clinical cardiovascular identification with limited data and fast forward simulation
DESCRIPTION
Clinical cardiovascular identification with limited data and fast forward simulation. 6 th IFAC SYMPOSIUM ON MODELLING AND CONTROL IN BIOMEDICAL SYSTEMS C. E. Hann 1 , J. G. Chase 1 , G. M. Shaw 2 , S. Andreassen 3 , B. W. Smith 3 - PowerPoint PPT PresentationTRANSCRIPT
Clinical cardiovascular identification with Clinical cardiovascular identification with limited data and fast forward simulationlimited data and fast forward simulation
6th IFAC SYMPOSIUM ON MODELLING AND CONTROL IN BIOMEDICAL SYSTEMS
C. E. Hann1, J. G. Chase1, G. M. Shaw2, S. Andreassen3, B. W. Smith3
1Department of Mechanical Engineering, University of Canterbury, Christchurch, New Zealand2 Department of Intensive Care Medicine, Christchurch Hospital, Christchurch, New Zealand3 Centre for Model-based Medical Decision Support, Aalborg University, Aalborg, Denmark
• Problem: Cardiac disturbances difficult to diagnose– Limited data
– Reflex actions
• Solution: Minimal Model + Patient-Specific Parameter ID– Interactions of simple models to create complex dynamics
– Primary parameters
– Identification must use common ICU measurements
– E.g. increased resistance in pulmonary artery pulmonary embolism, atherosclerotic heart disease
• However: Identification for diagnosis requires fast parameter ID– Must occur in “clinical real-time”
– Limits model and method complexity (e.g. parameter numbers, non-linearities, …)
Diagnosis and TreatmentDiagnosis and Treatment
Heart ModelHeart Model
D.E.’s and PV diagramD.E.’s and PV diagram
2
2232
2
1
1121
1
21
L
RQPPQ
L
RQPPQ
QQV
2
0
)375.0(80
)(02
)(
),1())(1()()(
t
VVdes
ete
ePteVVEteP
Fast Forward SimulationFast Forward Simulation
• Formulate in terms of Heaviside functions:
• No looking for sign changes and restarting DE solver
done automatically
• E.g. filling (P1>P2, P2 ):
(close on flow)
2
22322322
1
11211211
2211
)()5.0)()((
)()5.0)()((
)()(
L
QRPPQHPPHHQ
L
QRPPQHPPHHQ
QQHQQHV
0)(,1
0)(,0))((
tK
tKtKH
)(
1tan)(tan
1
2
1))(( 11
tKtKtKH
0,,0
0,,1)5.0)()((
112
112121
QPP
QPPQHPPHH
Fast Forward SimulationFast Forward Simulation
• Ventricular interaction solve every time step (slow)
Substitute:
• Linear in Vspt Analytical solution very fast
)1())(1()()(
)1())(1()()(
)1())(1()()(
)(,0,
)(,0,
)(,0,,
,0
,0
,0
sptrvrvf
sptlvlvf
sptsptspt
VVrvfsptrvrvfes
VVlvfsptlvlvfes
VVsptsptdsptsptes
ePteVVEte
ePteVVEte
ePteVVEte
rvfrvrvfV
lvflvlvfV
sptsptsptV
bVaebVaebVae rvrvflvlvfsptspt ,,
Method (20 Heart beats) CPU time (s) Speed increase factor
Event solver 101.9
First Heaviside 36.8 2.8
Simpler Heaviside 18.8 5.4
Simpler Heaviside + Vspt formula 3.1 32.9
Reflex ActionsReflex Actions• Vaso-constriction - contract veins
• Venous constriction – increase venous dead space
• Increased HR
• Increased ventricular contractility
Varying HR as a linear function of Pao
Simple interactions to create overall complex dynamic behaviour in the full system model
Disease StatesDisease States
• Pericardial Tamponade:– Build up of fluid in pericardium– Decrease: dead space volume V0,pcd
• Pulmonary Embolism:– Increase: Rpul
• Cardiogenic shock:– Not enough oxygen to myocardium (e.g. from blocked coronary artery)– Decrease: Ees,lvf, Increase: P0,lvf A more complex set of changes/interactions
• Septic shock:– Blood poisoning– Decrease: Rsys = systemic resistance
• Hypovolemic shock:– Severe decrease in total blood volume = sum of individual volumes
Current Status:• Clinical results for Pulmonary Embolism• Simulated results for others
A Healthy Human BaselineA Healthy Human Baseline
Output Value
Volume in left ventricle 111.7/45.7 ml
Volume in right ventricle 112.2/46.1 ml
Cardiac output 5.3 L/min
Max Plv 119.2 mmHg
Max Prv 26.2 mmHg
Pressure in aorta 116.6/79.1 mmHg
Pressure in pulmonary artery 25.7/7.8 mmHg
Avg pressure in pulmonary vein 2.0 mmHg
Avg pressure in vena cava 2.0 mmHg
Simulation of Disease StatesSimulation of Disease States
• Pericardial Tamponade Ppu – 7.9 mmHgCO – 4.1 L/minMAP – 88.0 mmHg
• Pulmonary Embolism: Rpul increases as embolism induced
• All other disease states similarly capture physiological trends and magnitudes
~60-75% change in mean value
Integral Method - ConceptIntegral Method - Concept
Discretised solution analogous to measured data
• Work backwards and find a,b,c
• Current method – solve D. E. numerically or analytically
8.0,2.0,5.0
1)0(,)sin(
cba
xctbaxx• (simple example with analytical solution )
R
PPq 21
12 13 14 15 16 17 18 191.4
1.45
1.5
1.55
1.6
1.65
1.7
1.75
1.8
1.85
time
x
))sincos(
)(()1(
1)(
22
322
ccatbatab
acaabcaeaa
tx at
- Find best least squares fit of x(t) to the data
- Non-linear, non-convex optimization, computationally intense
• integral method
– reformulate in terms of integrals
– linear, convex optimization, minimal computation
Integral Method - ConceptIntegral Method - Concept
0t• Integrate both sides from to t ( ) ,)sin( ctbaxx 4
0t
t
t
t
t
t
t
t
t
t
t
t
t
ttcttbdtxatxtx
dtcdttbdtxatxtx
dtctbaxdtx
0
0 0 0
00
)())cos()(cos()()(
1)sin()()(
))sin((
000
0
• Choose 10 values of t, between and form 10 equations in 3 unknowns a,b,c
40t 6
10,,1),()()())cos(1( 000
itxtxttctbdtxa iitt i
Integral Method - ConceptIntegral Method - Concept
)()(
)()(
)cos()cos(
)cos()cos(
010
01
010100
0110
10
0
1
0
txtx
txtx
c
b
a
ttttdtx
ttttdtx
tt
tt
• Linear least squares (unique solution)
Method Starting point CPU time (seconds) Solution
Integral - 0.003 [-0.5002, -0.2000, 0.8003]
Non-linear [-1, 1, 1] 4.6 [-0.52, -0.20, 0.83]
Non-linear [1, 1, 1] 20.8 [0.75, 0.32, -0.91]
• Integral method is at least 1000-10,000 times faster depending on starting point
• Thus very suitable for clinical application
Adding Noise & IDAdding Noise & ID
• Add 10% uniform distributed noise to outputs to identify
• Conservative value that is larger than what has been encountered clinically
• Sample at 100-200 Hz
• Capture disease states, assume Ppa, Pao, Vlv_max, Vlv_min, chamber flows.
Identifying Disease State Using Identifying Disease State Using All Variables - SimulatedAll Variables - Simulated
Change True value
(ml)
Optimized Value
Error (%)
First 180 176 2.22
Second 160 158 1.25
Third 140 138 1.43
Fourth 120 117 2.50
Fifth 100 100 0
• Pericardial Tamponade: Find all parameters, look for change in V0,pcd
• Pulmonary Embolism: Find all parameters, look for change in Rpul
Change True value
(mmHg s ml-1)
Optimized Value Error (%)
First 0.1862 0.1907 2.41
Second 0.2173 0.2050 5.67
Third 0.2483 0.2694 8.50
Fourth 0.2794 0.2721 2.60
Fifth 0.3104 0.2962 4.59
• All other variables effectively unchanged from baseline
• All other variables effectively unchanged from baseline
• Cardiogenic shock: Find all parameters, look for change in [Ees,lvf, P0,lvf]
Change True values Optimized Value
Error (%)
First [2.59,0.16] [2.61,0.15] [0.89,5.49]
Second [2.30,0.19] [2.30,0.18] [0.34,4.39]
Third [2.02,0.23] [2.02,0.21] [0.43,8.03]
Fourth [1.73,0.26] [1.70,0.24] [1.48,9.85]
Fifth [1.44,0.30] [1.43,0.27] [0.47,9.39]
Change True value
(mmHg s ml-1)
Optimized Value
Error (%)
First 1.0236 1.0278 0.41
Second 0.9582 0.9714 1.37
Third 0.8929 0.8596 3.73
Fourth 0.8276 0.8316 0.49
Fifth 0.7622 0.7993 4.86
• Septic Shock: Find all parameters, look for change in Rsys
Identifying Disease State Using Identifying Disease State Using All Variables - SimulatedAll Variables - Simulated
• All other variables effectively unchanged from baseline
• All other variables effectively unchanged from baseline
• Hypovolemic Shock: Find all parameters, look for change in total stressed blood volume
Change True value
(ml)
Optimized Value
Error (%)
First 1299.9 1206.5 7.18
Second 1177.3 1103.7 6.26
Third 1063.1 953.8 10.28 (hmm!)
Fourth 967.8 1018.9 5.28
Fifth 928.5 853.4 8.10
Identifying Disease State using Identifying Disease State using All Variables - SimulatedAll Variables - Simulated
• All other variables effectively unchanged from baseline
New for MCBMSNew for MCBMS! Clinical Results! Clinical Results
• Pulmonary embolism induced in pigs (collaborators in Liege, Belgium)
• Identify changes in Rpul including reflex actions and all variables
Left Ventricle (30 mins) Right Ventricle (30 mins)
• Use only: Pao, Ppa, min/max(Vlv, Vrv) to ID all parameters
• Far fewer measurements than in simulation
Animal Model Results – PV loopsAnimal Model Results – PV loops
Left Ventricle (120 mins)
Right Ventricle (120 mins)
Left Ventricle (180 mins)
Right Ventricle (120 mins)
Animal Model Results Over TimeAnimal Model Results Over Time
SeptumVolume(Pig 2)
Right ventricleexpansion index(Pig 2)
Pulmonary resistance(Pig 2)
Reflex actions(Pig 2)
PV loops (Pig 2)
Rpul – All pigs
ConclusionsConclusions
• Minimal cardiac model simulate time varying disease states– Accurately captures physiological trends and magnitudes
– Accurately captures a wide range of dynamics
– Fast simulation methods available
• Integral-based parameter ID patient specific models– Simulation: ID errors from 0-10%, with 10% noise
– Animal model: Pressures (Total Error) =2.22±2.17 mmHg (< 5%)
Volumes (Total Error) = 2.37±2.01 ml (< 5%)
• Identifiable using a minimal number of common measurements– Rapid ID method, easily implemented in Matlab
– Rapid ID = Rapid diagnostic feedback
• Future Work = septic shock (Dec 2006), and other disease states (2007)
AcknowledgementsAcknowledgements
Questions ???
Engineers and Docs
Dr Geoff Chase Dr Geoff Shaw
The Danes
Prof Steen Andreassen Dr Bram Smith
The Honorary Danes
Belgium Honorary Kiwis
Dr Thomas Desaive
Christina Starfinger