Slide 1

1 Computing the electrical activity in the human heart Aslak Tveito Glenn T. Lines, Joakim Sundnes, Bjrn Fredrik Nielsen, Per Grttum, Xing Cai, and Kent Andre Mardal Simula Research Laboratory/ The University of Oslo

Slide 2

2 Motivation/Background The mathematical model Solution strategy Solving linear systems CPU-models Results Future work

Slide 3

3 For each heartbeat, electrically charged atoms (ions) move in and out of the heart muscle cells in a complicated pattern.The most important ions are Na +, K +, Ca ++, and Cl -. The movement of the ions cause the muscle fibers to contract and the heart to pump. Normal electrical activity is important for the pumping function of the heart. The electrical activity of the heart may be recorded on the surface of the body. The recording is called an electrocardiogram (ECG). K+K+ Na + K+K+ K+K+ K+K+ K+K+ The heart

Slide 4

4 ECG 1887 The first ECG is recorded in London on Augustus Wallers dog Jimmy. 1911 A commercial ECG machine is constructed by Willem Einthoven. 1943 The lead positions are standardized. 2000 Worldwide, approximately 1 million ECGs are recorded every day.

Slide 5

5 Heart infarctions 20% of deaths in the western world are due to hearth infarctions and consequences thereof. About 50% of patients admitted to surveillance units with acute chest pain suffer from a heart infarction. ECG is the most important tool to diagnose heart infarction and heart cramp (angina). In some parts of the heart, the sensitivity of ECG is as low as 60%. Normal interpretation of ECG only gives crude estimates: large infarction, small infarction.

Slide 6

6

Slide 7

7 Computational domain sketch in 2D

Slide 8

8 Cardiac electrical activity The heart consists of billions of electrically charged cells. During a heartbeat these cells leak ions, which changes the polarity of the cells (depolarization). The electrocardiogram records this process on the body surface. Pathological conditions in the heart can be diagnosed from abnormal deviations in the ECG.

Slide 9

9 Electrical potential in the torso J~ current destiny u~ electrical potential M~ conductivity tensor Conservation of current, and Ohms law gives and, in addition

Slide 10

10 Electrical potential in the heart Intracellular (space within the cells) J i ~ current density u i ~ electrical potential M i ~ conductivity tensor q i ~ electrical charge Extracellular (space outside the cells) J e ~ current density u e ~ electrical potential M e ~ conductivity tensor q e ~ electrical charge Transmembrane potential v=u i - u e

Slide 11

11 Conservation laws for intra and extra cellular domains: where I ion models the ionic current.

Slide 12

12 From these two equations, the BiDomain model is derived The BiDomain model was developed by Gesolowitz, Miller, Schmilt, and Tung in the early 70s. The equations have been studied by a series of researchers (Colli Franzone et al, Henriques, Trayanova et al, Huang et al, )

Slide 13

13 Model for ionic current (I ion ) The appropriate model for the ionic current will depend upon the application. In the simplest type is is just a function of v: In more realistic models it depends on several factors: where s is a vector of variables including concentrations of ions and permeability of ion channels. These variables are typically governed by an ODE system:

Slide 14

14 The BiDomain Model

Slide 15

15 Boundary conditions

Slide 16

16 Cell models of ionic currents Hodgkin, huxley Noble Beeler, Reuter DiFrancesco, Noble Luo, Rudy Winslow et al

Slide 17

17

Slide 18

18 The BiDomain Model

Slide 19

19 Operator splitting Split in two parts with appropriate boundary conditions.

Slide 20

20 Consider the linear problem with appropriate boundary conditions.

Slide 21

21

Slide 22

22 The discretization leads to a linear system of equations on the form where A is a matrix, u h is the FEM-solution, and f h is the discretized version of f. Let N be the number of grid points. We solve this system using the CG-method and the MG-method. The stopping criterion is in a relative l 2 -norm on the residual.

Slide 23

23 Numerical experiments on an IBM-RS6000 show that the CPU-time needed by the CG and UG methods are and Suppose we want to solve (2) on a grid relevant for the heart.

Slide 24

24 Numerical experiments indicate that we need In the heart. So for 1cm, we need 50 points. In 1cm, we need Suppose the heart is 300cm, then we need

Slide 25

25 Solving one linear system of the form (2) using the CG method requires Which is about 7.5 hours. The MG method requires about or about 7 minutes. We need order optimal methods!

Slide 26

26 Preconditioners The number of iterations for the CG method is bounded by Since, for elliptic problems, the convergence is very slow as seen above. However, we may solve instead of

Slide 27

27 The goal of preconditioning is to find B such that is independent of h and where can be solved in O(N) operations. Then the number of iterations needed to solve (5) is independent of h and thus the entire solution process uses only O(N) operations.

Slide 28

28 It can be shown that if there exist constants c 0, c 1 independent of h, such that then the condition number of is bounded independently of h.

Slide 29

29 Analysis of a BiDomain preconditioner Define The system (7) becomes

Slide 30

30 The Crank-Nicholsons scheme Weak form Define

Slide 31

31

Slide 32

32 Linear operators

Slide 33

33

Slide 34

34 Spectral equivalence Two operators, T n and S n are spectrally equivalent if there are constants c 1 and c 2, independent of h, such that If T n and S n (V n gV n ) are -symmetric and positive definite, then

Slide 35

35 Preconditioner Define (from (12)) and, the preconditioner Our aim is to prove that where c is independent of h and rt.

Slide 36

36 Assumptions We assume that there are constants independent of h and rt such that The last assumption simply state that the matrix generated by the Laplace operator is spectrally equivalent to the matrices generated by the variable coefficient problems. Note that, since we have from (18)

Slide 37

37 Upper bound We start by proving that for a suitable constant c.

Slide 38

38

Slide 39

39

Slide 40

40

Slide 41

41 Lower bound

Slide 42

42

Slide 43

43 From (22) and (23), we have so

Slide 44

44 Numerical experiments We want to compare the numerical solution of the Bidomain equations using CG without preconditioning, and with the optimal preconditioner. Matrix Preconditioner

Slide 45

45 The table shows average number of iterations and CPU-time using 10 time-steps (fixed rt). Number of nodes No preconditioningOptimal preconditioning Iterations CPU-timeIterations CPU-time 25.66552023.82 2016.6 102.0171061181.94 2278.91 406.78521251453.63 22318.97 1.627.140 22

Slide 46

46 Models of CPU-time as a function of the number of computational nodes, N: With we expect

Slide 47

47 Parallel solution using Domain Decomposition Number of nodes: Number of unknowns per timestep: CPU-measurements for one time-step

Slide 48

48 How fast can we now solve the entire problem?

Slide 49

49