Você está na página 1de 37

Accepted Manuscript

Efficient Probabilistic Finite Element Analysis of a Lumbar Motion Segment

Dana J. Coombs, Paul J. Rullkoetter, Peter J. Laz

PII: S0021-9290(17)30361-5
DOI: http://dx.doi.org/10.1016/j.jbiomech.2017.07.002
Reference: BM 8290

To appear in: Journal of Biomechanics

Accepted Date: 3 July 2017

Please cite this article as: D.J. Coombs, P.J. Rullkoetter, P.J. Laz, Efficient Probabilistic Finite Element Analysis
of a Lumbar Motion Segment, Journal of Biomechanics (2017), doi: http://dx.doi.org/10.1016/j.jbiomech.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
An original article

Submitted to Journal of Biomechanics

Efficient Probabilistic Finite Element Analysis of a

Lumbar Motion Segment

Corresponding Author:
Coombs, Dana J.
DePuy Synthes
1301 Goshen Parkway
West Chester, PA 19380, USA

Rullkoetter, Paul J.
Center for Orthopaedic Biomechanics
University of Denver
2390 South York Street
Denver, CO 80208

Laz, Peter J.
Center for Orthopaedic Biomechanics
University of Denver
2390 South York Street
Denver, CO 80208

Keywords: finite element analysis, lumbar, disc, nucleus pulposis, annulus fibrosis,
Holzapfel, Gasser, Ogden, Mooney Rivlin, spinal ligament, calibration, probabilistic,
Monte Carlo, Descriptive sampling, Sobol sampling, functional spinal unit

Abstract Word Count: 250

Word Count (Introduction-Discussion): 3,498

Page 1 of 30

Finite element models of the lumbar spine are useful in assessing biomechanics

and performance of implants. Models are often developed using the anatomy of an

individual subject. Average mechanical property values for the annulus and other soft

tissue structures are typically utilized from the literature, as data for the same subject

are not available. However, these properties can have significant variability. While

probabilistic methods enable the impact of soft tissue property variability on spine

mechanics to be assessed, they often require lengthy computation times. Accordingly,

the objective of this study was to develop efficient methods to perform Monte Carlo

simulations of a finite element model of the L4 L5 functional spinal unit considering

variability in the properties of the soft tissue structures. Distributions for the soft tissue

properties included the stiffness of spinal ligaments and parameters of a Holzapfel-

Gasser-Ogden constitutive material model of the disc. Variance reduction sampling

methods, including the Sobol and Descriptive sampling techniques, were assessed for

efficiency and accuracy in comparison to traditional random Monte Carlo sampling.

Comparisons were based on output torque-rotation curves at the 10th and 90th

percentile for flexion, extension, axial rotation, and lateral bending. The Descriptive

sampling technique best matched the random sampling technique, at the extremes of

rotation, with a 3.6% mean difference. This was achieved with a 10X reduction in the

number of iterations and computation time. Improvements in efficiency and maintained

accuracy enable intersubject variability to be considered in a variety of biomechanical

evaluations, including design-phase screening of orthopedic implants.

Page 2 of 30

Mechanics of the spine are important to evaluate pathology, such as disc

degeneration, and implant performance through comparisons of healthy, diseased and

repaired spines. Experimental tests provide direct measurements, but are expensive

and time consuming. Computational modeling is a complimentary way to evaluate spine

mechanics (Ayturk et al., 2011; de Visser et al., 2007; Eberlein et al, 2004; Ezquerro et

al., 2004; Ezquerro et al., 2011; Guan et al, 2006; Lu et al., 1996; Schmidt et al., 2007;

Wong et al., 2003). For example, a subject-specific model provides a consistent method

to compare implant performance (Bono et al., 2007; Bowden et al., 2008; Chiang et al.,

2006; Dooris et al., 2001; Goel et al., 2005; Polikeit et al., 2003; Rohlmann et al., 2005;

Tsuang et al., 2009; Vadapalli et al., 2006; Xiao et al., 2012; Zhong et al., 2009). While

both models and experiments assess torque-rotation behavior, some measures can be

extracted from a model that cannot be easily measured experimentally including bone

strain, facet contact force, disc pressure, and annulus fibrosis (AF) strain. Deterministic

models are typically based on anatomy generated from medical image data. In some

cases, the soft tissue representation is based on tissue tests from the same specimen,

but in most cases, the ligaments are defined as the average values reported in the

literature (Ayturk et al., 2011; Guan et al., 2006).

Computational models can also be used in probabilistic analysis to represent a

population of subjects. In general, probabilistic approaches represent input parameters

as distributions and predict output distributions and bounds of performance, while also

identifying the sensitivity factors indicating which input parameters, or combination of

Page 3 of 30
parameters, are most influential. Previous studies have applied Monte Carlo (MC)

simulation to capture the variability in spine biomechanics based on the variability of

the soft tissue properties and anatomy (Lee et al., 2005; Barnes et al., 2011). Modeling

has also considered the failure of a vertebral body based on the variability of stress

measures in the bone (Ahman et al., 2010; Rohlmann et al., 2010). In addition,

probabilistic studies have assessed the mechanics associated with total disc

replacement (TDR) with variability in implant alignment in the disc space and key

geometric features of the TDR (Rohlmann et al., 2009).

As the traditional MC method uses a random sampling technique, it typically

requires 100s to 1000s of trials to reach convergence and is thereby computationally

expensive. Other variance reduction sampling techniques, such as Descriptive sampling

(Isight Component Guide, Saliby 1990) and Sobol Sampling (Burhenne et al., 2011, Isight

Component Guide) can predict output distributions with similar accuracy in fewer

iterations. Efficient probabilistic methods have been demonstrated in a cervical spine

model to assess mechanical property uncertainty on injury mechanisms (Thacker et al.,


Probabilistic FE modeling provides opportunities to understand the impact of

input parameters, and can extend the findings for a few subjects to better represent the

population. Recently, Coombs et al. (2016) performed a comprehensive summary of the

experimental literature and calibrated the material properties for models of tissue-level

specimens in order to establish a probabilistic representation of the AF and spinal

ligaments. The objective of this study was to demonstrate efficient methods to perform

Page 4 of 30
probabilistic predictions of spine mechanics considering variability in the properties of

soft tissue structures using Monte Carlo simulations of a finite element model of the L4

L5 functional spinal unit. Variance reduction sampling methods, including the Sobol and

Descriptive Sampling techniques, were assessed for efficiency and accuracy in

comparison to traditional random Monte Carlo sampling. Comparisons were based on

output torque-rotation curves at the 10th and 90th percentile for flexion, extension,

axial rotation, and lateral bending.


2.1 Finite Element Model

A specimen-specific finite element (FE) model was developed in Abaqus (Simulia,

Dassault Systemes, Johnston, RI), using an implicit solver, for the L4-L5 FSU from a

healthy 33 year old male with no evidence of disc degeneration (Fig. 1) (Rao, 2012,

Coombs et al., 2013). The vertebral geometry was segmented from a CT scan using

ScanIP (Simpleware, Exeter, UK) and bones were represented by 3-noded triangular

rigid elements (Element type=R3D3). The AF was meshed with 8-noded hexahedral

elements (Element type=C3D8R) and the nucleus pulposis (NP) was represented with an

8-noded fluid-filled membrane (Element type= SFM3D4R). The AF was modeled using

the Holzapfel-Gasser-Ogden (Holzapfel et al., 2000) anisotropic, hyperelastic

constitutive model. This constitutive model was chosen because the embedded fibers

are mathematically represented, which allows parameters defining fiber stiffness,

direction and dispersion, to be modified. The AF was divided into four quadrants

Page 5 of 30
representing an anterior, posterior, and right and left lateral quadrants due to the

variation in material properties in the AF (Kurtz et al., 2006). The orientation of the

elements in the AF was defined with the normal direction of the outer surface of the

disc as the radial direction, the anatomical inferior/superior axis, and the remaining

tangential direction derived via cross-product. The superior and inferior surfaces of the

disc were kinematically coupled to a reference point at the center of the L4 and L5

vertebral bodies, respectively. Because the vertebrae were defined as rigid, this results

in the same behavior as a tied constraint between the disc surface and the vertebral

endplates. The articulating facet surfaces were rigid and represented by 8-noded

hexahedral elements (Element type=C3D8R) using linear pressure over closure contact

behavior to capture the compliance of the facets to improve computational efficiency

(Rao et al., 2009). Seven passive ligaments were defined in the model using non-linear

tension only connector elements. The ligaments included the anterior longitudinal

ligament (ALL), posterior longitudinal ligament (PLL), supraspinous ligament (SSL),

intraspinous ligament (ISL), intertransverse ligament (ITL), facet capsular ligament (FCL)

and ligamentum flavum (LFL). The ligament attachment sites were based on dissection

performed after testing and agreed with literature-based descriptions (Panjabi et al.,


The test frame applied pure moments of 10Nm at the L4 vertebrae while

allowing passive translation at the L5 vertebrae. Rotation was measured using a Vicon

motion capture system (Vicon, Denver, CO). The resulting measurements were 5.92 of

extension, 8.66 of flexion, 6.01 of lateral bending, and 1.75 of axial rotation (Rao,

Page 6 of 30
2012; Coombs et al., 2013). These rotations were applied to the model and the reaction

moments were measured and compared.

2.2 Probabilistic Representation of Soft Tissue Structures

Probabilistic representations of the ligament properties were developed from a

comprehensive summary of direct mechanical test data; distributions of stiffness for the

7 spinal ligaments were defined with mean and standard deviation and used as input in

the current study (Table 1). The data compiled were focused at the L4 L5 level;

however, data from other levels were used in cases when the L4 L5 data were not

available. To minimize the number of parameters used in a Monte Carlo simulation, a

single stiffness parameter was defined for the stiffness of the linear load displacement

behavior for each ligament. Two inflection points were defined to represent the toe in

region based on the mechanical test data found in literature. The location of the points

was defined as a function of the stiffness parameter (Figure 1, Table 2).

Properties for the AF were defined for the four quadrants; anterior, posterior,

and left and right lateral. The left and right lateral quadrants were assumed to be

identical. Properties for the AF regions were defined as probabilistic representations

and based on a comprehensive summary of the experimental literature and calibrated

models of tissue-level specimens. Probabilistic analyses were performed to establish

the distribution of parameters for the Holzapfel-Gasser-Ogden anisotropic, hyperelastic

constitutive model (Holzapfel et al., 2000) to match the distribution found in the

Page 7 of 30
literature summary. The Holzapfel-Gasser-Ogden model is defined by five parameters

(C10, D, k1, k2, ) with the following strain energy function.

C10 affects the stiffness of the ground substance. k1 affects the stiffness of the fibers in

the material. k2 determines the nonlinearity of the fiber stiffness (Abaqus Theory Guide

Chapter 4.6.3, release 6.12). This is a phenomenological constitutive model and the

fiber stiffness values do not represent individual fibers or indicate the number of

discrete fibers. The compressibility parameter, D, was defined as 1/(20*C 10), which is

equivalent to defining the initial bulk modulus as twenty times the initial shear modulus.

Therefore, the compressibility parameter is nearly incompressible, as it is similar to a

Poissons ratio of 0.475 (Abaqus Users Guide Chapter 22.5.1, release 6.12). The fiber

dispersion parameter, , was set to 0 assuming the fibers are perfectly aligned to the

fiber direction. As a result, three input parameters (C10, k1, k2) defined the Holzapfel-

Gasser-Ogden relationship for each quadrant. The best agreement to literature data

was a lognormally distributed C10 parameter and normally distributed k1 and k2

parameters. Mean and standard deviations for the parameters are reported for each

quadrant (Table 1).

An initial assessment of the FE model was performed using mean values of the

soft tissue parameters and compared the torque rotation behavior and range of motion

to experimental data reported in literature.

Page 8 of 30
2.3 Probabilistic Methods

Monte Carlo simulation methods were implemented in Isight (Simulia, Dassault

Systemes, Johnston, RI) to determine the uncertainty of the torque-rotation curves for

flexion, extension, axial rotation, and lateral bending based on the uncertainty of the

soft tissue parameters. The random sampling technique with 500 iterations established

the baseline. Sampling error and convergence were calculated to assess the quality of

the predictions. To assess the number of trials required to achieve convergence, the

Monte Carlo simulation was evaluated by calculating the cumulative 10th and 90th

percentile of the reaction moment at the end of the applied rotation and then

calculating the percent error between the current iteration cumulative percentile and

the previous iteration.

The Descriptive and Sobol variance reduction techniques were evaluated. The

Descriptive Sampling technique is analogous to the Latin Hypercube technique in that

the space defined by each random input parameter is divided into subsets of equal

probability and the analysis is performed with the value for each subset of each random

parameter selected only once. The Sobol sampling technique uses a quasi-random

sequence to generate samples of input parameters more uniformly than random and

Descriptive Sampling while considering previously sampled points to avoid clusters and

gaps (Burhenne et al., 2011). The Sobol sequence generates numbers as binary

fractions of appropriate length from a set of special binary fractions. The Descriptive

Sampling and Sobol Sampling techniques were used with 50 and 25 iterations and

compared to the random sampling technique. Comparisons between methods were

Page 9 of 30
performed for the 10th percentile and 90th percentile torque rotation curves. Error was

computed as the sum of the squared differences in reaction moment between the

efficient prediction and the baseline results established with 500 iterations of random

sampling. The error was calculated using 20 evenly distributed points along the torque-

rotation curves for each of the applied rotations.

2.4 Sensitivity Analysis

A sensitivity study was also performed to identify which input parameters most

affected the reaction torque. Initially, the Parameter Study functionality in Isight was

utilized; this analysis independently perturbed each parameter at several intervals to

assess its impact on the output. Further, correlations, a measure of sensitivity, were

calculated for the 500 random sampling data points using the Pearson Product-Moment

Correlation Coefficient. Visualizing the sampling data also identified that some of the

relationships between the input parameters and the reaction torques were nonlinear.


3.1 Comparison of Predicted Range of Motion to Experimental Data

The FE model with ligament and disc parameters defined at their mean values fit

well within the ranges of motion reported in the literature. Guan et al. (2006) reported

torque rotation curves and their statistics for 10 L4-L5 FSUs (50.6 +/- 13.2 years old,

max 68, min 27) loaded with a pure moment in flexion, extension, lateral bending and

axial rotation. Campbell et al. (2011) reported similar data with 9 L4-L5 FSUs (mean

Page 10 of 30
65.5 years old, max 75, min 48). Qualitatively, the model falls between the

experimental ranges of +/- 1 standard deviation for all degrees of freedom with the

exception of extension. This is also true for the subject specific test data measured from

the specimen that was used to define the anatomy of the FE model (Fig. 2). Symmetric

behavior was assumed for lateral bending and axial rotation.

Total range of motion was also compared at specified moments. Campbell et al.

(2011) also reported range of motion at 10Nm for flexion, extension, lateral bending,

and axial rotation. Yamamoto et al. (1989) reported similar data based on 10 L4-L5

FSUs (25 to 63 years). Panjabi et al. (1994) reported range of motion at 4Nm for

flexion, extension, and lateral bending based on 9 FSUs (35 to 62 years) and Guan et al.

(2006) reported similar data based on 10 L4-L5 FSUs (mean 50.6 +/- 13.2 years). The FE

data was within 1 standard deviation of a literature source for each motion except for

flexion-extension at 10Nm, which was between the mean valued reported by Campbell

et al. (2011) and Yamamoto et al. (1989) (Fig. 3). There was not a literature source to

compare axial rotation at 4Nm.

3.2 Probabilistic Results

For all degrees of freedom (flexion, extension, axial rotation, and lateral bend),

torque-rotation curves for the FSU at the 10th and 90th percentiles showed that Monte

Carlo simulations with the Descriptive and Sobol sampling methods compared well with

the baseline random sampling method with 500 iterations (Fig. 4). Results for the

Descriptive Sampling and Sobol sampling techniques were based on 25 and 50

Page 11 of 30
iterations. The sum of the squared error (SSE) quantified differences between the

torque-rotation curves for the various methods (Fig. 4, Table 3). The SSE was lower for

certain methods in specific degrees of freedom (e.g. Sobol with 50 iterations for lateral

bending). The Descriptive sampling technique with 50 iterations had the smallest mean

SSE across all degrees of freedoms (2.74 (Nm)2) and for flexion-extension (1.10 (Nm)2)

which describes the largest motions in the spine.

Computation time for each iteration was approximately 70 minutes for all

degrees of freedom performed in parallel, each on a single core. As a result, the MC

simulation with 500 iterations took 24 days, underscoring the need for more efficient

approaches. The 25 and 50 iteration methods took approximately 1.2 days and 2.4

days, respectively. In Monte Carlo simulation with random sampling, the number of

trials and the location (mean or tail) in the distribution influenced the accuracy of the

results. For 500 iterations, confidence intervals bounding the predictions were

computed for the 10th and 90th percentile torque-rotation curves at the maximum

applied rotation (Table 4) and provide a context for error comparison with the variance

reduction methods (Haldar et al., 2000). The convergence error for flexion, extension,

axial rotation, and lateral bending was less than 0.5% at 500 iterations, which supported

the us the results as the baseline for comparison with the variance reduction sampling


3.3 Sensitivity Analysis

Page 12 of 30
The sensitivity analysis identified the key input parameters, and was influenced

by the degree of freedom for the applied rotation (flexion, extension, axial rotation, and

lateral bending) and the relative location/contribution of the soft tissue structures (Fig.

5a). For example, the torque-rotation behavior during flexion was most sensitive to the

stiffness of the SSL and ISL. During extension, the torque-rotation behavior was most

sensitive to the stiffness of the ALL and the C 10 parameter of the anterior disc quadrant.

During lateral bending, the torque-rotation behavior was most sensitive to the stiffness

of the ITL and C10 parameter of the lateral disc quadrant. Input parameters identified

based on correlations from the Monte Carlo simulation generally matched the

parameters from the perturbation study (Fig. 5). However, the fiber angle parameter

had the greatest impact on the torque-rotation behavior for the extension, axial

rotation, and lateral bending rotations. To investigate further, correlations in the input

and output data from the baseline random sampling analysis (500 trials) were computed

and relationships between fiber angle and the reaction torque were investigated. By

plotting the reaction moments as a function of fiber angle (Fig. 6), the point clouds

exhibited trends for extension and axial rotation. For these degrees of freedom, the

relationship is nonlinear with increases in fiber angle leading to increased reaction

moment and a resulting greater sensitivity. The correlations and relationships between

fiber angle and reaction moment were less strong for lateral bending and flexion.


Page 13 of 30
The aim of this study was to assess whether variance reduction techniques, like

the Descriptive or Sobol sampling, could accurately and more efficiently predict the

torque-rotation behavior of a lumbar L4-L5 FSU FE model in comparison to traditional

Monte Carlo simulation with random sampling. This study focused on the variability in

torque rotation behavior due to variability in the soft tissue representations of the AF

and ligaments. A probabilistic representation from Coombs et al. (2016) established the

distributions from a literature review of direct mechanical test data and FE analyses of

test specimens calibrated to the experimental results. Traditionally, FE modeling of the

spine uses average values for soft tissue properties. The use of the probabilistic

representation to represent the soft tissue properties in the predictions of the current

study are particularly relevant as they enable consideration of the impact of intersubject


Baseline in the current study was established with a traditional Monte Carlo

simulation using random sampling with 500 iterations and required approximately 4

days of computation time. Descriptive sampling and Sobol sampling techniques were

shown to reduce the computation time with similar levels of accuracy. The Descriptive

sampling technique with 50 iterations was recommended as the results compared well

to the baseline Monte Carlo simulation (500 trials); the total SSE error was 21.88 (Nm) 2

and required 1.2 days or about a 90% reduction in computation time. A deformable

representation of this model was also run and increased the computation time by 10.

This further demonstrates the need for an efficient sampling method if using a

deformable model.

Page 14 of 30
The probabilistic FE modeling approach uniquely allows investigators to assess

how the experimental variability in small tissue specimens propagates to influence the

mechanics of an FSU or the entire spine. Leveraging the efficient methods, the ability to

quantify bounds of variability on output parameters, e.g. torque-rotation for various

motions, can be useful in establishing ranges for healthy normal populations, assessing

pathologies (e.g. disc degeneration) and informing implant design and surgical

approach. Further, in addition to assessing torque or reaction moment, the impact on

other output measures of interest, such as disc pressure, facet contact or annulus strain,

can be similarly assessed. These measures would need to also be compared to physical

biomechanics data.

Using a sensitivity analysis, the most important variables influencing the specific

outputs were identified for each applied rotation. As expected, the reaction moment

was generally most sensitive to ligaments that had the greatest moment arms

considering the direction of the applied motion. This observation was also true for the

quadrants of the AF. Surprisingly, the reaction moments were most sensitive to the AF

fiber angle in the extension and axial rotation degrees of freedom. The distribution for

fiber angle was derived from Guerin et al. (2006) from direct experimental

measurement. This variability was determined from outer anterior annulus specimens

from 8 discs. The reaction moments were not as sensitive at lower values of fiber angles

because the component of the force generated by the fiber angle resisting the rotation

is smaller. This finding underscores the importance of accurately measuring the fiber

angle when performing tests. Further biomechanical studies to better characterize the

Page 15 of 30
variation present in fiber angle and understand the relationship between AF fiber angle

and the torque rotation response are recommended. Further studies should also be

conducted to understand the sensitivity of fiber angle and fiber stiffness in degenerated

specimen. The Holzapfel Gasser Ogden model is a phenomenological constitutive model

and the parameter values reported in this study reproduce the torque-rotation curves.

This material model was defined with initially nearly incompressible behavior. It is

possible that another set of parameter values with a purely incompressible material

would provide similar load displacement behavior, which could decrease the sensitivity

of the fiber angle.

The finite element model utilized in the study was based on the anatomy of a

single specimen (33 years, male, 59 kg). The specimens anatomy was compared to

anatomy data derived from 157 healthy spines with a mean age of 26.8 years (Gilad et

al., 1986). In general, the specimens L4-L5 anatomy was slightly smaller than the

average reported by Gilad et al. (1986). The majority of the measurements were within

1 standard deviation of the mean (Table 5), with the exception of the anterior/posterior

(A/P) width of the superior L5 endplate. The anterior and posterior FE model disc height

was greater than the mean disc height. Therefore, the disc in the FE model was taller

and narrower than the literature means, causing the disc to be more flexible in flexion

and extension. This observation could explain why the model is less stiff with mean

parameter values than the typical torque rotation curves. Furthermore, the donor used

to define the FE model was 177.8cm tall compared to the mean height of 174.7cm from

Gilad et al. (1986).

Page 16 of 30
In this study, the bony anatomy was approximated as a rigid body. As the model

outputs are focused on kinematics and torque-rotation behavior, they should not be

influenced by the stiffness of the vertebral bodies. Other mechanics-focused studies

have made similar assumptions (Thacker et al., 2001; Coombs et al. 2013). While

variability in the soft tissue properties was considered, future work could include

variability in the bony anatomy, facet cartilage geometry, and even loading conditions.

Statistical shape modeling shows potential in characterizing anatomic variation across a

population and efficiently generating virtual subjects for finite element analysis

(Hollenbeck et al., 2013). The bony anatomy could also be treated as deformable and

the variability in bone stiffness could be included. This would require significant effort

because the modulus of bone spatially varies and needs to be accurately correlated to

three dimensional data such as a CT scan.

A limitation of the probabilistic representation of the AF properties is that they

were based on data from a healthy normal population. Further work could be

performed to characterize the properties of degenerated discs and their variation,

which may be useful when assessing pathology or developing spinal implants. Studies

have shown that disc degeneration influenced the torque-rotation behavior of the FSU;

range of motion generally decreased with more severe degeneration (Rohlmann et al.,


In closing, the implementation of variance reduction sampling methods enabled

probabilistic analyses to be completed in 10% of the time required for the traditional

random sampling method, while maintaining accuracy. The improved computational

Page 17 of 30
efficiency enables consideration of the uncertainty present in the population and can be

employed for timely design phase assessments of spinal implant designs and

treatments, including the evaluation of stresses or strains in an implant, bone or soft

tissue structures, kinematic evaluations of motion preserving implants, and optimal

positioning of the implant in the spinal anatomy.


This study was supported in part by DePuy Synthes. The authors gratefully acknowledge

the model development work performed by Dr. Milind Rao and experimental testing

performed by Dr. Randy Ching at the University of Washington.


Dr. Coombs is an employee of DePuy Synthes.

Page 18 of 30

Abaqus Users Guide (release 6.12), 2012, Simulia, Dassault Systemes, Hyperelastic
behavior of rubberlike materials, Chapter 22.5.1

Ahmad, Z., Akramin, M., 2010. Probabilistic stress analysis of the human lumbar spine
extended finite element method. In Proceedings of the 14th Asia Pacific Regional
Meeting of International Foundation for Production Research, Melaka.

Ayturk, U., Puttlitz, C., 2011. Parametric convergence sensitivity and validation of a finite
element model. Comp. Meth. Biomech. Biomed. Eng. 14, 695-705.

Barnes, K., Armstrong, J., Agarwala, A., Petrella, A., 2011. Probabilistic study of a lumbar
motion segment: sensitivity to material and anatomic variability. In Proceedings of the
ASME 2011 Summer Bioengineering Conference, Parts A and B, Farmington,
Pennsylvania. Paper No. SBC2011-53846, pp. 863-864.

Bono C., Khandha, A., Vadapalli, S., Holekamp, S., Goel, V., Garfin, SR., 2007. Residual
sagittal motion after lumbar fusion: a finite element analysis with implications on
radiographic flexion-extension criteria. Spine 32, 417-22.

Bowden, A., Guerin, H., Villarraga, M., Patwardhan, A., Ochoa, J., 2008. Quality of
motion considerations in numerical analysis of motion restoring implants of the spine.
Clin. Biomech. 23, 536-44.

Burhenne, S., Jacob, D., Henze, G., 2011. Sampling based on Sobol sequences for Monte
Carlo techniques applied to building simulations. In Proceedings of the 12th Conference
of International Building Performance Simulation Association, Sydney.

Campbell, JR., Imsdahl, S., Ching, RP., 2012. Evaluation of a synthetic L2-L5 spine model
for biomechanical testing. In Proceeding of Canadian Biomechanics Society, Vancouver,
BC, Canada.

Chazal, J., Tanguy, M., Bourges, M., Gaurel, G., Escande, G., Guilot, M., Vanneuville, G.,
1985. Biomechanical properties of spinal ligaments and a histological study of the
supraspinal ligament in traction. J. Biomech. 18, 167-176.

Chiang, M., Zhong, Z., Chen, C., Cheng, C., Shih, S., 2006. Biomechanical Comparison of
Instrumented Posterior Lumbar Interbody Fusion with One or Two Cages by Finite
Element Analysis. Spine 31, E682E689.

Coombs, D. J., Bushelow, M., Laz, P. J., Rao, M., Rullkoetter, P. J., 2013. Stepwise
validated finite element model of the human lumbar spine. Proceedings of the ASME

Page 19 of 30
2013 Conference on Frontiers in Medical Devices: Applications of Computer Modeling
and Simulation Washington, DC, USA.

Coombs, D. J., Rullkoetter, P. J., Laz, P. J., 2016. Quantifying variability in lumbar L4 L5
soft tissue properties for use in FE analysis. ASME Journal of Verification, Validation,
and Uncertainty Quantification. J. Verif. Valid. Uncert. 1, 031007.

de Visser, H., Adam, C., Crozier, S., Pearcy, M., 2007. The role of quadratus lumborum
asymmetry in the occurrence of lesions in the lumbar vertebrae of cricket fast bowlers.
Med. Eng. Phys. 29, 877-885.

Dooris, A., Goel, V., Grosland, N., Gilbertson, L., Wilder, D., 2001. Load-Sharing Between
Anterior and Posterior Elements in a Lumbar Motion Segment Implanted With an
Artificial Disc. Spine 26, E122-E129.

Ebara, S., Iatridis, J., Setton, L., Foster, R., Mow, V., Weidenbaum, M., 1996. Tensile
Properties of Nondegenerate Human Lumbar Anulus Fibrosis. Spine, 21, 452-61.

Eberlein, R., Holzapfel, G., Frohlich, M., 2004. Multi-segment FEA of the human lumbar
spine including the heterogeneity of the annulus fibrosus. Comp. Mech. 34, 147-163.

Ezquerro, F., Simon, A., Prado, M., Perez, A., 2004. Combination of finite element
modeling and optimization for the study of lumbar spine biomechanics considering the
3D thoraxpelvis orientation. Med. Eng. Phys. 26, 11-22.

Ezquerro, F., Vacas, G., Postigo, S., Prado, M., Simon, A., 2011. Calibration of the finite
element model of a lumbar functional spinal unit using an optimization technique based
on differential evolution. Med. Eng. Phys. 33, 89-95.

Fujita, Y., Duncan, N., Lotz, J., 1997. Radial tensile properties of the lumbar annulus
fibrosis are site and degeneration dependent, J. Orthop. Res. 15, 814-819.

Gilad, I., Nissan, M., 1986. A study of vertebra and disc geometric relations of the
human cervical and lumbar spine, Spine 11, 154-157.

Goel, V., Grauer, J., Patel, T., Biyani, A., Sairyo, K., Vishnubhatla, S., Matyas, A,. Cowgill,
I., Shaw, M., Long, R., Dick, D., Panjabi, M., Serhan, H., 2005. Effects of Charite artificial
disc on the implanted and adjacent spinal segments mechanics using a hybrid testing
protocol. Spine 30, 2755-2764.

Guan, Y., Yoganandan, N., Moore, J., Pintar, F., Zhanga, J., Maimana D., Laud, P., 2006.
Momentrotation responses of the human lumbosacral spinal column. J. Biomech. 40,

Page 20 of 30
Guan, Y., Yoganandan, N., Zhang, J., Pintar, F., Cusick, J., Wolfla, C., Maiman, J., 2006.
Validation of a clinical finite element model of the human lumbosacral spine. Med. Bio.
Eng. Comp. 44, 633-41.

Guerin, H., Elliott, D., 2006. Degeneration affects the fiber reorientation of human
annulus fibrosus under tensile load. J. Biomech. 39, 14101418.

Haldar, A., Mahadevan, S., 2000. Probability, Reliability and Statistical Methods in
Engineering Design, John Wiley & Sons, Inc., New York, pp. 258.

Hollenbeck, J.F.M., Cain, C., Fattor, J., Fitzpatrick, C.K., Rullkoetter, P.J., Laz, P.J., 2014.
Statistical modeling to characterize disc degeneration in the lumbar spine. 7th World
Congress of Biomechanics, Boston, MA.

Holzapfel, G. A., T. C. Gasser, R. W. Ogden, 2000. A New Constitutive Framework for

Arterial Wall Mechanics and a Comparative Study of Material Models. J. Elasticity 61, 1

Holzapfel, G. A., Schulze-Bauer, C. A., Feigl, G., Regitnig, P., 2005. Single lamellar
mechanics of the human lumbar anulus fibrosus. Biomechan. Model. Mechanobiol. 3,

Iida, T., Abumi, Kotani, Y., Kaneda, K., 2002. Effects of aging and spinal degeneration on
mechanical properties of lumbar supraspinous and interspinous ligaments, The Spine J.,
2, 95100.

Isight Component Guide (release 5.7), Simulia, Dassault Systemes, pp 561-562.

Kurtz, S.M., Edidin, A.A., 2006. Spine Technology Handbook, Elsevier Academic Press,
Burlington, MA.

Lee, K.K., Teo, E.C., 2005. Material sensitivity study on lumbar motion segment (L2-L3)
under sagittal plane loadings using probabilistic method. J. Spinal Disorders &
Techniques 18, 163-170.

Lu, Y., Hutton, W., Gharpuray, V., 1996. Do bending, twisting, and diurnal fluid changes
in the disc affect the propensity to prolapse? A viscoelastic finite element model. Spine
21, 2570-2579.

Neumann, P., Keller, T.S., Ekstrom L., Perry L., Hansson, T. H., Spengler, M., 1992.
Mechanical Properties of the human lumbar anterior longitudinal ligament, J. Biomech.
25, 1185-1194.

Page 21 of 30
OConnell, G., Guerin, H., Elliott, D., 2009. Theoretical and uniaxial experimental
evaluations of human annulus fibrosis degeneration. J. Biomech. Eng. 131, 111007-1 -

Panjabi, M.M., Greenstein, G., Duranceau, J., Nolte, L.P., 1991. Three-dimensional
quantitative morphology of lumbar spinal ligaments. J. Spinal Disorders 4, 54-62.

Panjabi M., Oxland T., Yamamoto I., Crisco J., 1994. Mechanical behavior of the human
lumbar and lumbosacral spine as shown by three-dimensional load-displacement
curves. J. Bone Joint Surg. 76, 413-424.

Pintar, F. A., Yoganandan, N., Myers, T., Elhagediab, A., Sances, A. Jr., 1992.
Biomechanical properties of human lumbar spine ligaments. J. Biomech. 25, 1351-1356.

Polikeit, A., Nolte, LP., Ferguson, J., 2003. The effect of cement augmentation on the
load transfer in an osteoporotic functional spinal unit: finite-element analysis. Spine 28,

Rao, M., 2012. Explicit finite element modeling of the human lumbar spine. PhD.
Thesis, Electronic Theses and Dissertations. Paper 906. University of Denver

Rao, M., Patrella, A., Baldwin, M., Laz, P., Rullkoetter, P., 2009. Efficient Probabilistic
Finite Element Modeling for Evaluation of Spinal Mechanics. In Proceeding of the 55th
Annual Meeting of the Orthopedic Research Society, Las Vegas, NV.

Robertson, D., Willardson, R., Parajuli, D., Cannon, A., Anton, E., 2013. The lumbar
supraspinous ligament demonstrates increased material stiffness and strength on its
ventral aspect. J. Mech. Behav. Biomed. Mater. 17, 3443.

Rohlmann, A., Zander, T., Bergmann, G., 2005. Effect of Total Disc Replacement with
ProDisc on Intersegmental Rotation of the Lumber Spine. Spine, 30, 738-743.

Rohlmann, A., Mann, A., Zander T., Bergmann, G., 2009. Effect of an artificial disc on
lumbar spine biomechanics: a probabilistic finite element study. European Spine J. 18,

Rohlmann, A., Boustani, H., Bergmann, G., Zander, T., 2010. A probabilistic finite
element analysis of the stresses in the augmented vertebral body after vertebroplasty.
European Spine J. 19, 15851595.

Rohlmann, A., Zander, T., Schmidt, H., Wilke, H., Bergmann, G., 2006. Analysis of the
influence of disc degeneration on the mechanical behaviour of a lumbar motion
segment using the finite element method. J. Biomech. 39, 24842490.

Page 22 of 30
Saliby, E., 1990. Descriptive sampling: A better approach to Monte Carlo simulation. J.
Operational Res. Soc. 41, 1133-1142.

Schmidt, H., Kettler, A., Heuer, F., Simon, U., Claes, L., Wilke, H., 2007. Intradiscal
Pressure, Shear Strain, and Fiber Strain in the Intervertebral Disc under Combined
Loading. Spine 32, 748 755.

Thacker, B., Nicolella, D., Kumaresan, S., Yoganandan, N., Pintar, F., 2001. Probabilistic
finite element analysis of the human lower cervical spine. American Society of
Mechanical Engineers, Bioengineering Division (Publication) BED01/2000; 48, 237-238.

Tsuang, Y., Chiang, Y., Hung, C., Wei, H., Huang, C., Cheng, C., 2009. Comparison of cage
application modality in posterior lumbar interbody fusion with posterior
instrumentation - A finite element study. Med. Eng. Phys. 31, 565-570.

Vadapalli, S., Sairyo, K., Goel, V., Robon, M., Biyani, A., Khandha, A., Ebraheim, N., 2006.
Biomechanical Rationale for Using Polyetheretherketone (PEEK) Spacers for Lumbar
Interbody FusionA Finite Element Study. Spine 31, E992-E998.

Wagner, D., Lotz, J., 2004. Theoretical model and experimental results for the nonlinear
elastic behavior of human annulus fibrosus. J. Orthop. Res. 22, 901909.

Wong, C., Gehrchen, P., Darvann, T., Kaer, T., 2003. Nonlinear Finite-Element Analysis
and Biomechanical Evaluation of the Lumbar Spine. IEEE Trans. Med. Imag. 22, 742-746.

Xiao, Z., Wang, L., Gong, H., Zhu, D., 2012. Biomechanical evaluation of three surgical
scenarios of posterior lumbar interbody fusion by finite element analysis. Biomed. Eng.
Online 11.

Yamamoto, I., Panjabi, M., Crisco, T., Oxland, T., 1989. Three-dimensional movements
of the whole lumbar spine and lumbosacral joint. In Proceeding of the 12th Annual
Meeting of the International Congress of Biomechanics, Los Angeles, CA.

Zhong, Z., Chen, S., Hung, C., 2009. Load and displacement controlled finite element
analyses on fusion and non-fusion spinal implants. Proc. Inst. Mech. Eng. H J. Eng.
Med. 223, 143-157.

Page 23 of 30

Figure 1. Finite element model of L4 L5 FSU model with detailed representation of the
disc showing anterior, posterior, and lateral regions of the annulus fibrosis, and a
representative load vs. displacement graph of the ALL with mean and +/- 1 standard
deviation of stiffness.

Figure 2. Comparison of model-predicted torque-rotation behavior using average

parameter values to the literature (Guan et al., 2006; Campbell et al., 2011).

Figure 3. Total range of motion compared to available literature data at 4Nm and 10Nm
for flexion-extension (FE), lateral bending (LB) and axial rotation (AR) degrees of
freedom. Model predictions are based on average parameter values.

Figure 4. Comparison of torque-rotation behavior for Monte Carlo results at 10th and
90th percentile, using random, Descriptive and Sobol sampling.

Figure 5. Parameter sensitivity assessed via parameter perturbation in Isight (a) and
correlations based on Monte Carlo simulation data (b). Results shown for flexion,
extension, axial rotation and lateral bending reaction moments at the maximum range
of motion.

Figure 6. Scatter plots of annulus fiber angle vs. reaction moments for flexion, extension,
axial rotation and lateral bending with Pearson Correlation Coefficients. Each data point
represents one of the 500 Monte Carlo iterations.

Table 1. Summary of input parameters representing ligaments and annulus fibrosis

(Coombs et al., 2016)

Table 2. Equations for Ligament Load vs. Displacement Behavior, Npar = number of
elements in parallel, Nser = number of elements in series, K = stiffness

Table 3. Sum of squared error (Nm)2 comparison of Descriptive and Sobol sampling to
baseline random sampling (500 trials) for flexion-extension, axial rotation, and lateral
bending degrees of freedom

Table 4. Confidence interval (alpha = 0.95) bounds for the reaction moment (Nm) at the
maximum applied rotation. Calculation is based on traditional random Monte Carlo
sampling with 500 trials.

Page 24 of 30
Table 5. Comparison of FSU anatomic measurements between specimen/model and
statistical measurements from Gilad et al. (1986).

Page 25 of 30
Table 1. Summary of input parameters representing ligaments and annulus fibrosis
(Coombs et al., 2016)

Parameter Distribution Mean Standard Source

Neumann et al (1992)
ALL Stiffness (N/mm) Normal 55.39 17.87 Pintar et al. (1992)
Chazal et al. (1985)
Pintar et al. (1992)
PLL Stiffness (N/mm) Normal 31.30 22.48
Chazal et al. (1985)
Pintar et al. (1992)
LFL Stiffness (N/mm) Normal 23.23 8.67
Chazal et al. (1985)
FCL Stiffness (N/mm) Normal 30.60 1.50 Pintar et al. (1992)
ITL Stiffness (N/mm) Normal 35.35 7.20 Chazal et al. (1985)
Pintar et al. (1992)
ISL Stiffness (N/mm) Normal 24.68 15.75 Chazal et al. (1985)
Iida et al. (2002)
Pintar et al. (1992)
Chazal et al. (1985)
SSL Stiffness (N/mm) Normal 20.55 9.96
Iida et al. (2002)
Robertson et al. (2013)
Anterior Annulus Fibrosis Holzapfel-Gasser-Ogden parameters
Guerin et al. (2006)
C10 (MPa) Lognormal 0.0670 0.050
Wagner et al. (2004)
k1 (MPa) Normal 1000.0 500.0 Fujita et al. (1997)
Holzapfel et al. (2005)
OConnell et al. (2009)
k2 Normal 4809.5 2113.9
Ebara et al. (1996)
Posterior Annulus Fibrosis Holzapfel-Gasser-Ogden parameters
C10 (MPa) Lognormal 0.134 0.100 Fujita et al. (1997)
k1 (MPa) Normal 2000.0 1000.0 Holzapfel et al. (2005)
k2 Normal 5296.3 3208.7 Ebara et al. (1996)
Lateral Annulus Fibrosis Holzapfel-Gasser-Ogden parameters
C10 (MPa) Lognormal 0.130 0.100 Fujita et al. (1997)
k1 (MPa) Normal 1500.0 750.0 Holzapfel et al. (2005)
k2 Normal 5849.2 3119.7 Ebara et al. (1996)
Annulus Fibrosis Fiber angle from transverse plane
Normal 25.44 4.22 Guerin et al.(2006)

Page 26 of 30
Table 2. Equations for Ligament Load vs. Displacement Behavior, Npar = number of
elements in parallel, Nser = number of elements in series, K = stiffness

Ligament Displacement Points (mm) Load Points (N)








Page 27 of 30
Table 3. Sum of squared error (Nm)2 comparison of Descriptive and Sobol sampling to
baseline random sampling (500 trials) for flexion-extension, axial rotation, and lateral
bending degrees of freedom.

Axial Lateral
Extension Extension
Rotation Bending Mean
Sampling Percentile Percentile
Percentile Percentile Error
Technique (Nm)2 (Nm)2
(Nm)2 (Nm)2 (Nm)2
10th 90th 10th 90th 10th 90th 10th 90th
Descriptive 2.78
0.09 3.22 0.38 4.81 0.58 10.95 1.93 0.25
25 Iterations
Descriptive 2.74
0.15 1.31 1.88 1.04 4.28 7.98 0.33 4.91
50 Iterations
Sobol 10.10
0.06 6.78 20.07 25.33 2.67 23.48 0.62 1.82
25 Iterations
Sobol 4.77
0.04 8.46 9.18 5.15 0.04 14.16 0.56 0.53
50 Iterations

Page 28 of 30
Table 4. Confidence interval (alpha = 0.95) bounds for the reaction moment (Nm) at the
maximum applied rotation. Calculation is based on traditional random Monte Carlo
sampling with 500 trials.

10th Percentile 90th Percentile

Applied Motion
Upper Bound Lower Bound Upper Bound Lower Bound
Extension 3.31 3.14 9.73 9.22
Flexion -35.39 -37.34 -20.54 -21.67
Axial Rotation 4.13 3.91 22.16 21.00
Lateral Bending 7.10 7.49 11.72 12.37

Page 29 of 30
Table 5. Comparison of FSU anatomic measurements between specimen/model and
statistical measurements from Gilad et al. (1986).

FE Model Gilad et al., 1986

Measurement (mm)
L4 L5 Disc L4 L5 Disc
Inferior A/P width of
32.7 30.5 NA 34.9 2.8 33.9 2.7 NA
Anterior height of vertebral
28.9 26.2 NA 27.4 2.2 28.3 2.1 NA
Superior A/P width of
31.5 32.4 NA 34.3 2.7 34.2 2.7 NA
Posterior height of vertebral
27.3 24.0 NA 27.1 2.3 25.7 2.5 NA
Anterior disc height NA NA 14.1 NA NA 12.0 1.8
Posterior disc height NA NA 8.23 NA NA 7.7 1.5

Page 30 of 30