Modeling Critical Illness

1
Modeling Critical Illness
Gilles Clermont MD, MSc
Center for Inflammation and Regenerative Modeling
(CIRM) and The CRISMA Laboratory Critical Care
Medicine School of Medicine University of
Pittsburgh
2
Objectives

• The clinical problems facing Critical Illness
• Our work at the Center for Regenerative and
  Inflammatory Modeling
• The challenges We face

3
The Goal of Critical Care ?
4
The big challenges in Critical Illness

• Timely diagnosis
• Outcome prediction
• Development of targetted (personalized) therapies
• Modulation of the inflammatory response

5
Root causes of these challenges

• Insufficient, inaccurate data
• What to measure
• Point-of-care technologies
• Insufficient interpretative framework
• Uncertain biological mechanisms
• Biological variability
• Dearth of in silico disease models
• Insufficient mathematics

6
Critical Illnesses
Trauma/Shock
Severe infections (sepsis)
Inflammation
Acute coronary syndrome
Stroke
7
Current Paradigm of Injury/Recovery
Stress response

• Innate immunity
• Coagulation
• Metabolism

Anti- Inflammation
Inflam- mation
Recovery
Time
8
The canvasof opportunitites
Physiologic manifestations
Metabolic manifestations

• Early mediators
• TNF
• IL-1
• IL-10
• IL-6
• Cells

• Infection
• Pathogen
• Toxins

• Detection
• Dendritic cells
• Macrophages

Organ dysfunction
Late mediators
Death
Coagulation
9
Opportunities for Immune Support
Anti-LPS
Anti-TNF
Anti-IL-1
Anti-IL-10
Anti-HMGB1
Blood Purification
Immunologic Support
HMGB1
RECOVERY
IL-10
IL-6
TNF
INSULT
IL-1
Time
10
What about immunomodulation in sepsis and trauma?

• 25 years of global disaster
• Simplistic rationales
• Ineffective products
• Poor patient selection
• gt70 phase II trials, gt 1B dollars invested
• Two possible leads
• Low dose anti-inflammatory treatment
• Activated protein C
• Contradictory results!

11
Treating sepsis the strength of the Consensus
2 5 3 2 18
Good or Bad?
12
The inflammatory response
Huang Q, Science 2001
13
Calvano, Nature 2005
14
Modeling Inflammation at the CIRM

• Multiple models of acute inflammation (sepsis,
  trauma/hemorrhage, biowarfare agents,
  phonotrauma, wound healing), organ
  damage/dysfunction, and healing/regeneration
• Qualitative and quantitative predictions
• Probing mechanisms
• Have simulated device usage and guided device
  design
• Have outlined an iterative strategy for rational
  drug design and administration
• Have carried out simulated clinical trials in the
  settings of sepsis and trauma, including
  biowarfare applications

15
Small models of inflammation

• Top-down
• Understand the biology
• Biological plausibility
• High-level map of the biology
• Building blocs for more complex models

16
Reduced models of inflammation
17
Transients for 3 possible regimen
Health
Aseptic death
Septic death
18
Bifurcation analysis on Kpg
Septic death
Aseptic death
Heatlh
19
2-D bifurcation diagram
Opportunity
20
Manipulating anti-inflammatories
21
Why complicate things

• To produce a calibrated
• To intervene in the dynamics in a realistic
  way, more realistic handles are needed
• Not all modules need to be equally detailed
• The analysis of large models
• May rapidly become intractable
• May not yield useful results

22
Simulating Inflammatory Disorders at the CIRM
Research Biological Mechanisms
Develop Representative Models
Collect Biomarker Data
Calibrate Models to Data
Use Model for Predictions
23
A Unified Inflammatory Response
24
Simulations of infectious agents with bioterror
potential
Shock 2007
JTB 2007
25
Probing mechanisms
26
In silico design of RCTs
27
Why model a clinical trial

• Can accommodate host factors
• Can accommodate pathogen factors
• Allows what-if scenarios
• Evaluate consequence of an intervention based on
  supposed mechanism of action
• Allows systematic exploration of
• Dosing
• Duration of treatment, frequency of
  administration
• Drug combinations

28
Simulating a clinical trial

• Host factors
• Barrier defenses
• Genetic diversity in host response
• Current responsiveness status of the host
• Pathogen factors
• Virulence
• Site of infection
• Size of inoculum
• Stage of disease

29
Why model a clinical trial

• Suggests key biological markers to follow
• Allows interpretation of predictions
• Of patterns of biological markers
• Of survival benefit/detriment
• No randomization failures

30
Anti-TNF treatment for sepsisA simulation study
Clermont et al. 2004
31
Outcome by subgroups
Pathogen virulence (Quartile)
Anti-inflammatory responsiveness (Quartile)
32
Validating in silico simulation?

• Face validity of the disease model
• Knowledge of key driving factors
• Disease model includes these factors
• Intervention model
• Mechanism of action of the proposed treatment
• PK/PD data
• Predictive ability on empirical data
• Controlled experiments
• Existing trial data

33
Disease model

• Must account for uncertain mechanisms
• Model structure recapitulates biology
• Predictors in a statistical model
• Equations/rules in white box models
• Must make best use of observations at hand which
  are often incomplete
• Within a given model structure, develop an
  understanding of the breath of parameter
  realizations that fit data equally well
• Uncertainty in the relative importance of
  mechanisms/interactions
• Many model realizations are necessary

34
Who should be treated?

• Medical Decision Making
• Statistical model predicting probabilities of
  each of 4 outcomes
• Select therapeutic strategy

• Compare outcomes
• Model prediction (actual) vs. Statistical
  prediction

Disease starts
Disease detected
1st lab values
2nd lab values
60 minutes

• This is still not personilized therapy
• Standard therapeutic regimen
• Model predictions assumed true

35
Predictors of response to therapy
36
Value of a predictive statistical model
37
Value of a predictive model

• Strategy I treated only patient predicted to be
  saved by treatment
• 265 patients only would have been treated
• 22 patients that would have been helped did not
  received it (error of omission 2.2)
• 2 patients would have in fact been harmed (error
  of commission 0.2)
• 146 patients were not harmed by the treatment
  (14.6)
• Net benefit over indiscriminate administration
• 122 patients saved
• 735 doses avoided

38
Value of a predictive model

• Strategy II treated those predicted to be helped
  and those predicted to die anyway
• 397 patients treated
• 10 errors of omission
• 2 errors of commission
• 146 patients not harmed by treatment
• Net benefit over discrimate administration
• 134 live saved
• 603 doses avoided
• Strategy II over strategy I
• 12 lives gained
• 132 additional doses spent
• 11 doses/life-saved

39
Some core theoretical challenges

• The variability problem
• The inverse problem
• The cogent-reduction problem

40
The variability problem

• Can inter-individual variability be characterized
  in some fashion?
• Sepsis as an example
• How can mathematical models to capture
  variability?
• Can meaningful diagnostic and therapeutic
  insights be achieved in the presence of such
  variability?

41
Observed Variability

• Measurements
• Disease dynamics
• Less so in the relative timing of events
• Host biological idiosyncrasies
• Genetic
• Milieu

42
Day 1 cytokine levels in non-septic patients for
prediction of severe sepsis
TNF
IL-6
2.1
4.0
2.0
3.8
TNF at baseline and 95 CI (ln pg/ml)
3.6
IL6 at baseline and 95 CI
1.9
3.4
1.8
p 0.0009
p 0.0087
3.2
1.7
Severe sepsis(n268)
No SS(n1073)
Severe sepsis(n268)
No SS(n1073)
Analysis restricted to day 1 levels of those
patients who do NOT have severe sepsis on first
day
43
Day 1 levels and survival
TNF
IL-6
5.0
4.6
Baseline IL6 with 95 CI
4.2
3.8
plt0.0001
3.4
Dead(n212)
Alive(n1410)
IL-10
2.8
2.6
Baseline IL10 with 95 CI
2.4
plt0.0001
2.2
Dead(n212)
Alive(n1410)
44
Day 1 cytokine levels in patients who develop ARF
and those that do not
TNF
ARF RIFLE-I or F
Pg/ml ? SEM
plt0.0001
ARF (n258)
No ARF(n1544)
IL-6
IL-10
Pg/ml ? SEM
Pg/ml ? SEM
p0.0285
plt0.0001
ARF (n258)
No ARF(n1544)
ARF (n258)
No ARF(n1544)
45
IL-6, IL-10 patterns v. Outcomes
46
Trajectories

• Trajectories are consistent within individuals
• Shapes are often consistent across individuals

Suffredini et al. 1996
47
Cytokines by Outcome (60 days)
48
Trajectory Analysis
Log IL-10
Log IL-6
H
H
M
M
L
L
1 2 3 4 5 6 7
1 2 3 4 5 6 7
Day
Day
49
Mixed model trajectory analysis

• Describes the course of a behavior over time
• Identifies distinctive groups of individual
  trajectories in the population
• Profiles the characteristics of group members
• Applications
• CD4 counts (Zeger, et al. Biometrics 1994)
• Aggressive behavior (Nagin, et al. Psych Methods
  1999)

50
Some good news

• Standard (although by no means elementary)
  statistical techniques identify classes of
  patients
• Physiology
• Omics
• Qualitative patterns, but not magnitude of
  response, often preserved across individuals
• Within species

51
Mathematical models in translational research

• Black box models
• The data tells the whole story (statistical
  models, neural networks, other data mining-based
  models, e.g. SVMs)
• White box models
• Mechanistic (Agent-based or differential
  equation-based models)
• Grey box models
• Flexible mechanistic models
• Informed statistical models
• Others

52
One model one patient
M
M
M
53
Disease models

• Must account for uncertain mechanisms
• Model structure recapitulates biology
• Predictors in a statistical model
• Equations/rules in white box models
• Must make best use of observations at hand which
  are often incomplete
• Within a given model structure, develop an
  understanding of the breath of parameter
  realizations that fit data equally well
• Uncertainty in the relative importance of
  mechanisms/interactions
• Many model realizations are necessary

54
Parameters vs. structure
55
?DAMPs ?Th2
56
Patient-specific metamodel
M1
M2
Mn
E(Mn) Metamodel or Ensemble
Where the individual models vary in
their mathematical structure and parameters
57
Population-level Ensemble
E(Mn)
E(Mp)
E(Mq)

• Empirical rather than phenomenological

58
Targetted TherapyModel Predictive Control (MPC)

• Base System
• Real patient
• Metamodel

• INPUT
• Actual data

• OUTPUT
• Actual data
• Predicted data

• Sensor
• Error between
• actual/desired

Actuator
Controller

• The desired output is health
• The MPC method uses actual data and model
  simulations to estimate output the discrepancy
  is estimated (Sensor)
• The MPC method suggest an optimal intention
  strategy which is time dependent (Actuator)

59
MPC Schematic
Babatunde A. Ogunnaike and W. Harmon Ray.
Process Dynamics, Modeling, and Control (Topics
in Chemical Engineering). Oxford University
Press New York, 1994. pg. 997.
Past
Future
Reference Trajectory, R
R(k2)
R(k1)
Predicted Output, p
R(k)
p(k1)
p(k)
Measured Output, M
R(k-2)
M(k)
u(km-1)
Control Action, u
p(k-2)
M(k-2)
uk
m move horizon h prediction horizon k
current simulation time step
?U
k
k1
k2
km-1
kh
Horizon
60
Tailored Standard
Patient 20
Patient 405

• Early treatment, frequent titration are key
• Measurement frequency congruent with the time
  scale of the process being modulated is key

61
The inverse problem
62
Observe Low Blood pressure
Observe Blood pressure
Normal
Give a fluid challenge
Mid-range
Low
time
63
Model reduction Bottom-up models

• Map lumped observables to biologically relevant
  functions/modules
• Component aggregation
• Identify aggregates (SVD PCA)
• Time aggregation
• Spatial segregation
• Data-driven vs. Knowledge-driven

64
Model reduction Top-down models

• Function vs components may be biased
• What we think we know
• What we can measure
• Which parts of the model can/should be abstracted
  
• Could this be driven by constraints imposed by
  biological laws?

65
Cogent model reduction

• Biological challenges
• Necessity of increasingly quantitative,
  reductionist research (evidence-based modeling)
• Computational challenges
• Core theoretical challenges
• Lucid conceptualization of the multi-scale
  concept
• Mathematical theory and methods
• Educational challenges

66
Knowledge and successful translation
67
Inflammation Modeling is a Team Sport
68
www.iccai.org
69
Inflammation Modeling is a Team Sport
Developing Trans-disciplinary Mathematical
Modeling Teams for the study of Acute
Inflammation A review of the Experiences of Four
Research Groups Gary An MD (1), C. Antony
Hunt PhD (2), Gilles Clermont MD (3, 6), Edmund
Neugebauer, PhD (4) and Yoram Vodovotz PhD (5,
6) 1) Department of Surgery, Northwestern
University Feinberg School of Medicine 2)
Biosystems Research Group, University of
California, San Francisco 3) Department of
Critical Care Medicine, University of
Pittsburgh 4) Institute for Research in Operative
Medicine, University of Witten/Herdecke 5)
Departments of Surgery and Immunology, University
of Pittsburgh 6) Center for Inflammation and
Regenerative Modeling, McGowan Institute for
Regenerative Medicine, University of Pittsburgh,
Pittsburgh, PA J. Crit. Care, submitted for
publication