The modelling framework used in this study integrated gait analysis data and muscle forces computed using a hybrid electromyography (EMG)-assisted method [8, 9] with an anatomically based FE model of the hip complex derived from the visible human dataset and the International Union of Physiological Sciences Physiome Project (Figure 1). This study was conducted in three stages.
Muscle force computation
One healthy male subject (age: 28 years, height: 183 cm, mass: 67 kg) volunteered for this investigation and gave informed, written consent. The project was approved by the Human Research Ethics Committee at the University of Western Australia. The motion data acquired from the subject were static anatomical poses and dynamic gait trials. During all trials, the three-dimensional location of 27 retro-reflective markers placed on the subject's lower extremities, pelvis and trunk was recorded (at 250 Hz) using a 12-camera motion capture system (Vicon, Oxford, UK). During the Dynamic1 trials, ground reaction forces and EMG data were collected (2000 Hz) synchronously with marker trajectories using an in-ground force plate (AMTI, Watertown, USA), and bipolar electrodes with a telemetered EMG system (Noraxon, Scottsdale, USA), respectively. EMG data were collected by placing a pair of electrodes in the correspondence of 16 muscle groups including: hip adductors, gluteus maximus, gluteus medius, gracilis, tensor fasciae latae, lateral hamstrings, medial hamstrings, sartorius, rectus femoris, vastus medialis, vastus lateralis, gastrocnemius medialis, gastrocnemius lateralis, peroneus group, soleus and tibialis anterior.
Both ground reaction forces and marker trajectories were low-pass filtered with a fourth-order Butterworth filter. Cut-off frequencies (between 2 and 8 Hz) were determined by a trial-specific residual analysis . EMGs were processed by band-pass filtering (10–450 Hz), then full-wave rectifying and low-pass filtering (6 Hz). The resulting linear envelopes were normalised with respect to the peak-processed EMG values obtained from the entire set of recorded trials . The human movement data were collected from eight repeated trials of walking at a speed of 1.3 ± 0.25m/s. The gait data was used within an EMG-informed methodology for the prediction of the forces in the musculotendon units (MTUs) crossing the hip, knee and ankle joints [8, 9]. The activity of deeply located MTUs that could not be experimentally measured (iliacus and psoas MTUs) was predicted using a standard static optimization-based approach. Specifically, MTU forces were resolved by minimising the sum of squared MTU activations subject to matching the experimental joint moments at the hip. The amplitude of the experimentally available EMG linear envelopes that were recorded were adjusted (i.e. increased or decreased) to further minimise the discrepancy between experimental and predicted joint moments and account for limitations in EMG processing [8, 9]. Both static optimization and EMG amplitude adjustment were based on a simulated annealing optimization algorithm .
Adjusted and predicted EMG linear envelopes were then used to drive a forward dynamic EMG-driven musculoskeletal model of the human lower extremity [9, 13]. In this model, the musculoskeletal geometry of the human lower extremity was based on a previously presented rigid-body model in which the MTU kinematics was represented by lines of action with MTU-to-bone wrapping points and surfaces [14, 15]. The MTU dynamics was based on a Hill-type muscle model that included non-linear passive tendons in series to parallel non-linear active and passive fibre contractile elements . The musculoskeletal model was comprised of 34 MTUs and six DOF in the lower extremity including: hip flexion-extension, hip adduction-abduction, hip internal-external rotation, knee flexion-extension, ankle plantar-dorsi flexion and ankle subtalar flexion. The proposed musculoskeletal model was scaled and then calibrated to the actual subject [9, 13]. In brief, the scaling procedure used the software opensim (The National Center for Simulation in Rehabilitation Research (NCSRR), Stanford, CA, USA)  to scale a generic model of the human musculoskeletal geometry to match the individual subject's anthropometry. In this process, virtual markers were created and placed on the generic musculoskeletal geometry model on the basis of the position of the experimental markers. The anthropomorphic properties of the anatomical segments and MTUs were then linearly scaled on the basis of the relative distances between experimental markers and their corresponding virtual markers. The adjusted segment and MTU properties included: anatomical segment length, width, depth, center of mass location and mass moment of inertia, as well as MTU insertion, origin and MTU-to-bone wrapping points.
The model calibration step was used to identify a set of intrinsic MTU parameters that vary non-linearly across individuals because of anatomical and physiological differences [9, 13]. These included: tendon slack length, optimal fibre length and maximum isometric force. A nominal set of parameters taken from [14, 16] was initially used in the model to predict the MTU force and the resulting joint moment as a function of EMG signals during a set of calibration trials. The initial parameter set was repeatedly refined using a simulated annealing algorithm  until the mismatch between predicted and experimental joint moments was minimised. The specific hip-crossing muscles modelled, which were used as input to the FE model, were the adductors (brevis, longus and magnus), biceps femoris, gluteus maximus, medius and minimus, gracilis, iliacus, psoas, rectus femoris, sartorius, semimembranosus, semitendinosus and tensor fascia late (Figure 2).
Figure 2. Hip musculotendon forces from electromyography-assisted gait simulation during the stance phase of the gait cycle.
Download figure to PowerPoint
Figure 1(left) shows the geometric model of the pelvis, which was semi-automatically constructed from CT data of a cadaveric specimen  and stored within the Physiome database . These meshes use high-order cubic Hermite elements that are suited to contact mechanics as they allow for the surface normal to be C1 continuous across elements . For the femur mesh, we used an existing model from the Physiome repository and fit this to the cadaveric CT data in two steps. Firstly, using host-mesh fitting , we morphed the Physiome femur to roughly match the cadaveric CT data at key anatomical points (Figure 1 right). Blue landmark points of the femur were pulled towards the target red points. Secondly, the newly morphed femur surface was refit to accurately match the complete CT dataset.
The pelvis model was previously validated using an Instron test as shown in Figure 3 . Specifically, three cadaveric embalmed pelvises were each loaded with 600 N (single leg stance of a 60 kg subject) and strain gauged. Finite element models were then generated from the CT scans. Material properties for cancellous bone were obtained from the CT scans and assigned to the FE mesh using a spatially varying field embedded inside the mesh. In the acetabular region, the model showed good agreement between predicted surface strains and experimental strains with an R 2 of 0.9.
Figure 3. (Top) Cadaveric hip harvested; Instron compression testing within the acetabulum and equivalent finite element model predicted strains with fixed regions highlighted (posterior superior iliac spine and pubic symphysis) and (Bottom) Correlation of finite element analysis predicted strain and experimentally measured strain.
Download figure to PowerPoint
The material properties of the model were adapted from Hounsfield units (HU) [17, 21] from CT as shown at the numerical integration points in Figure 4. The specific relationship between material property and HU is detailed in Appendix A. A ‘St Venant Kirchoff’ model , which is a linear elastic constitutive law that uses the large deformation Green strain tensor, was used for both bone and cartilage. A constant cortical bone stiffness of 16.7 GPa with Poisson's ratio of 0.3 [1, 23]; a spatially varying trabecular bone stiffness ranged from 200 MPa to 2.2 GPa with Poisson's ratio of 0.3 ; and a cartilage stiffness of 10 MPa with Poisson's ratio of 0.4 were used in the model . The pelvis was fixed at the posterior superior iliac spine and the pubic symphysis, which are the attachment points for the two symmetric pelvis halves. The pelvis was loaded via a ground reaction force through the femur with muscle attachments crossing the hip. The femur was left free to resolve for an equilibrium position at each solution step.
Figure 4. Cortical and trabecular material properties from Hounsfield units visualised at finite element integration points.
Download figure to PowerPoint
The muscle forces computed using the EMG-assisted method were applied to the FE model using muscle belly centroid paths. The muscle paths were identified from the visible human male dataset  (Figure 5) and the origin and insertion areas on bone were digitised (Figure 6) with the assistance of an orthopaedic surgeon. These points were then stored in a coordinate frame that is fixed to the meshed bone so they would remain unchanged during morphing and transformation routines. During the gait cycle the action lines of each muscle was used to direct the force onto the pelvis and femur (Figure 5). At the attachment point to the bone, the net force was distributed over the insertion/origin areas (Figure 6). The cartilage was modelled as a uniform layer of elements adjacent to the bone. The high-order cubic Hermite interpolation allowed for non-linear deformations with a single layer of elements.
Figure 6. Muscle contact areas identified from the visible human male and orthopaedic advice. (O) indicates origin and (I) indicates insertion.
Download figure to PowerPoint
The bone and cartilage contact problem was solved using the bioengineering software cmiss (Auckland Bioengineering Institute, Auckland, New Zealand) (www.cmiss.org). A series of finite elastic quasi-static simulations were computed for one cycle of gait. Although bone typically undergoes less than 10% strain (allowing a linear elasticity approximation), the hip cartilage strains are higher and finite elasticity was employed instead. The governing equations for solid mechanics used in the problem are given in Appendix B. The strain and resulting von Mises stress was computed over one gait cycle, and the average was used for this analysis. Remodelling in the hip was based on the strain-based mechanostat approach [26, 27] where strain stimulus less than 1000 μϵ lowers bone density and strain above 3000 μϵ increases bone density. Strains between these values maintain bone density (a homeostatic state). The piecewise stimulus for remodelling, ϵs, is related to the von Mises strain, ϵVM, by
For this study, we used the von Mises scalar measure for bone (cartilage was not included in the analysis). The von Mises strain accounts for all principal components and is not directionally biased.
The outputs from the simulation included the pelvic von Mises bone stress, bone remodelling stimulus pattern and the contribution of each muscle to the bone remodelling stress stimulus in the acetabulum by perturbing each muscle individually.