data with selected clinical data such as medications, mechanical ventilation settings, clinical laboratory test results, temperature and urine output, and severity of illness scores, such as the Physiologic Index of Mortality (13) and Glasgow Coma Scale (14), and outcome scores, such as the Glasgow Outcome Score (15) and Pediatric Overall Performance Category (16). Physiologic Signal Processing Techniques. Preliminary analysis is conducted using several signal visualization techniques. These include nonstationary analysis using moving windows to observe changes in the autocorrelation, power spectral density (i.e., spectrogram), beat morphology (Fig. 3), and wavelet correlation (i.e., scaleogram). All analysis is conducted using our biomedical signal processing toolbox (available June 2002 at http:// bsp.pdx.edu) written in MATLAB. Researchers may also use analysis software available at PhysioToolkit (http://physionet.org), a library of open-source software for physiologic signal processing and analysis. Metrics for Physiologic Signal Analysis. Time series analysis of physiologic waveforms may be accomplished using linear (time and frequency domains), nonlinear, and geometric/morphologic metrics. A detailed discussion of the various signal metrics is beyond the scope of this manuscript. The reader is referred to selected references listed in the bibliography for further information (2, 6, 17– 19).
RESULTS To date, we have successfully recorded real-time continuous time series data from ⬎170 patients from the PICU at Doernbecher Children’s Hospital (Table 1). Data are imported and stored in physiologic databases organized by specific disease states including brain injury, sepsis/septic shock, postoperative cardiothoracic surgery, status asthmaticus, diabetic ketoacidosis, poisoning/overdose, and terminal events. Sample multivariable clinically annotated physiologic time series are detailed in Figure 4. Examples of some of the different metrics used to characterize specific physiologic signals are shown in Figures 5 and 6.
DISCUSSION Relatively few systems exist to record and analyze physiologic systems for clinical research in ICUs (10). Most researchers have used portable monitoring equipment connected via an analog-to-digital Crit Care Med 2003 Vol. 31, No. 2
Figure 4. Clinically annotated multivariable physiologic data from a patient after undergoing cardiac surgery for repair of an endocardial cushion defect over a 120-min time epoch. Eight separate parametric signals are plotted plus a time-sensitive area for textual notations. bpm, beats per minute; ABP, arterial blood pressure; LAP, left atrial pressure; PAP, pulmonary arterial pressure; RESP, respiratory rate; Temp, temperature; RAP, right atrial pressure; SpO2, pulse oximetry.
board in a personal computer (20). Kropyvnytskyy et al. (11) recently reported a system for continuous recording and processing of physiologic data from braininjured patients in a neurosurgical ICU. This system is based on the WFDB software package developed for the MIT-BIH arrhythmia database (21) and collects electrocardiogram, systemic blood pressure, intracranial pressure, and a calculated cerebral perfusion pressure (mean arterial pressure minus intracranial pressure). Data are sampled at 500 Hz and stored in an ASCII file of 10 mins in duration (~2 MB). Tsui et al. (8) initially reported the system for acquiring, modeling, and predicting intracranial pressure in the ICU from which our current system evolved. Other ICU data acquisition systems and databases are listed in Table 2. To our knowledge, the Complex Systems Laboratory is the first system with the capability of remote data acquisition directly from multiple patient monitors simultaneously and the first continuous data acquisition system dedicated to critically ill children. There are a number of relatively large signal databases that contain physiologic signals from adult critical care patients Crit Care Med 2003 Vol. 31, No. 2
such as MGH/MF (Massachusetts General Hospital/Marquette Foundation) waveform (22), MIMIC (Multi-parameter Intelligent Monitoring for Intensive Care) (12), IMPROVE (Improving Control of Patient Status in Critical Care) (23, 24), and IBIS (Improved Monitoring for Brain Dysfunction in Intensive Care and Surgery) (25–27). These databases contain combinations of parametric data, waveform signals, beat labels, annotations, and relevant clinical data and use similar data acquisition systems. The ICU databases discussed above may be viewed on the Web at PhysioBank (www.physionet.org), a large and growing archive of well-characterized digital databases of physiologic signals and related data for use by the biomedical research community (18). PhysioBank currently includes databases of multivariable cardiopulmonary, neural, and other biomedical signals from healthy subjects and patients with a variety of conditions with major public health implications, including sudden cardiac death, congestive heart failure, epilepsy, gait disorders, sleep apnea, and aging. Associated with PhysioBank are PhysioNet, an online forum for discussion and exchange of re-
corded biomedical signals and opensource software and PhysioToolkit, a library of software for physiologic signal processing and analysis, interactive signal display, and new database creation. We plan to submit the data from published studies of our PICU patients to PhysioBank in the near future. Availability of these databases offers the potential to decrease the time required for development of new clinical monitoring methods (10) and for researchers to test new metrics on known databases (18). A 6-hr sample of waveform signals acquired from a patient with a traumatic brain injury is available on the Biomedical Signal Processing Laboratory’s web site (http://bsp.pdx.edu). Other than for dysrhythmia detection, current clinical ICU monitoring relies almost completely on time-averaged mean values of physiologic signals from 3- to 5-sec data epochs. Most monitoring systems also allow the trend of these timeaveraged means to be viewed over periods up to 24 hrs. Both of these modalities presuppose that the physiologic systems being monitored behave in a linear fashion. In linear systems, the response (output) is proportional to the sum of the 437
Figure 5. Top, example of how the intracranial pressure (ICP) signal from a patient with traumatic brain injury was divided into five nonoverlapping segments. We assumed that segments 1– 4 occurred during a compensated physiologic state and that segment 5 occurred during the transition to an uncompensated state immediately before a rapid increase in ICP. Reprinted with permission from McNames et al (7). Bottom, spectrogram of segment shown in Figure 3. The spectrogram is an estimate of the power spectral density vs. time of a nonstationary signal. The frequency range was chosen to show only the cardiac component of the ICP signal. Note that there is a clear decrease in the cardiac component at the leading edge of the ICP elevation (segment 5, light gray), whereas the mean ICP remains unchanged, and there is a clear increase afterward (black). Further analysis from seven subjects found that in 26 of 33 ICP elevations (79%), we could successfully detect a physiologic transition zone before a rapid rise in ICP. This suggests that it may be possible to develop a predictor of rapid and significant elevations in ICP in patients with traumatic brain injury using signal analysis metrics. It is likely that these novel biomedical signal metrics will enable detection of heretofore undetectable physiologic changes in the ICP regulatory system that immediately precede, by a finite time period, an uncompensated physiologic state. Reprinted with permission from McNames et al (7)
responses from all of the stimuli (inputs), and the system can be fully explained and predicted by analyzing its response to each stimulus independently (17). Although this is likely a valid approximation for some physiologic states and ranges, investigators have recently sug438
gested that many, if not all, physiologic systems behave in a nonlinear rather than linear manner. However, it is clear that all systems studied in nature are, in fact, nonlinear. In nonlinear systems, the response is not a sum of the individual responses from each stimulus. Small
changes can result in a disproportionately large and abrupt response. Examples of nonlinear physiologic processes with disproportionate stimulus-response plots include the oxyhemoglobin dissociation curve and the baroreflex. The nonlinear nature of these systems requires a holistic analysis that includes the interconnections and coupling between organ systems. Examples of nonlinear coupled systems include the interaction of pacemaker cells in the heart and neurons in the brain (17) and the autonomic nervous system and the heart in acute brain injury (28). Thus, there is a clear underlying rationale for use of metrics other than time averaged means to describe physiologic systems. This approach may be particularly well suited in the ICU where organ dysfunction and disturbed interorgan communication are the final common pathway from many critical illnesses and injuries (4, 5, 29 –32). There are a number of limitations associated with most physiologic data acquisition systems that affect the resultant database and its uses. Routine patient care activities often interfere with accurate signal acquisition as a result of motion artifact or the addition or removal of sensors. Technical properties of the sensors, the bedside monitors, or the recording system may result in less than ideal data. For example, the arterial blood pressure from an indwelling arterial catheter may have different signal characteristics depending in which artery it is placed (radial vs. femoral), how long it has been in place (catheter “fling” or dampening due to thrombus formation), the underlying disease and physiologic state of the patient (hypotension or hypertension), and the length and gauge of catheter used, the location of the pressure transducer, and how often it is used for blood sampling. Other types of artifact caused by movement or clinical interventions can often be identified by distinguishing features. One of the best features for this purpose is rapid onset. Interventions and movement artifact generally have an onset of less than a second, whereas most physiologic changes occur over a period of at least 5 secs. Segments that contain artifact can also be identified and isolated if they lack known and necessary signal properties such as a dominant cardiac component at the heart rate frequency or if they exceed the physiologically possible signal range (e.g., a blood pressure that is negative or exceeds a realistic physiologic limit). Crit Care Med 2003 Vol. 31, No. 2
Figure 6. Spectrogram of pulmonary arterial pressure signal demonstrating a drop-off in low-frequency signal of ~0.1 Hz up to 200 secs before the onset of pulmonary hypertension. This may signal a change in the physiologic state undetectable by viewing the mean value. Also note oscillations in the pulmonary artery spectrogram at the heart rate and respiratory rate. ICP, intracranial pressure.
Table 2. Continuous intensive care unit (ICU) physiologic data acquisition systems and databases CSL Type of data Type of ICU Type of data Continuous signals Hemodynamic
MGH/MF
MIMIC
IMPROVE
IBIS
Pediatric ICU
Adult ICU, OR, CC
Adult ICU
Adult ICU
Adult ICU and OR
ABP, PAP, CVP LAP, VP, IABP Impedence, EtCO2 ICP None Beat labels, events
ABP, PAP, SpO2
ABP, PAP, CVP
ABP, PAP, CVP
Respiratory Neurologic Intermittent data Annotations
ABP, CVP, PAP RAP, LAP, SpO2 Impedence, EtCO2 ICP, JvO2 1 Hz parametric Limited
AWP EtCO2 None 1 Hz parametric Patient alarms, monitor alarms
AWF, AWP, O2, EtCO2 EEG 0.03 Hz parametric Patient state, nursing actions, artifacts
AWF, AWP, O2, EtCO2, AA EEG, EP, ERP 0.03 Hz parametric Patient state, free text
Database size No. of recordings No. of patients
170 170
250 225 adult, 25 pediatric 1.5
100 100
59 50
48 (ICU), 52 (OR) 100
24–48
24
3
Average length, hrs
48–72
OR, operating room; CC, cardiac catheterization lab; ABP, arterial blood pressure; CVP, central venous pressure; PAP, pulmonary arterial pressure; RAP, right atrial pressure; LAP, lafet atrial pressure; SpO2, oxygen saturation; VP, ventricular pressure; IABP, intra-aortic baloon pressure; AWP, airway pressure; AWF, airway flow; EtCO2, end tidal CO2; AA, anesthetic agent; ICP, intracranial pressure; EEG, electroencephalogram; EP, evoked potentials; ERP, event related potential.
Although there is no perfect technique for isolating artifact, clinical annotations by an observer stationed in the patient’s room is probably the most accurate. The IMPROVE database provides detailed timed annotations regarding patient state, nursing actions, and other patient disturbances by a team of physicians who remained at the bedside throughout the Crit Care Med 2003 Vol. 31, No. 2
~24-hr recording (23, 24). Annotation of the recording during data collection provided for comparison of any changes in the recorded waveforms or measurements against physician observations. The IBIS database is from essentially the same team as IMPROVE but from three hospitals, and it concentrates on neuromonitoring during critical care (25–27).
To address the problem of signal artifact, we plan two approaches. First, we plan to start a clinically annotated database with all activity recorded by video monitor or a researcher stationed at the bedside recording patient activity, nursing care, and therapies similar to that used in the IMPROVE and IBIS databases (23–27). Second, and the 439
W
e have successfully developed a real-
time, continuous physiologic data acquisition system that can capture, store, and archive data from pediatric intensive care unit patients for subsequent time series analysis of dynamic changes in physiologic state.
method currently in place, we use standard signal processing methods to reduce or eliminate the impact of artifact to the extent possible. These methods include notch filters to eliminate periodic equipment noise, standard frequency-selective linear filters, adaptive filters, and nonlinear filters, including wavelet de-noising. Almost counterintuitively, we suggest that the artifact in our database may prove valuable for some purposes because it is representative of the artifact typically encountered in critical care settings. This can help ensure that new methods of patient monitoring are robust to artifact that may be absent in databases created from carefully designed studies. The laboratory and its components detailed in this article were a novel solution to a unique problem at the time of inception. Until that time, we had relied on a computer cart equipped with Hewlett Packard monitors, an analog-to-digital board, a personal computer, and specialized recording software (HRView, Boston Medical Technologies, Boston, MA) that was limited to obtaining data from one subject at a time while at the bedside (20). On building our new PICU in 1998, we used existing Hewlett Packard (now Philips) hardware (Philips Merlin bedside monitors, and the programmable PDS server) with specially designed software from the University of Pittsburgh designed so that the PDS could communicate with the bedside monitors and send the digital information to a HPUX workstation in our laboratory. Currently, Phil440
ips no longer manufactures a programmable PDS, so duplication of our exact laboratory setup is not possible. An alternative method that uses existing technology is to place an RS232 board (Philips M1170A Opt J13) that plugs directly into the Merlin monitor and allows for analog signal output. An upgrade board (Philips M1170A J11) allows for eight channels of analog output per bedside monitor. Each RS232 board would need a cable connected to an analog-to-digital board in a central location where the signals could be acquired and stored using appropriate computer media. It would be possible to develop software for bed selection and automatic acquisition similar to that used in our lab. We hope that by providing a detailed description of the requisite components for a high-fidelity physiologic signal data acquisition laboratory that both industry and researchers will either design comparable hardware and software or modify our existing setup to reach the next generation of physiologic signal analysis with enhanced data acquisition, archiving, and analysis capabilities. For researchers who use an existing physiologic signal data acquisition system or who plan to develop a new system, we suggest that commercially available automated peak detection systems (i.e., detection of heart beats or pulses) should be carefully examined for sensitivity and specificity. In our experience, these programs do not detect peaks with sufficient sensitivity or specificity nor do they adequately identify or handle signal artifact. We prefer a combination of a programmable peak detection algorithm with manual editing capabilities. In a 1988 study of postoperative pediatric cardiac surgery patients, Gordon et al. (33) stated that current patient monitoring practices may not always detect transient or evolving cardiovascular instability. They suggested that recognizing and potentially reversing a pattern of severe cardiovascular stress or loss of regulatory capacity before clinical manifestations may provide a potent diagnostic and prognostic tool in the ICU. These ideas now seem prescient. The computerized data acquisition and analysis system presented here is the most recent step in the study of critical illness and injury from a dynamic perspective. Future development of the system will include digital video or bedside monitoring of patients to allow differentiation of various external perturbations (movement, noxious stimuli, drug treatments, noise,
etc.), real-time physiologic analysis of linear and nonlinear metrics to describe specific pathophysiologic ICU disease states and processes and to predict sudden clinical deterioration, and assessment of the dynamic pharmacophysiologic effectiveness of certain drugs.
REFERENCES 1. Mizelle KM, Goldstein B: Advances in intensive care ECG and cardiac event monitoring. In: Cardiac Electrophysiology Review. Andersen K (Ed). New York, McGraw-Hill Professional, 1999, pp 245–248 2. Goldberger AL: Applications of chaos to physiology and medicine. In: Applied Chaos. Kim JH, Stringer J (Eds.) New York, WileyInterscience, 1992, pp 321–331 3. Belair J, Glass L, an der Heiden U, et al: Dynamical disease: Identification, temporal aspects and treatment strategies of human illness. Chaos 1995; 5:1–7 4. Buchman TG: Physiologic stability and physiologic state. J Trauma 1996; 41:599 – 605 5. Goldstein B, Fiser D, Kelly MM, et al: Decomplexification in critical illness and injury: The relationship between heart rate variability, severity of illness, and outcome. Crit Care Med 1998; 26:352–357 6. Goldstein B, Buchman TG: Heart rate variability in intensive care. J Intensive Care Med 1998; 13:252–265 7. McNames J, Veiga CC, Aboy M, et al: Precursors to Rapid Elevations in Intracranial Pressure: Proceedings of the 23rd Annual International Conference of the IEEE. 2001; 4:3977–3980 8. Tsui FC, Ching-Chung L, Sun M, et al: Acquiring, modeling, and predicting intracranial pressure in the intensive care unit. Biomed Eng 1996; 8:96 –108 9. Pelosi G, Emdin M, Carpeggiani C, et al: Impaired sympathetic response before intradialytic hypotension: A study based on spectral analysis of heart rate and pressure variability. Clin Sci 1999; 96:23–31 10. Korhonen I, van Gils M, Gade J: The challenges in creating critical-care databases. IEEE Eng Med Biol Mag 2001; 20:58 – 62 11. Kropyvnytskyy I, Saunders F, Schierek P, et al: A computer system for continuous longterm recording, processing, and analysis of physiologic data of brain injured patients in ICU settings. Brain Inj 2001; 15:577–583 12. Moody GB, Mark RG: A database to support development and evaluation of intelligent critical care monitoring. IEEE Comp Cardiol 1996:675– 680 13. Shann F, Pearson G, Slater A, et al: Paediatric index of mortality (PIM): A mortality prediction model for children in intensive care. Intensive Care Med 1997; 23:201–207 14. Teasdale G, Jennett B: Assessment of coma and impaired consciousness: A practical scale. Lancet 1974; 2:81– 4 15. Jennett B, Bond M: Assessment of outcome
Crit Care Med 2003 Vol. 31, No. 2
16. 17.
18.
19.
20.
21.
22.
after severe brain damage. Lancet 1975; 4:480 – 484 Fiser DH: Assessing the outcome of pediatric intensive care. J Pediatr 1996; 121:68 –74 Goldberger AL: Non-linear dynamics for clinicians: Chaos theory, fractals, and complexity at the bedside. Lancet 1996; 347: 1312–1314 Goldberger AL, Amaral LAN, Glass L, et al: Components of a new research resource for complex physiologic signals. Circulation 2000; 101:e215– e220 Heart rate variability: Standards of measurement, physiological interpretation, and clinical use. Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology. Circulation 1996; 93:1043–1065 Goldstein B, Woolf PD, DeKing DE, et al: Power spectral analysis of heart rate variability and plasma catecholamine levels after postural change and cold pressor testing in man. Pediatr Res 1994; 36:353–363 Moody G, Mark R: The MIT-BIH arrhythmia database on CD-ROM and software for use with it. Comput Cardiol 1990; 17:185–188 MIT-BIH Database and Software Catalog. Available at: ecg.mit.edu/dbinfo.html#
Crit Care Med 2003 Vol. 31, No. 2
23.
24.
25.
26.
27.
28.
29.
DBSoftwarePackage. Accessed December 7, 2002. Korhonen I, Ojaniemi J, Nieminen K, et al: Building the IMPROVE data library. IEEE Eng Med Biol Mag 1997; 16:25–32 Thomsen CE, Gade J, Nieminen K, et al: Collecting EEG signals in the IMPROVE data library: Data acquisition and visual analysis tools for obtaining prolonged recordings in intensive care. IEEE Eng Biol Med Mag 1997; 16:33– 40 Gade J, Korhonen I, van Gils M, et al: Technical description of the IBIS Data Library: Improved monitoring for brain dysfunction in intensive care and surgery. Comput Methods Programs Biomed 2000; 63:175–186 Jakob SM, Nieminen K, Karhu J, et al: IBIS data library: Clinical description of the Finnish database. Comput Methods Programs Biomed 2000; 63:161–166 Mandersloot GF, Pottinger RC, Weller PR, et al: The IBIS Project: Data collection in London. Comput Methods Programs Biomed 2000; 63:167–174 Toweill D, Sonnenthal K, Kimberly B, et al: Linear and nonlinear analysis of hemodynamic signals during sepsis and septic shock. Crit Care Med 2000; 28:2051–2057 Godin PJ, Buchman TG: Uncoupling of bio-
30.
31.
32.
33.
34.
logical oscillators: A complementary hypothesis concerning the pathogenesis of multiple organ dysfunction syndrome. Crit Care Med 1996; 24:1107–1116 Ellenby MS, McNames J, Lai S, et al: Uncoupling and recoupling of autonomic regulation of the heart beat in pediatric septic shock. Shock 2001; 16:274 –277 Winchell RJ, Hoyt DB: Spectral analysis of heart rate variability in the ICU: A measure of autonomic function. J Surg Res 1996; 63: 11–16 Yien HW, Hseu SS, Lee LC, et al: Spectral analysis of systemic arterial pressure and heart rate signals as a prognostic tool for the prediction of patient outcome in the intensive care unit. Crit Care Med 1997; 25: 258 –266 Gordon D, Herrera VL, McAlpine L, et al: Heart-rate spectral analysis: A noninvasive probe of cardiovascular regulation in critically ill children with heart disease. Pediatr Cardiol 1988; 9:69 –77 Aboy PM, McNames J, Goldstein B: Automatic detection algorithm of intracranial pressure waveform components. Proceedings of the 23rd Annual International Conference of the IEEE. 2001; 3:2231–2234
441
