Thomas Heldt

Multiparameter Intelligent Monitoring in Intensive Care II: A public-access intensive care unit database*

Objective:We sought to develop an intensive care unit research database applying automated techniques to aggregate high-resolution diagnostic and therapeutic data from a large, diverse population of adult intensive care unit patients. This freely available database is intended to support epidemiologic research in critical care medicine and serve as a resource to evaluate new clinical decision support and monitoring algorithms. Design:Data collection and retrospective analysis. Setting:All adult intensive care units (medical intensive care unit, surgical intensive care unit, cardiac care unit, cardiac surgery recovery unit) at a tertiary care hospital. Patients:Adult patients admitted to intensive care units between 2001 and 2007. Interventions:None. Measurements and Main Results:The Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC-II) database consists of 25,328 intensive care unit stays. The investigators collected detailed information about intensive care unit patient stays, including laboratory data, therapeutic intervention profiles such as vasoactive medication drip rates and ventilator settings, nursing progress notes, discharge summaries, radiology reports, provider order entry data, International Classification of Diseases, 9th Revision codes, and, for a subset of patients, high-resolution vital sign trends and waveforms. Data were automatically deidentified to comply with Health Insurance Portability and Accountability Act standards and integrated with relational database software to create electronic intensive care unit records for each patient stay. The data were made freely available in February 2010 through the Internet along with a detailed users guide and an assortment of data processing tools. The overall hospital mortality rate was 11.7%, which varied by critical care unit. The median intensive care unit length of stay was 2.2 days (interquartile range, 1.1–4.4 days). According to the primary International Classification of Diseases, 9th Revision codes, the following disease categories each comprised at least 5% of the case records: diseases of the circulatory system (39.1%); trauma (10.2%); diseases of the digestive system (9.7%); pulmonary diseases (9.0%); infectious diseases (7.0%); and neoplasms (6.8%). Conclusions:MIMIC-II documents a diverse and very large population of intensive care unit patient stays and contains comprehensive and detailed clinical data, including physiological waveforms and minute-by-minute trends for a subset of records. It establishes a new public-access resource for critical care research, supporting a diverse range of analytic studies spanning epidemiology, clinical decision-rule development, and electronic tool development.

Integrating Data, Models, and Reasoning in Critical Care

Modern intensive care units (ICUs) employ an impressive array of technologically sophisticated instrumentation to provide detailed measurements of the pathophysiological state of each patient. Providing life support in the ICU is becoming an increasingly complex task, however, because of the growing volume of relevant data from clinical observations, bedside monitors, mechanical ventilators, and a wide variety of laboratory tests and imaging studies. The enormous amount of ICU data and its poor organization makes its integration and interpretation time-consuming and inefficient. There is a critical need to integrate the disparate clinical information into a single, rational framework and to provide the clinician with hypothesis-driven displays that succinctly summarize a patients trajectory over time. In this paper, we present our recent efforts towards the development of such an advanced patient monitoring system that aims to improve the efficiency, accuracy, and timeliness of clinical decision making in intensive care

The cardiac output from blood pressure algorithms trial.

OBJECTIVE The value of different algorithms that estimate cardiac output (CO) by analysis of a peripheral arterial blood pressure (ABP) waveform has not been definitively identified. In this investigation, we developed a testing data set containing a large number of radial ABP waveform segments and contemporaneous reference CO by thermodilution measurements, collected in an intensive care unit (ICU) patient population during routine clinical operations. We employed this data set to evaluate a set of investigational algorithms, and to establish a public resource for the meaningful comparison of alternative CO-from-ABP algorithms. DESIGN A retrospective comparative analysis of eight investigational CO-from-ABP algorithms using the Multiparameter Intelligent Monitoring in Intensive Care II database. SETTING Mixed medical/surgical ICU of a university hospital. PATIENTS A total of 120 cases. INTERVENTIONS None. MEASUREMENTS CO estimated by eight investigational CO-from-ABP algorithms, and CO(TD) as a reference. MAIN RESULTS All investigational methods were significantly better than mean arterial pressure (MAP) at estimating direction changes in CO(TD). Only the formula proposed by Liljestrand and Zander in 1928 was a significantly better quantitative estimator of CO(TD) compared with MAP (95% limits-of-agreement with CO(TD): -1.76/+1.41 L/min versus -2.20/+1.82 L/min, respectively; p < 0.001, per the Kolmogorov-Smirnov test). The Liljestrand method was even more accurate when applied to the cleanest ABP waveforms. Other investigational algorithms were not significantly superior to MAP as quantitative estimators of CO. CONCLUSIONS Based on ABP data recorded during routine intensive care unit (ICU) operations, the Liljestrand and Zander method is a better estimator of CO(TD) than MAP alone. Our attempts to fully replicate commercially-available methods were unsuccessful, and these methods could not be evaluated. However, the data set is publicly and freely available, and developers and vendors of CO-from-ABP algorithms are invited to test their methods using these data.

Model-based estimation of cardiac output and total peripheral resistance

We describe a novel model-based approach to estimate cardiac output (CO) and total peripheral resistance (TPR) continuously from peripheral arterial blood pressure (ABP) waveforms. Our method exploits the intra-beat and inter-beat variability in ABP to estimate the lumped time constant of a beat-to-beat averaged Windkessel model of the arterial tree, from which we obtain an uncalibrated estimate of CO. To estimate absolute CO, we determine the lumped arterial compliance using calibration data, and assuming either constant or state-dependent compliance. We applied our method to a porcine data set in which stroke volume was measured with an ultrasonic flowmeter. We obtain root-mean-square normalized errors of 11-13% across all pigs, lower than those obtained on the same data set using various other estimation methods. The CO estimates, and TPR estimates derived from them track intravenous drug infusions quite closely.

A multidimensional approach to pain assessment in critically ill infants during a painful procedure.

Objectives:Inferring the pain level of a critically ill infant is complex. The ability to accurately extract the appropriate pain cues from observations is often jeopardized when heavy sedation and muscular blocking agents are administered. Near-infrared spectroscopy is a noninvasive method that may provide the bridge between behavioral observational indicators and cortical pain processing. We aimed to describe regional cerebral and systemic hemodynamic changes, as well as behavioral reactions in critically ill infants with congenital heart defects during chest-drain removal after cardiac surgery. Methods:Our sample included 20 critically ill infants with congenital heart defects, less than 12 months of age, admitted to the cardiac intensive care unit after surgery. Results:Cerebral deoxygenated hemoglobin concentrations significantly differed across the epochs (ie, baseline, tactile stimulus, noxious stimulus) (P=0.01). Physiological systemic responses and Face Leg Activity Cry Consolability (FLACC) pain scores differed significantly across the events (P<0.01). The 3 outcome measures were not found to be associated with each other. Mean FLACC pain scores during the painful procedure was 7/10 despite administration of morphine. Midazolam administration accounted for 36% of the variance in pain scores. Discussion:We demonstrated with a multidimensional pain assessment approach that significant cerebral, physiological, and behavioral activity was present in response to a noxious procedure in critically ill infants despite the administration of analgesic treatment. Considering that the sedating agent significantly dampened pain behaviors, assessment of cerebral hemodynamic in the context of pain seems to be an important addition.

The ear as a location for wearable vital signs monitoring

Obtaining vital signs non-invasively and in a wearable manner is essential for personal health monitoring. We propose the site behind the ear as a location for an integrated wearable vital signs monitor. This location is ideal for both physiological and mechanical reasons. Physiologically, the reflectance photoplethysmograph (PPG) signal behind the ear shows similar signal quality when compared to traditional finger transmission PPG measurements. Ballistocardiogram (BCG) can be obtained behind the ear using 25mm×25mm differential capacitive electrodes constructed using fabric. The BCG signal is able to provide continuous heart rate and respiratory rate, and correlates to cardiac output and blood pressure. Mechanically, the ear remains in the same orientation relative to the heart when upright, thus simplifying pulse transit time calculations. Furthermore, the ear provides a discreet and natural anchoring point that reduces device visibility and the need for adhesives.

Numerical analysis of blood flow through a stenosed artery using a coupled multiscale simulation method

A global system model of the systemic circulation is combined with a local finite element solution to simulate blood flow in a stenosed coronary artery. Local fluid dynamic issues arise in connection with the detailed flow patterns within the stenosed coronary artery while the global system model is used to simulate the response of the rest of the circulation to the local perturbation. A PISO type finite element technique is employed to compute the local blood flow. The Navier-Stokes equations are solved with the assumption of viscous incompressible flow across the stenosed coronary artery. A detailed lumped parameter model simulates the characteristics of the coronary circulation and is imbedded in a coarse-grained lumped parameter model of the entire cardiovascular system. These two methods are coupled in that the lumped parameter calculations provide the time-dependent boundary conditions for the local finite element calculation. In turn, the local fluid dynamical computation provides estimates for the pressure drop across the stenosis, which is subsequently used to refine the lumped parameter calculation. Results are obtained for an axisymmetric coronary artery model with a stenosis of 90% area reduction over one cardiac cycle. Numerical results show that the flow rate and resistance are strongly coupled. Compared with the flow rate distribution computed from the global simulation with constant resistance, the coupled solution predicts a flow rate with only slight changes. The high flow rate during diastole increases the stenosis pressure drop and resistance. In turn, this increased resistance of the stenosis slightly reduces the flow rate computed in the lumped parameter simulation.

Automated Quantitative Analysis of Capnogram Shape for COPD–Normal and COPD–CHF Classification

We develop an approach to quantitative analysis of carbon dioxide concentration in exhaled breath, recorded as a function of time by capnography. The generated waveform-or capnogram-is currently used in clinical practice to establish the presence of respiration as well as determine respiratory rate and end-tidal CO2 concentration. The capnogram shape also has diagnostic value, but is presently assessed qualitatively, by visual inspection. Prior approaches to quantitatively characterizing the capnogram shape have explored the correlation of various geometric parameters with pulmonary function tests. These studies attempted to characterize the capnogram in normal subjects and patients with cardiopulmonary disease, but no consistent progress was made, and no translation into clinical practice was achieved. We apply automated quantitative analysis to discriminate between chronic obstructive pulmonary disease (COPD) and congestive heart failure (CHF), and between COPD and normal. Capnograms were collected from 30 normal subjects, 56 COPD patients, and 53 CHF patients. We computationally extract four physiologically based capnogram features. Classification on a hold-out test set was performed by an ensemble of classifiers employing quadratic discriminant analysis, designed through cross validation on a labeled training set. Using 80 exhalations of each capnogram record in the test set, performance analysis with bootstrapping yields areas under the receiver operating characteristic (ROC) curve of 0.89 (95% CI: 0.72-0.96) for COPD/CHF classification, and 0.98 (95% CI: 0.82-1.0) for COPD/normal classification. This classification performance is obtained with a run time sufficiently fast for realtime monitoring.

Bayesian Networks for Cardiovascular Monitoring

Bayesian Networks provide a flexible way of incorporating different types of information into a single probabilistic model. In a medical setting, one can use these networks to create a patient model that incorporates lab test results, clinician observations, vital signs, and other forms of patient data. In this paper, we explore a simple Bayesian Network model of the cardiovascular system and evaluate its ability to predict unobservable variables using both real and simulated patient data

Blood pressure variability: can nonlinear dynamics enhance risk assessment during cardiovascular surgery?

As the population ages, increasing numbers of elderly patients with multiple co-morbid conditions are presenting for high-risk cardiovascular surgical procedures. The commensurate increase in perioperative major adverse events (MAEs) increases mortality by 1.4 to 8-fold,1 with an estimated 1 billion dollars annually spent on managing these complications.2 Current MAE risk prediction indexes3,4 are typically based on static or “snapshot” measures, such as the presence or absence of a co-morbid condition like hypertension. Unfortunately, these indexes have failed to adequately predict which high-risk patients will have MAEs5–7 possibly, at least in part, because they do not take into consideration the complex (nonlinear), time-varying features of physiological hemodynamic signals. Furthermore, a “one-size-fits-all” risk prediction model approach is unlikely to accurately identify patients at high risk5–7 particularly at extremes of age and predicted risk.8–13 A major motivation for the program outlined here is that current risk prediction tools may be improved by incorporating dynamical properties of physiologic signals, thereby enhancing: (a) individual patient risk assessment and counseling, (b) design of timely interventions to prevent disabling or fatal complications (e.g., stroke, renal failure, atrial fibrillation and myocardial infarction), and (c) the accuracy of comparisons of provider and hospital performances. Toward this end, our goal is to develop a real-time blood pressure variability (BP variability) index or set of indexes incorporating a patients own baseline and evolving pathophysiologic characteristics into current “snapshot” scoring systems.4,5,14 One of the most important physiologic signals obtained in the perioperative period is the continuously recorded systemic BP signal.15 While the optimization of BP is a major perioperative target there is no universally accepted guideline for defining hypotension. Furthermore, hypotensive episodes, are dynamic, not static phenomena and not only vary from patient to patient but also within a patient at different surgical stages. Therefore, measures of BP variability, quantified using different metrics, have been the focus of considerable interest. For example, in one study,16 BP variability was defined as the time spent above or below a target systolic blood pressure range of 95–135 mm of Hg, and an increased BP variability value was associated with higher 30-day mortality. In another study, BP variability was defined as the root mean successive square difference of a moving 5 second time period. In this investigation17, decreased intracranial pressure and BP variability were shown to predict long-term adverse outcome after aneurysmal subarachnoid hemorrhage. An intuitive limitation of traditional measures of variability is the fact that they do not take into consideration the temporal structure of a sequence of measurements. For example, the following two sequences: A = {1 2 3 2 1 2 3 2 1 2 3 2 1} and B ={1111222222333}16, have the same variability, as measured by amplitude of range and standard deviation, but completely different structures. In fact, while sequence A defines a triangular wave, sequence B is a step function. Measures that are sensitive to the temporal organization of a signal have been essential in characterizing and discriminating different dynamical systems. Here we assess BP fluctuation (variability) dynamics via two complementary metrics: 1) traditional standard deviation of BP time series and 2) the degree of complexity of their dynamics. The motivating framework for quantifying the degree of complexity of nonlinear physiologic signals, such as BP, is that complexity reflects the degree of robustness/resilience of the underlying control mechanisms, and it decreases with aging and pathology (http://physionet.org/tutorials/cv/, accessed Oct 21, 2013). The term nonlinear may be unfamiliar to readers of physiologic and clinical journals. Linear systems exhibit two properties: proportionality and superposition. Proportionality, as implied by the term, means that there is a straight-line relationship between input and output. Superposition indicates that one can completely understand the system (e.g., a Rube Goldberg-type device) by breaking it down into multiple sub-components. In contrast, the sub-components of a non-linear system do not “add up” to the whole because of either “constructive” or “destructive” interactions between those sub-components. In these cases, reductionist strategies will fail to provide full understanding of a given system.18,19 Furthermore, in nonlinear systems, unanticipated (“off-target”) effects are likely since small input changes may induce major changes in the output (the “so-called “butterfly effect”). Pilot Study: Overview In this pilot study, we tested the feasibility of: i) acquiring BP waveform data of sufficient length and quality for nonlinear complexity analyses, and ii) converting the data from a proprietary to an open-source format. Our specific hypothesis is that the complexity of the dynamics of systolic arterial (SAP), diastolic arterial (DAP) and pulse pressure (PP) from the post-bypass period is lower for the group of patients with MAEs (cases) than for a control group with comparable risk but no MAEs. We included pulse pressure dynamics in light of evidence that abnormalities in pulse pressure has been independently associated with up to 3-fold increase in MAEs following cardiac surgery.20