☆ This work was supported in part by the National Science Foundationgrant 0830-370-V601.

e30 Meeting Abstracts for the Society for Complexity in Acute Illness (SCAI)

Objective: Sepsis is a medical condition defined by systemicinflammation and the presence of infection. Systemic inflamma-tion may progress to multiple-organ dysfunction and eventualdeath. Animal models of lethal sepsis have demonstrateddefective neutrophil recruitment to the site of infection,accompanied by high levels of recruitment to healthy organs.We propose that a compartmental model of sepsis will provideinsight into misregulation of neutrophil trafficking in sepsis.Methods: We constructed a compartmental model of peritonealsepsis using the rule-based modeling platform BioNetGen. Themodel includes 3 compartments: peritoneum, blood, and othertissues. Pathogen, macrophage, endothelial cells, and neutrophilspopulate the model. Macrophages detect pathogen in the perito-neum and produce the proinflammatory cytokine interleukin 1 (IL-1). Interleukin 1 activates local endothelial cells, leading toneutrophil adhesion. Adherent neutrophils migrate into the tissuewhere they destroy pathogen and secrete IL-1 receptor antagonist(IL-1RA). Interleukin 1 receptor antagonist counteracts theproinflammatory effects of IL-1 and accelerates the return tobaseline as the infection is cleared. IL-1 and IL-1RA diffusebetween compartments and are eliminated in the bloodstream byaction of the kidney and liver.

Parameters were obtained from the literature, where possible, orselected from a biologically realistic range. The rule-based modelwas expanded into a full reaction network and simulated as a set ofordinary differential equations. A comparison study of sublethaland lethal peritoneal infection was performed by initializing variousdoses of pathogen in the peritoneal compartment. A lethal modelwas defined by a pathogen dose sufficiently large to overwhelm thehost response. Blood purification protocols were simulated byincreasing the rate of elimination of IL-1 and IL-1RA from theblood compartment.Results: Simulations of sublethal sepsis resulted in rapid recruit-ment of neutrophils to the site of infection and minimal recruitmentto healthy tissue. Neutrophil counts in the peritoneum increased asthe initial pathogen load was increased, but a saturation effect wasobserved as the pathogen load approached lethal levels. In sublethalsimulations, endothelial activation and neutrophil recruitmentrapidly returned to baseline as the pathogen was cleared and theanti-inflammatory mediator dominated. Simulations of lethal sepsisresulted in extensive neutrophil recruitment to the healthy tissuecompartment, whereas recruitment to the infection site did notincrease over sublethal levels. This correlated with high levels ofcirculating mediators, systemic endothelial activation, and reducedlevels of circulating neutrophils.

Simulations of blood purification protocols resulted in theclearance of pathogen doses that were lethal in the absence ofpurification. Rescue by blood purification was only observed inpathogen doses up to 10% higher than the lethal threshold. Bloodpurification resulted in lower levels of circulating mediators,reduced systemic endothelial activation, reduced neutrophil recruit-ment to peripheral tissues, and increased recruitment to the site ofinfection. This result replicates experimental findings that hemoad-sorption, a blood purification therapy, increases short-term survivalin a rat model of sepsis.Conclusions: Simulations of our compartmental, rule-basedmodel of sepsis reveal that systemic neutrophil recruitmentcontributes to lethality through depletion of circulating pools and

subsequent reduction of recruitment to the primary site ofinfection. Furthermore, blood purification simulations suggest amechanism through which intervention decreases systemicneutrophil recruitment, increases recruitment to the infection,and improves survival.

doi:10.1016/j.jcrc.2009.06.033

An agent-based modeling framework for ontology integration: Towardformal executable knowledge representation☆

Gary An a, Miles Parker b, Scott Christley c

aDepartment of Surgery, Northwestern University, University of California,

IrvinebMetascape, LLC, University of California, IrvinecDepartment of Mathematics and Computer Science, University of

California, Irvine

Objectives: In today's high throughput, data-rich environment,technology has been able to augment various aspects of thescientific process. However, to date, there has been sparse work ontechnological enhancement of the intuitive processes of hypothesisformulation and conceptual model verification. These “discovery”steps can be augmented via the instantiation of thoughtexperiments in a “virtual sandbox” where researchers can bothformally represent their conceptual models and instantiate themdynamically such that the behavioral consequences of their beliefscan be examined and communicated. This has most recently beentermed executable biology but has its roots in the simulationcommunity. There have been nascent projects using object-oriented and Petri-net modeling to translate ontological informa-tion into dynamic models; this abstract introduces an agent-basedframework to aid in bridging the gap between ontologies andexecutable models.Methods: We present an agent-based framework (ABMF) toconcatenate ontologic descriptions of components and functionsand facilitate the development of an executable layer ofknowledge representation. The ABMF integrates terms expressedin standard BioPortal ontologies into modules that incorporate thecharacteristics and properties needed to construct an agent-basedmodel. The modules of the ABMF are organized in a series oforthogonal hierarchies that draw upon the structures ofontologies, object-oriented programming, and agent-based mod-eling. The representation of a conceptual model in the ABMFprovides a knowledge structure that can then be transformed intoexecutable models.Results: We describe an ABMF use-case of a specific multiscaleconceptual model: the pathophysiology of the gut-lung axis insystemic inflammation related to epithelial tight junction integrity.The conceptual model crosses scales of organization extendingfrom gene activation to clinically observed pulmonary failure. Thecomponents of the conceptual model are classified from 9 differentNCBO BioPortal ontologies and concatenated using the ABMF intoa single knowledge representation.Conclusion: We present an example of how an ABMF may beimplemented to extend current knowledge representation

e31Meeting Abstracts for the Society for Complexity in Acute Illness (SCAI)

capabilities. The intent is not to suggest that an agent-basedframework is “the” format for an executable ontology layer; theexperience with classic ontologies suggests that it is impossible toproduce a definitive knowledge representation tool. It is fullyexpected that other researches will develop similar types of tools,using different modeling paradigms. However, we believe that anABMF demonstrates a robust, evolvable approach and willprovide an essential translational step toward the goal of havingas many steps as possible of the discovery process be automated.Eventually, this type of architecture would allow the automatedgeneration of models that can be picked through by the researcher,providing a truly evolutionary solution to the synthetic aspect ofbiomedical research.

doi:10.1016/j.jcrc.2009.06.034

Table 1 The aROC values and Brier scores of the differentexperiments

Input aROC BS

Admission 0.650 0.226Dynamic coefficients 0.721 0.182Combined model 0.720 0.187Nurses 0.724 0.241Doctors 0.788 0.212

Fluctuation-dissipation theorem provides a simple analyticalrelationship between post-stress heart rate recovery andheart rate variability during the stressAnton Burykin a, Yan Lu b, Michael W. Deem b, Timothy G. Buchman a

aWashington University, Saint Louis, MObRice University, Houston, TX

Objectives: The autonomic nervous system modulates both thedynamics of heart rate (HR) recovery (HRR) after a cardiac stresstest (eg, treadmill exercise test) and the HR variability (HRV) understeady-state (“free-running”) conditions. Also, both reduced HRVand prolonged HRR are believed to be predictors of mortality. Thus,we hypothesized that there is a relationship between post-stressHRR time constant (Toff) and an HRV measure.Methods: Although all previous studies have used conventionalstatistical tools (such as correlation coefficients) to explore empiricalcorrelations between HRR time constant and different HRV indicesduring the exercise, we adopted an alternative strategy, deriving such arelationship theoretically rather than empirically (statistically) inferringit from the data. We applied the fluctuation-dissipation theorem (FDT),which relates the system response to a relatively small perturbation tothe fluctuations in the stationary state of the system.

We test our theoretical results with 20 HR data sets recordedduring and after the spontaneous breathing trial (SBT), consideredas a stressor, from 16 mechanically ventilated critically ill patientsin an intensive care unit.Results: The FDT predicts the exponential shape of HRR andprovides a simple analytical relationship linking post-stress HRRtime constant with a standard HRVmeasure, namely, the correlationcoefficient of the Poincare plot of the HR dynamics during the stress(SBT). For 17 of 20 data sets, FDT correctly predicts the values ofpost-SBT HRR time constant using HRV during the SBT. The caseswhen the FDT-based relationship fails correspond to nonexponen-tial shapes of HRR.Conclusion: The relationship between the microscopic fluctuations(HRV) during the stress and the macroscopic response (HRR) afterthe stress was terminated can be interpreted as an example of FDTor, equivalently, Onsager Regression Hypothesis, which states thatthe same time constant characterizes both the decay of spontaneousmicroscopic fluctuations (HRV) at steady state and the macroscopicrelaxation (HRR) to a new steady state after a relatively smallexternal perturbation (stress) was withdrawn.

The approach is specific neither to cardiac physiology nor totransitions between mechanical and spontaneous ventilation as aspecific stress. It may therefore have wider applicability (both inintensive care unit and elsewhere) to many physiologic systemssubjected to modest stresses.

doi:10.1016/j.jcrc.2009.06.035

Dynamic information improves discharge prediction after cardiac surgeryK. Van Loon a, F. Guiza b, G. Meyfroidt c, J.-M. Aerts a, H. Blockeel b,

G. Van den Berghe c, D. Berckmans a

aDivision Measure, Model & Manage Bioresponses, Katholieke

Universiteit Leuven, Kasteelpark Arenberg 30, B-3001 Leuven, BelgiumbDepartment of Computer Sciences, Celestijnenlaan 200a, B-3001 Leuven,

BelgiumcDepartment of Intensive Care Medicine, University Hospital

Gasthuisberg, Herestraat 49, B-3000 Leuven, Belgium

Objectives: To predict the probability of ICU discharge the dayafter scheduled cardiac surgery and to examine the independentcontributions of data available upon ICU admission and dynamicdata of the first 4 hours of ICU stay.Methods: All 461 adult patients who were admitted to our 56-bedsurgical ICU after scheduled cardiac surgery between the January18 and December 4, 2007, were selected as development cohort.All data in this study were available in the computerized medicalchart of the patient (Metavision, iMD-Soft). We used thefollowing ICU admission data: age, sex, body mass index,weekday and hour of admission, APACHE II score, history ofdiabetes, use of temporary epicardiac pacing, presence ofendocarditis, preoperative lung function, heart rate, bloodpressure, baseline serum creatinine, type of surgery, andintraoperative blood loss. Three signals with a sample intervalof one measurement per minute were selected for time seriesanalysis: heart rate, systolic arterial blood pressure, and oxygensaturation measured by pulse oximetry (SpO2). The dynamicfeatures of these signals were calculated using multivariateautoregressive models, cepstral coefficients, detrended fluctuationanalysis results, and approximate entropy values. We used 4 hoursof data of each of these signals. Models were built using aGaussian process classifier to predict the probability of ICUdischarge on the day after surgery. Three separate models weredeveloped: a first model used only the ICU admission data, asecond model used only the coefficients from the dynamic dataanalysis, and a third model was based on a combination of bothadmission data and dynamic coefficients. Models were validatedin a previously unseen validation cohort of 116 patients. Areaunder the receiver operating characteristic curve (aROC) was used