Background and Significance

Scope of the Problem: In the United States of America, the number of PICUs, NICUs, and non ICU beds available for children will increase to 5,500, 20,000, and 110,000, respectively, by 2010. Rendering care to these critically ill and hospitalized children represents a large, significant, and growing cost to the nation’s health care expense. The importance of critical care monitoring and the role of Rapid Response Teams in patient safety is recognized as a priority by the Joint Commission on Accreditation of Hospital Organizations (JCAHO).6 The critical care monitoring industry alone is a $1.2 billion per year business, growing at 3.9% per year, providing advanced critical care monitoring capabilities worldwide. Ninety percent of market share is adult, 8% neonatal and 2% pediatric. The need for continued improvement and advancement of monitoring devices for critically ill and hospitalized children is significant.

PICU Monitoring Capabilities: The most important and essential function of a PICU is to provide advanced monitoring capabilities to A) evaluate the clinical status of patients and B) track their response to a wide range of interventions. Morbidity and mortality have diminished dramatically in PICUs across the US due to advancements in medical and surgical therapeutics, life saving pharmaceuticals, and training of health care providers. Despite tremendous advancements in computer technology and bioinformatics, critical care monitoring devices have not provided any significantly new information to the bedside caregivers. These standard monitoring parameters include measurement of: intermittent or continuous core temperature; continuous heart rate with arrhythmia and ST segment analysis; intermittent oscillometric or continuous invasive arterial blood pressure; continuous invasive central venous pressure; continuous pulsed oximetry; and continuous end tidal carbon dioxide concentrations. Current monitoring systems do not present these data in a format that fosters understanding of the complex dynamic changes occurring instantaneously or over time in a biological system. Simple, time domain measures of heart rate and respiratory rate are displayed such as: 3-5 second mean values; mean values of parameters plotted against time; or alarms for out of range values determined by manufacturers or modified by bedside care providers. These values are often not scaled to reflect age or diagnosis adjusted norms for children. The need for advancement of these monitoring devices and for the introduction of new technologies is great.

Recent Advancements in Monitoring Technology: In recent years, critical care monitoring technology has been advanced by: 1) developing monitors of different sizes to accommodate children of all ages with critical illness; 2) improving technology to allow for more accurate and precise measurements of physiologic parameters; and 3) developing simple, easy to use and read devices. One “advanced” multivariate measure of patient instability is VARITREND (Spacelabs Medical, Redmond, WA, USA) which is used in the neonatal ICU to trend heart rate, oxygen saturation, and apnea in the time domain. HERO (Medical Predictive Science Corporation, Charlottesville, VA USA) is a new medical device developed in the neonatal ICU which is FDA cleared for the identification of transient decelerations and reduced baseline variability in heart rate in neonates with sepsis. A pediatric trial failed to demonstrate significant benefit in critically ill patients in the PICU. BIOSIGN Patient Status Index (Oxford BioSignals LTD, UK) is a new monitor used on adult hospital wards to identify patients who are decompensating by analyzing changes in vital signs (heart rate, blood pressure, respiratory rate, temperature and oxygen saturation). This monitor is capable of triggering a “rapid response team” to aid the patient. Despite this capability, no improvement in patient outcome was shown in a randomized trial of the technology. For PICU patients, essentially no “new” information has been made available to bedside clinicians to guide management decisions. We have created REALTROMINS to be an advanced, multivariate, medical device that will provide a continuously updated prediction of the risk of mortality in critically ill and hospitalized children and display the factors contributing to this elevated risk to the bedside caregivers in order to guide the medical management of these patients.

A variety of advanced organ specific monitoring devices have emerged over recent years with mixed utility and acceptance in the PICU. These devices include: continuous cardiac output measurements (PICCO, Pulsion Medical AG, Munich, Germany); thoracic electrical bioimpedance continuous cardiac output (Bio Z® (Cardiodynamics, San Diego, CA USA) and TEBCO, (Hemo Sapiens, Sedona, AZ USA); continuous in vivo arterial blood gas analysis (pH, PCO2, PO2, HCO3, base excess, oxygen saturation) (VIA LVM Monitor; Metracor Technologies, Inc San Diego, CA USA); continuous pulmonary mechanics and ventilator parameters (Servo I, Siemens AG, Munich, Germany); continuous venous oxygen saturation measurement (Baxter Healthcare Corporation, Deerfield, IL USA); continuous processed electroencephalographic analysis (A2000 Bispectral Index monitoring system, Aspect Medical Systems, Natick, MA USA), and other devices. These technologies have provided additional information to the bedside clinician, but most have failed to demonstrate improvement in patient morbidity or mortality. In fact, one “accepted” form of invasive monitoring (Swan Ganz catheter) may result in increased mortality in critically ill adult patients. Ultimately, each of these monitoring devices provides a unidimensional time domain appraisal of a specific organ function that must be assimilated, interpreted, and acted upon by the bedside physician, nurse, or allied health care worker. The human interface remains the best, most important, and yet, biased and flawed monitoring device. We have created a new tool, REALTROMINS, to assist the bedside clinician in this regard.

Bioinformatics in the PICU: The field of bioinformatics has emerged as a result of the vast array of information that must be acknowledged, processed, and acted upon in caring for critically ill patients. Computerized ICU charting programs (CareVue, Phillips Medical Systems Andover, MA, USA) have been developed and are being slowly deployed in PICUs across the country. Despite their tremendous potential for improving the quality and quantity of information provided to health care providers, no demonstrable impact on patient outcomes has been shown. Despite this, nursing satisfaction, increased direct RN-patient contact, decreased nursing charting time, and increased availability of information across an institution have all been accomplished.10-15 Computerized physician order entry has recently emerged as an important new technology due to the number of medication errors that occur in hospitals and the significant morbidity and mortality that results. This technology has emerged as a proposed mechanism for improving healthcare and controlling costs according to the LEAPFROG Group (The Leapfrog Group, Washington, DC, USA). However, a recent study found an INCREASED mortality rate when deployed in a leading tertiary care PICU. These two technologies have the potential to positively impact patient care by improving the quality and quantity of information the bedside care providers have to make better patient care decisions. Arguably, no new information has been created, but rather “old” information has been reformatted and made more accessible to health care providers. REALTROMINS will be a “new” bioinformatics technology in the PICU, NICU, and outside the ICU to continuously assess the changing severity of illness of pediatric patients, and track the success or failure of medical interventions ordered by physicians. Bedside caregivers can quickly determine which therapies are leading to a decrease in risk of mortality while those that increase risk should trigger new approaches to the clinical problem. By continuously displaying the overall risk score (REALTROMINS) and the individual components that contribute to that increased risk, patient outcomes should improve.

Risk of Mortality Prediction in the PICU: The pediatric risk of mortality score (PRISM) was developed by Dr. Pollack and published in 1988.43 It remains the most referenced article in the pediatric critical care literature. It was originally composed of 14 physiologic variables obtained on admission to the PICU and two general diagnostic/demographic variables which were entered into a logistic regression equation which computed severity adjusted risk of mortality. It was recalibrated and revised in 1996 as PRISM III.3 PRISM III has 17 physiologic variables and 4 general diagnostic/demographic variables which are entered into a proprietary logistic regression equation. It is based on principles developed in adult (APACHE III) and pediatric (PSI) severity of illness scoring systems. It remains the gold standard for benchmarking the performance of PICUs around the country. A second, nonproprietary, risk of mortality scoring system has been developed and is gaining acceptance (PIM2-Pediatric Index of Mortality). Despite the importance of a pediatric risk of mortality score, numerous limitations exist. PRISM III and PIM2 are static scoring systems that must be updated and recalibrated infrequently. Several reports suggest that PRISM may not reflect the rapidly advancing and improving care delivered in PICUs today or reflect the risk of mortality of certain subpopulations of patients (i.e. surgical and medical pediatric cardiac patients).7 Most importantly, PRISM III and PIM2 fail to guide the care of individual critically ill patients, since risk of mortality is computed on admission and not updated during the course of stay in the PICU. This limitation prevents the bedside health care providers from evaluating the changing status of the critically ill pediatric patient or their response to a wide variety of therapeutic interventions.

Physician’s Ability to Prognosticate Mortality: Physicians ability to prognosticate mortality in the PICU and other ICUs has been studied fairly extensively over the last decade. Marcin first reported on prognostication AND certainty in the PICU by comparing physician (attending, fellow, and resident) and nurse predictions of mortality with how certain the providers were of the prediction when compared to PRISM III. Attendings had the best predictive ability while fellows, residents, and nurses consistently OVER predicted mortality in critically ill children. Only attendings and fellows were calibrated to PRISM III in predicting mortality across the full spectrum of mortality (low, medium, high). Mortality predictions did not improve after 48 hours when compared to estimates at 24 hours, highlighting a previously described inability to learn from sequential clinical information over time. Weighting the mortality prediction with a measure of certainty did minimally help the accuracy of the prediction and yet all health care providers were over confident of their certainty. Marcin followed this study up by assessing the addition of subjective estimates of mortality provided by attending physicians to the PRISM III physiology based estimate creating a new Bayesian estimate. This combined mortality prediction was minimally improved over either the subjective or actuarial estimates.

Subjective mortality predictions in non PICU patients has been assessed and the results were varied with studies showing either high or low predictive abilities of physicians when compared to computer models depending on the experience of the physician, categories of patients studied, or the severity of illness of the patients. In the neonatal ICU, both nurses and physicians were able to equally predict mortality with a high degree of accuracy when compared to the Score for Neonatal Acute Physiology (SNAP). One short coming of the subjective estimate was that it was poorly calibrated across the spectrum of mortality risk (low, medium, high) with a consistent bias of overestimating mortality. This demonstrated pessimism in outcome prediction has been termed “reverse ego bias.” In the adult ICU, physicians and slightly less so, nurses, were as accurate as the APACHE II score in predicting mortality. Outside of the ICU setting, physicians ability to estimate survival in terminally ill adult patients was particularly poor.60 In this study, only 20% of the survival predictions were accurate (defined as being ± 33% of the actual length of survival). Sixty three percent of predictions were overly optimistic while 17% were overly pessimistic. Older physicians were most accurate in their predictions. Non oncology medical specialists and physicians with the longest relationship with patients were overly pessimistic. In contrast, physicians with the most recent contact with patients had overly optimistic predictions of survival as reported to investigators (but not directly to the patient). These studies highlight the subjective bias that is introduced by health care providers when attempting to prognosticate mortality. While certain groups of physicians caring for certain subpopulations of patients are AS ACCURATE as computer generated predictive models, no subjective estimates are BETTER than the predictive models. Finally, combining subjective and actuarial estimates has some improvement in accuracy over either prediction alone, albeit small.

Another critical point to be made is that all severity of illness scoring systems (PRISM III, PIM II, APACHE III) and REALTROMINS predict mortality independent of physician (or other caregiver) input into the model. While this may appear to be a limitation of the scoring systems, it in fact remains an essential part of the models. For REALTROMINS, the action, inaction, or inappropriate action of the physician effects the physiologic parameters being measured (ie heart rate variability) in ways that the model recognizes rather than the physician influencing the model directly. Indirectly health care providers do “influence” the REALTROMINS model by the frequency and specific laboratory tests being ordering. Patients who are felt to be stable are not subjected to routine laboratory evaluation, whereas hypercritical patients have frequent laboratory evaluations (ie arterial blood gas determinations). Since REALTROMINS is a multivariate model with inputs of the most salient laboratory tests, it can be “influenced” by the concerns of the ordering physicians and nurses caring for the patient.

Heart Rate Variability Analyses: Spectral analysis (also called, power spectral analysis or auto spectral analysis) of heart rate variability has been established in both adults and children as an extremely useful tool for quantifying the dynamic, physiologic changes occurring in a wide variety of critical illnesses, such as shock, severe head trauma, post operative congenital heart disease, acute myocardial infarction, diabetes, etc. The concept is that healthy biological systems are in a natural state of “chaos” and when critical illness develops, this natural variability is lost. Analysis of the low frequency (LF) and high frequency (HF) spectral content of heart rate variability, with calculation of a ratio of these two areas has been predictive of mortality in children. These measurements are indicators of sympathetic and parasympathetic neuromodulation of the heart. These analyses have also been weakly negatively correlated with: 1)PRISM (r2=0.292), 2)Pediatric overall performance category, and 3)Pediatric cerebral performance category scores, in a multidisciplinary PICU.21 Despite the potential of spectral analysis of heart rate variability for predicting outcome, this has not been accepted in pediatric critical care medicine as a standard monitoring parameter. It remains limited to research applications. Current computer, engineering, and bioinformatics technologies have advanced such that continuous monitoring of multiple physiologic parameters in a 16 bed PICU have been accomplished and described. We have created an analogous system in our PICU using another vendor (Spacelabs Healthcare, Redmond, WA, USA). We believe that displaying only time domain measurements of these important physiologic parameters limits essential information necessary for decision making by bedside clinicians. In addition, displaying measures of variability of these parameters is as important as the individual values themselves. Coupling these measurements into a multivariate predictor of mortality (REALTROMINS) will improve the predictive capability of continuous physiologic monitoring.

Signal Processing: Physiological signals encode information which characterizes the spectrum from normal to disease conditions in living systems. Physicians are trained to recognize diagnostic patterns in these signals. Signal processing methodologies enhance diagnostic capabilities though either detection of information not otherwise perceptible, or quantification of physiological measures leading to diagnostic differentiation of sickness from health. Signal processing techniques may be broadly categorized as spectral and time series methodologies, while wavelets span the two approaches. Time series methods may be linear or non-linear, stationary or non-stationary, parametric or non-parametric. Signal processing methods presume stationarity, while non-linear methods, such as bispectral analysis, have been applied to diagnostics in recent years. Neural networks are trained to “recognize” patterns derived from physiological signals using characteristics derived from any signal processing methodology, which are linked to health and disease states, or trends toward either. Applications of bispectral methods can be used to detect nonlinear trends in critically ill patients. Heart rate variability using linear signal processing methods have been developed for continuous real time analysis of spectral bands which characterize autonomic activity. This approach provides a quantitative measure of autonomic activity as a trend as patients progress toward health or disease. Detection of trends can be enhanced using cross spectral quantification of interactions between physiological signals determined continuously in real time.53,40 This technique is one of many useful tools utilized in the signal processing strategy developed for REALTROMINS for determining signal artifact (noise) from significant changes in clinical state (signal). These methodologies are being applied in REALTROMINS. The success of REALTROMINS is due to the multidisciplinary collaboration with the Department of Biomedical Engineering at UNC-Chapel Hill.

Neural Networks and Datamining: Data mining is an extension of statistical hypothesis testing. In the past, assumption failure and required sample sizes have restricted the ability of computational methods to identify significant trends. With the advent of high-speed computers and new machine learning algorithms, neural network development and data intensive mining have become highly useful techniques. The basic underpinning of data mining can be summarized as “Predictive.” Non stationary effects are not seen in human clinical trials and this may have hampered progress in adapting old techniques to new clinical data problems. We are utilizing advanced neural network techniques which are part of the SAS Enterprise Miner Software.