Su

Pa

b

c

a

ARRA

KARTPP

1

rc51swupcp

0d

Artificial Intelligence in Medicine 54 (2012) 29– 41

Contents lists available at SciVerse ScienceDirect

Artificial Intelligence in Medicine

jou rn al h om epage: www.elsev ier .com/ locate /a i im

tatic and dynamic pressure prediction for prosthetic socket fitting assessmenttilising an inverse problem approach

hilip Sewell a,∗, Siamak Noroozia, John Vinneya, Ramin Amalib, Stephen Andrewsc

School of Design, Engineering & Computing, Bournemouth University, Poole, Dorset, BH12 5BB, UKFaculty of Environment and Technology, University of the West of England, Bristol, Bristol, BS16 1QY, UKDisablement Services Centre, Southmead Hospital, Bristol, BS10 5NB, UK

r t i c l e i n f o

rticle history:eceived 21 June 2010eceived in revised form 5 September 2011ccepted 8 September 2011

eywords:rtificial neural networksehabilitationrans-tibialrostheticressure measurement

a b s t r a c t

Objective: It has been recognised in a review of the developments of lower-limb prosthetic socket fittingprocesses that the future demands new tools to aid in socket fitting. This paper presents the resultsof research to design and clinically test an artificial intelligence approach, specifically inverse problemanalysis, for the determination of the pressures at the limb/prosthetic socket interface during stance andambulation.Methods: Inverse problem analysis is based on accurately calculating the external loads or boundaryconditions that can generate a known amount of strain, stresses or displacements at pre-determinedlocations on a structure. In this study a backpropagation artificial neural network (ANN) is designedand validated to predict the interfacial pressures at the residual limb/socket interface from strain datacollected from the socket surface. The subject of this investigation was a 45-year-old male unilateraltrans-tibial (below-knee) traumatic amputee who had been using a prosthesis for 22 years.Results: When comparing the ANN predicted interfacial pressure on 16 patches within the socket withactual pressures applied to the socket there is shown to be 8.7% difference, validating the methodology.Investigation of varying axial load through the subject’s prosthesis, alignment of the subject’s prosthesis,and pressure at the limb/socket interface during walking demonstrates that the validated ANN is able to

give an accurate full-field study of the static and dynamic interfacial pressure distribution.Conclusions: To conclude, a methodology has been developed that enables a prosthetist to quantitativelyanalyse the distribution of pressures within the prosthetic socket in a clinical environment. This will aidin facilitating the “right first time” approach to socket fitting which will benefit both the patient in termsof comfort and the prosthetist, by reducing the time and associated costs of providing a high level of socket fit.

. Introduction

Every year the NHS receives approximately 5000 new patienteferrals to their prosthetic limb service following amputation orongenital deficiency. There are currently estimated to be over5,000 amputees in the UK, which equates to approximately 1 in000 of the UK population [1]. Accuracy of the socket fit remains theingle most important parameter affecting amputee satisfactionith their rehabilitation, as this is the interface between the resid-al limb and prosthesis. Fergason and Smith [2] state the socketortion of the trans-tibial (below-knee) prosthesis is of signifi-

ant importance in the overall outcome in the rehabilitation of theatient.

∗ Corresponding author. Tel.: +44 01202 961294; fax: +44 01202 965314.E-mail address: psewell@bournemouth.ac.uk (P. Sewell).

933-3657/$ – see front matter © 2011 Elsevier B.V. All rights reserved.oi:10.1016/j.artmed.2011.09.005

© 2011 Elsevier B.V. All rights reserved.

It is recognised that the quality of socket fit is directly relatedto the pressure distribution generated at the residual limb/socketinterface. The socket fit is considered to be of a high quality if theentire load is supported by the pressure tolerant regions (e.g. areasof thick tissue) of the limb and the pressure sensitive regions (e.g.areas where the bone is close to the surface) are non-load bearing.Mak et al. [3] state that achieving a comfortable interfacial pressuredistribution between a residual limb and prosthetic socket is veryimportant in socket design and fit. Although evidence-based prac-tice is paramount to the provision of optimum care and enhancingquality of life for the amputee, a limited amount of evidence isavailable to prescribing clinicians which would provide a clearunderstanding of what constitutes a “good socket fit” [4]. There-fore, the fitting process is “artisan” in nature and dependent on the

level of experience and the degree of skill of the fitting prosthetist.It comes as no surprise that the latest Audit Commission report [5]has revealed that a quarter of prosthesis users do not regularly usetheir artificial limbs. The main reasons, specifically detailed in the

3 igence in Medicine 54 (2012) 29– 41

ra(

ssattdavodaip

but[tTgarmQt

uova

tpcpttoa

(sleatatata

sitpcpwea

0 P. Sewell et al. / Artificial Intell

eport, for adult amputees not using their artificial limbs as muchs they would like are: too uncomfortable (29%); too much pain25%); does not fit (12%).

Current research is tending towards the development of newocket materials to enhance suspension and redistribute the pres-ures within the socket, developing new prosthetic componentrynd integrating computer aided design/computer aided manufac-ure (CAD/CAM) into the socket manufacturing process. However,he new materials and componentry that are being developedo not solve all issues with the prosthesis and have some dis-dvantages. They are unable to resolve issues related to limbolume fluctuation due to stump oedema and/or muscle atrophy,r eliminate all other pressure related problems (ulcers, sores, skinisorders, for example). They are also not suitable for all amputeesnd increase the unit cost of a prosthesis. CAD/CAM facilities, whilemproving the production rate, do not improve the quality of therosthesis fit.

Many approaches have been investigated to evaluate socket fitoth in stance and walking, such as interface stress investigationstilising transducers [6,7] and finite element analysis (FEA) [8–10];o evaluate the effect of limb volumetric fluctuations on socket fit11–13]; and to evaluate the effect on gait of prosthetic componen-ry such as the foot, pylon, knee joint, torque absorbers, etc. [14–16].he use of transducers for direct pressure measurement limits dataathering to a small number of sites within the prosthetic socketnd generally requires modification of the socket geometry. FEAequires detailed information of the socket and limb geometry andaterial properties, which cannot be gathered for every patient.uantitative gait analysis requires access to a gait analysis labora-

ory, a facility available to very few prosthetic clinics.These investigations have produced results that aid in the

nderstanding of the complex issues faced when fitting a socketr designing new socket technology. However, they have not pro-ided the prosthetist with any useful clinical tools to aid in thessessment of prosthetic socket fit.

A published review by the authors of “developments in therans-tibial prosthetic socket fitting process: a review of past andresent research” [17] concluded that new tools are needed to over-ome limitations of current socket fitting processes that will beractical for use in a clinical environment and that will fit into theime and cost constraints of the NHS. Eleven years later it is clearhat these conclusions are still valid as there has been little impactn clinical practice or socket technology witnessed in the UK orbroad.

There is currently only one commercial system on the marketTekscan®) which allows the pressure distribution in the prosthe-is to be measured through the insertion of a sensor mat at theimb/socket interface. Issues found with Tekscan® are that driftrrors are significantly affected by curvature; that cyclic drift isffected by the presence of a gel liner [18]; and that the calibra-ion method can have an effect on the accuracy [19]. Repeatabilitynd accuracy are therefore key issues. This research has concludedhat further studies should be performed with other curvaturesnd liners. It has also been recommended that investigators designheir own calibration curves for the Tekscan® system to improveccuracy.

This review of currently available methods of measuring pres-ure has shown that pressure data can only be collected by eithernserting a sensor into the socket/limb interface, which affectshe results collected, or by modifying the socket to insert theressure transducer, making it unsuitable for everyday use. Over-oming these limitations will enable a clinical tool, instead of a

urely research tool, to be developed. Therefore, the ideal approachould be to measure the pressure directly from strain transduc-

rs attached to the socket surface. However, this is not possibles the internal pressure cannot be related to the surface strain

Fig. 1. ANN training loop to minimise error.

through a closed form formula. Therefore, it is proposed by theauthors to utilise an artificial intelligence approach, specificallyinverse problem analysis, for the determination of the pressuresat the limb/prosthetic socket interface. This approach can find acomplex non-linear transfer function that relates the surface strainto the internal pressure.

Artificial neural networks (ANNs) have been utilised in otherareas of prosthetic design such as prosthesis alignment [20], assess-ment of gait patterns [21] and control of above-knee poweredprosthesis [22]. A study by Chahande and Billakanti [23] usedan ANN to provide the underlying bone data around the kneeto enhance surface skin topological images that can currently beacquired by devices such as a video laser scanner. The reviewof literature in this field has revealed that an ANN approach forthe prediction of interfacial pressures for trans-tibial amputees isnovel.

This paper presents the results of research to design, validateand clinically test the artificial intelligence approach for the deter-mination of the pressures at the limb/prosthetic socket interfaceunder different static loading conditions and during walking. Theaim of this multidisciplinary research is to provide prosthetists witha means of making an informed judgement of trans-tibial socket fitthrough the process of collecting and analysing quantitative pros-thesis socket pressure distribution data during all phases of the gaitcycle.

2. Inverse problem methodology design

Inverse problem analysis is based on accurately calculating theexternal loads or boundary conditions that can generate a knownamount of strain, stress or displacement at pre-determined loca-tions on a structure. An ANN, as an inverse problem solver, can beutilised to solve structural problems where the structure’s responseto the load is known but the load which causes this response is not.

To predict the interfacial pressures at the residual limb/socketinterface (i.e. ANN output data) from strain data collected from thesocket surface (i.e. ANN input data) the ANN requires a number ofexample input and output data for training (i.e. relating the ANNinputs to outputs using a transfer function and series of weight-ing values). New surface strain data can then be introduced to thetrained ANN (problem data) to give the predicted pressures (Fig. 1).

The data is fed through the ANN and it is adjusted using an acti-vation function (F) which is typically a sigmoid function, given bythe following:

F(u) = 11 + exp(−s ∗ u)

(1)

where s is a constant that affects the slope of the sigmoid function.In this study a modified sigmoid function, which has been shownelsewhere by the authors [24] to improve training speed for this

application, is applied as shown below:

MF(u) = 0.5a × (1 − f (u)) + 2F(u) × b (2)

igence in Medicine 54 (2012) 29– 41 31

wl

∀wi

cotiiyksatmancf

tida

2

papafba

gtwuaattceucgsawt

g9bggvap

Table 1Strain gauge specifications.

Type WFRA-6-11-3LT (Techni Measure, Studley, UK)Resistance 120 �Gauge factor 2.1

lected. Fig. 3 shows the typical change in strain over time froma gauge attached to the socket worn by the subject while theywere sitting. The results indicate that after 700 s (approximately

Fig. 2. Strain gauge based transducer.

P. Sewell et al. / Artificial Intell

here MF is the modified sigmoid function and, a and b are theimitation numbers that satisfy the follow condition:

i; 1 ≤ i ≤ m ⇒ 0.5a ≤ Yi ≤ 2b (3)

here m is the number of neurons in the ANN output layer and Ys the output of every neuron in the ANN output layer.

Once the relationship between the loading and response of theomponent has been established the inverse problem approachffers a high-speed solution for monitoring in-service loads onhe structure (e.g. prosthetic socket). New response data can bentroduced to the solver and the load on the component predictednstantaneously. The main advantages of an inverse problem anal-sis approach is that it can be utilised without the need for anynowledge of the constitutive laws or geometry of the limb skin, tis-ue, bone or socket material, which would be required for valid andccurate numerical analysis. These parameters are accounted for inhe function found by the ANN that relates the inputs to outputs. It

ust be noted that this methodology provides a unique solution forn individual prosthetic socket and a new transfer function wouldeed to be found for each new socket to be studied and/or whenhanges are made to the geometry of a socket for which a transferunction has previously be found.

The design of the inverse problem methodology for determininghe dynamic pressures at the limb/socket interface during walk-ng is detailed in the following sections. The key components to beesigned are the methods of acquiring training and problem datand the ANN architecture.

.1. Acquisition of training data

To find the relationship between the surface strains and internalressures the ANN requires a transducer that can reliably measure

strain response on the socket due to the internal pressure. Theressure and related strain data must then be stored as ANN inputsnd outputs pairs. A large number of these pairs were requiredor the ANN to accurately train (i.e. find an accurate relationshipetween the inputs and outputs). This requires a loading device forpplying known pressures inside the socket.

The transducer must produce a repeatable response (i.e. for aiven pressure magnitude and position the output should be iden-ical for every test). The most suitable technique for this purposeas found to be strain measurement. This method is regularlysed to monitor structures and would produce reliable and repeat-ble responses to loading the component. Strain measurement ischieved by attaching strain gauges to the socket surface. This hashe advantages of (i) producing a large response for a small load, (ii)he gauges are bonded to the component and therefore the gaugeannot slip, and (iii) the gauge response can be easily captured in anlectronic format. However, there are also several disadvantages ofsing strain gauges as part of a clinical tool which are (i) the gaugesannot be removed without being damaged, (ii) attaching strainauges is time consuming, (iii) strain gauges are relatively expen-ive, and (iv) the gauge requires an excitation voltage, amplificationnd signal conditioning to produce meaningful data. These issuesere overcome to make the use of strain gauges a viable option for

he collection of training data as part of the clinical tool.The final transducer utilised a resin encased stacked strain

auge rosette (giving three separate strain readings at 0◦, 45◦ and0◦) usually used for measuring strain in underwater applicationsonded to a plastic plate (Fig. 2). The specification of the strainauge utilised in the transducer can be found in Table 1. Investi-

ations found that the transducer could be bonded to the socketia the plastic plate using a standard hot glue gun. This provided

semi-permanent bond. The encased gauge was found to be wellrotected against damage and could be easily removed from the

Gauge length 6 mmGauge width 2.3 mm

socket. The remaining glue could be peeled from the socket surfacecausing no damage to the socket itself.

Once the socket is donned by the subject it will heat up due tothe increased temperature from the residual limb. Therefore, it wasimportant to investigate how long it would take the strain readingsfrom the transducers attached to the socket to reach a steady-state.This would determine the length of time required for the socket tobe worn by the patient before any meaningful data could be col-

Fig. 3. Change in strain over time before steady-state reached.

3 igence in Medicine 54 (2012) 29– 41

1tpsd

ssalrnffid(wor

st7psug

dsaaarcpie

ptlwhwfa

mtawc1

2

etAtn

tp

2 P. Sewell et al. / Artificial Intell

2 min) the strains had drifted considerably due to the heating ofhe socket. The strains had begun to reach a steady-state after thiseriod of time. It was concluded from this investigation that theubject would need to don the socket for at least 12 min beforeata could be collected.

The positions of the strain readings collected form the socketurface are also important. To produce efficient training data thetrain data should be captured at the sensitive regions (i.e. the straint those locations must vary significantly due to changes in pressureevel). In addition, the strain data collected must provide a uniqueesponse for each pressure distribution. If strain is collected fromon-sensitive regions and/or the strain data collected is not unique

or each pressure distribution the ANN is less likely to be able tond a function relating the input and output. In this study 10 trans-ucers giving 30 strain readings were attached to the socket surfaceFig. 4(a)). The location of these transducers on the socket surfaceas determined by using the results of a sensitivity study carried

ut by the authors [25] which gave an insight into the sensitiveegions to pressure within the socket.

A Windmill 751-SG strain monitoring and control data acqui-ition system (Windmill Software Ltd., Manchester, UK) was usedo capture the strain data with a resolution of ±1 microstrain. The51-SG package comprises Windmill 6 software, a USB unit whichrovides differential inputs to monitor 16 strain gauges at up to 80amples per second, and a strain gauge connection card. Eight USBnits can be connected to one laptop to monitor up to 128 strainauges.

The socket loading device (Fig. 4(b)) was designed for the stan-ardisation of the collection of training data in any type or size ofocket. The device has a loading arm that is fixed to a sliding mech-nism. Two springs are connected to the top of the arm. When therm is pushed the springs extend causing a reaction force in therm. The prosthesis can be attached to the device via its pylon to theotating platform. This allows the socket to turn through 360◦. Theombination of rotating platform and sliding mechanism allows aressure to be applied at any position within the socket. The load-

ng arm has been designed so that it will fit into the geometricalxtremes of any socket.

Strain gauges are attached to the loading arm near the pivotoint which has been calibrated against a known pressure so thathe software can directly read pressure values. A locking pin is alsoocated close to the pivot to hold the arm in the vertical position

hen it is being raised or lowered. Quick release locking clampsave been provided to stop any movement in the rotating platformhile data is being collected. Using markings on the rotating plat-

orm it is possible to quickly locate the centre of each patch andpply a pressure to it.

The pressure is applied to the internal surface of the socket byeans of a circular pad (Fig. 4(a)). To ensure the loading is normal to

he surface of the socket the pad is connected to a universal joint and rubber surface is attached to the face of pad to ensure complianceith the socket curvature. To calculate the pressure applied to the

entre of each patch the force is divided by the surface area of the2.5 mm diameter pad.

.2. ANN architecture

The architecture defines the number of neurons or processinglements (PEs) and the way they interconnect in the ANN’s struc-ure. Key architectural issues are (i) the number of layers in theNN, (ii) the number of PEs per layer, (iii) the type and parame-

ers of the PE which are usually the same throughout, and (iv) the

umber of calculations per iteration during learning and recall.

The optimum number of nodes in the hidden layer of the ANN,he learning rate and momentum constant were determined in arevious study by the authors [26] for a socket divided into 16

Fig. 4. (a) Pressure being applied to a socket patch and (b) the training device.

patches. Ten removable, reusable strain gauge based transducersgiving 30 strain readings were attached to the socket in areas sen-sitive to pressure change as discussed in Section 2.1. Therefore, theANN studied had 30 nodes in the input layer and 16 nodes in the

output layer to match the number of strains to be captured andpressures to be predicted from the prosthetic socket in a clinicalenvironment. The final specification of the ANN is shown in Table 2.

P. Sewell et al. / Artificial Intelligence

Table 2Architecture of the artificial neural network.

Architecture Feed forward backpropagationData process NormalisationNoise generator ±1 microstrain on 10% of training

patternsRange of pressures 0 to user-defined maximumNumber of nodes in input layer

(surface strains)30

Number nodes in output layer(internal pressures)

16

Number of nodes in hidden layer 16Number of training patterns 1420Number of testing patterns 50Number of problem patterns Depends on the number of patterns

collectedNumber of loops 50,000

warswtcpptweaoiotct

hhtfwmi

2

wtwiicdtvlwtt

cv

Learning rate 0.0005Momentum constant 0.0003

The number of required training patterns was found to be 1420hich meant that 1420 sets of 16 random pressures on each patch

nd the resultant strains caused by these random pressures wereequired to provide enough input/output data to find the relation-hip between them. 50 testing patterns were used. This numberas randomly selected and the patterns were generated in addi-

ion to the training patterns. It was found that the amount of dataollected could be dramatically reduced using the superpositionrinciple to generate training and testing patterns from the inde-endent parent patterns. The application of this theory was showno be valid through studying the linearity of the socket materialithin the expected pressure range, as discussed by the authors

lsewhere [26]. The theory of superposition states that the straint a point on a structure due to a series of loads is equal to the sumf the strains from each individual load case. Using this theory annfinite number of training patterns can be generated by applyingne known load to each location on the structure individually. Onceraining and testing data had been generated it was normalised byonverting all data to a range of between −0.5 and 0.5 to improvehe accuracy of the solution.

It is possible to over-train the ANN, which means that the ANNas been trained to respond to only one type of input. If this shouldappen then learning can no longer occur. In real-world applica-ions this situation is not very useful as a separate over-trained ANNor each new kind of input would be required. To ensure the ANNas not over-trained the training was supervised to ensure that theean square error of the testing data did not increase, which is an

ndication over over-training.

.3. Acquisition of problem data

Problem data is the captured strain data on the socket surfacehile the subject is wearing the socket. Problem data can be cap-

ured under different loading conditions (i.e. the prosthesis is setupith different alignments, the subject is sitting, standing or walk-

ng). Problem data can also be collected from the subject at severalntervals so that a comparison between the pressure distributionsan be made to aid in the assessment of socket fit over time and/orue to limb geometry changes. It is introduced to the trained ANN sohat it can predict the pressure levels within the socket due to theseariables. It is essential that the strain data is captured at identicalocations for both the training and problem data: if not the results

ill be invalid. Therefore this data should also be collected usinghe same transducers in the same position as for the collection of

raining data.

A portable system is required as the strain data was to beaptured in a clinical environment. The strain data was collectedia the Windmill USB unit which is attached to a laptop running

in Medicine 54 (2012) 29– 41 33

the bespoke data acquisition/ANN software. This software wasdeveloped by the authors in Visual C++ utilising Windmill dataacquisition Active X controllers and ANN theory to generate thecode.

3. Inverse problem methodology assessment results

Three case studies were performed to assess the developed sys-tem which are detailed in the following sections. The clinical trialswere carried out at the Disablement Services Centre, SouthmeadHospital, Bristol with the full consent of the hospital’s ethical com-mittee.

3.1. Subject profile and prosthesis setup

The subject of this investigation was a 45-year-old male uni-lateral trans-tibial traumatic amputee who had been using aprosthesis for 22 years. He had a cylindrical, fleshy residual limb,140 mm in length, and was fitted with a patellar tendon-bearingsocket with a supracondylar cuff suspension. The subject’s socketused in the studies was vacuum formed using Northplex (an opti-cally clear copolyester manufactured by North Sea Plastics Ltd.,Glasgow, UK) and a standard pylon and foot were attached (Fig. 5(a)and (b)).

3.2. Training data acquisition, ANN training and validation

The socket divided into 16 patches can be seen in Fig. 6. Train-ing data was collected using the loading device and the ANN trainedbefore the clinical studies began. Therefore, the analysis of the prob-lem data collected could be performed instantly in the clinic.

Lastly the ANN was trained to minimise the errors (Fig. 7). Thefinal testing error was approximately 12% after 30,000 loops of theANN, which took 30 min to reach. This meant that the ANN shouldpredict accurately to within 88% of the actual pressure value. Theaccuracy of the ANN in this study was found to be less than in aprevious study by the authors [26]. In the previous study it wasfound that the ANN was able to predict accurately to within 91.2%of the actual pressure value. The 3% difference in the reported accu-racies is due to the reduction in the number of inputs to the ANN inthis study. In the previous study 45 input strains were used to pre-dict 16 pressures. In this study 30 input strains (from the 10 strainrosettes) are used to predict the 16 pressures. While this reduc-tion in the number of inputs has had an impact of the accuracy ofthe ANN the “ANN Difference Method” developed and reported bythe authors [24] has been applied to the ANN output to improveits accuracy further. It has been shown that application of the ANNDifference Method to the output of a backpropagation ANN canreduce inherent errors that exist at the low and high ends of theANN solution envelope.

To investigate and validate the performance of the trained ANN16 known experimental pressures were applied to the socket usingthe loading device (Fig. 4) and the generated strains were acquired.The strains were presented to the ANN and the magnitude of thepressure on each of the 16 patches was predicted by the ANN. Thegenerated data was new and had not been seen by the ANN pre-viously. Fifty strain patterns were produced and given to the ANNas problem data. The output of the ANN was then compared withthe original applied pressures that generated those strain patterns.Fig. 8 shows a comparison of the actual applied pressure and the

predicted ANN pressure from two of the 50 patterns used to validatethe ANN.

The results in Fig. 8 show the ANN was adequately trained andcould predict the pressures on each of the 16 patches to a high

34 P. Sewell et al. / Artificial Intelligence in Medicine 54 (2012) 29– 41

F

dt

p

pefp

Fig. 6. (a) Anterior (front) and (b) posterior (rear) views of the prosthesis dividedinto 16 patches.

three quarter weight or full weight (Fig. 9). The amount of weight

ig. 5. (a) Subject’s prosthesis setup and (b) subject standing in the prosthesis.

egree of accuracy. The percentage difference was calculated usinghe following formula:

ercentage error = actual result − ANN resultANN result

(4)

The average percentage error difference between the actual andredicted ANN pressures were −10.7% and −9.2% (the negative

rror indicating that the ANN result is larger than the actual result)or Tests 1 and 2 respectively (Table 3) and overall −8.7% for the 50atterns.

Fig. 7. Training and testing errors graph.

3.3. Case Study 1 – interfacial pressure changes due to variedweight applied to the prosthesis

The first investigation performed was to study the strains cap-tured and pressures predicted by the ANN when the body weightapplied through the prosthesis was varied. The subject was asked tolift his prosthesis from the ground so that the strain reading couldbe zeroed. The subject was then asked to apply one of five bodyweights to the prosthesis, no weight, quarter weight, half weight,

applied through the limb was judged by the subject.The strains were collected and presented to the validated

trained ANN to determine the predicted pressures at the

P. Sewell et al. / Artificial Intelligence in Medicine 54 (2012) 29– 41 35

Fig. 8. Comparison of pressures applied to the sock

Table 3Percentage difference between the actual and predicted ANN pressures for Tests 1and 2.

Patch number Test 1 (%) Test 2 (%)

1 4.1 6.72 −29.6 8.43 2.9 −5.34 −7.1 −7.35 5.3 0.46 −25.2 −48.37 7.7 −28.18 −31.0 5.29 0.1 −2.910 −1.8 5.411 6.2 8.212 8.0 −4.513 −18.6 −64.414 −51.1 −2.615 −49.7 −1.516 8.8 −17.0

Average −10.7 −9.2

Table 4Pressure data for each patch due to variation in weight applied to the prosthesis.

Patch number No weight(kPa)

Quarter weight(kPa)

Half weigh(kPa)

1 0.0 0.0 0.0

2 7.1 171.5 401.1

3 9.8 0.0 0.0

4 17.9 0.0 0.0

5 11.0 37.4 65.7

6 13.0 39.9 46.0

7 7.3 0.0 17.4

8 0.0 0.0 0.0

9 40.9 386.9 642.1

10 12.9 83.1 212.5

11 0.0 223.7 517.3

12 25.4 0.0 0.0

13 0.0 0.0 0.0

14 12.0 73.2 124.8

15 44.9 16.3 0.0

16 4.2 0.0 0.0

et with those predicted by the trained ANN.

limb/socket interface. The comparison of pressure results on eachof the 16 patches for the different applied weights can be seen inTable 4.

3.4. Case Study 2 – interfacial pressure changes due to variedalignment of the prosthesis

An investigation was performed to study the effects on thestrains captured and pressures predicted when the left to rightalignment of the prosthesis was modified by the prosthetist. Thesubject was again asked to lift his prosthesis from the ground sothat the strain reading could be zeroed. The subject was then askedto apply half body weight to the prosthesis for each test. The amountof load applied through the limb was judged by the subject. Inthe first study the alignment was set to extreme socket adduction(Fig. 10(a)). However, extreme socket adduction, in this particular

subject’s case, did not produce the same high level of alignmentchange as would be expected in a standard socket set-up. Thiswas because the subject’s prosthesis ideal alignment was close toextreme socket adduction already due to his particular physiology.

t Three-quarterweight (kPa)

Full weight(kPa)

Average pressure(kPa)

0.0 0.0 0.0514.4 511.2 321.1

0.0 0.0 2.0122.4 167.0 61.5117.7 107.0 67.8154.5 181.5 87.0

82.2 94.4 40.30.0 0.0 0.0

1003.0 1112.0 637.0414.1 461.6 236.8815.8 873.9 486.1

0.0 0.0 5.10.0 16.5 3.3

269.0 293.7 154.50.0 0.0 12.20.0 0.0 0.8

36 P. Sewell et al. / Artificial Intelligence in Medicine 54 (2012) 29– 41

Fig. 9. Subject applying (a) no weight, (b) quarter weight, (c) half weight, (d) three quarter weight and (e) full weight to the prosthesis.

P. Sewell et al. / Artificial Intelligence

Table 5Pressure data for each patch due to variation in alignment.

Patch number Extremeadduction(kPa)

Ideal(kPa)

Extremeabduction (kPa)

1 23.4 0.0 0.02 478.0 574.2 675.73 0.0 0.0 0.04 0.0 14.7 34.65 67.1 49.0 37.86 74.8 68.7 6.47 38.8 38.5 11.88 0.0 0.0 0.09 553.0 605.6 741.3

10 246.3 249.7 259.511 512.0 415.6 325.512 0.0 0.0 0.013 0.0 0.0 0.0

m(o

14 147.4 118.4 90.115 0.0 144.2 465.816 0.0 0.0 0.0

In the second study the alignment was set to the ideal align-ent (Fig. 10(b)) and finally it was set to extreme socket abduction

Fig. 10(c)). The predicted pressure results from the validated ANNn each of the 16 patches can be seen in Table 5.

Fig. 10. Changes to left to right socket alignment for (a) extreme s

in Medicine 54 (2012) 29– 41 37

3.5. Case Study 3 – determination of interfacial pressure duringwalking

To assess the inverse problem methodology for determiningthe pressure at the limb/socket interface during walking, the sub-ject was asked to walk up and down in a straight line and atan even speed while wearing the prosthesis. Ideally, a treadmillwould have been used to ensure that a constant walking speed wasachieved. However, a treadmill was not available during this study.The strains on the socket surface were continuously captured dur-ing this time. Before collection the strains were zeroed while thesubject elevated his prosthesis. Fig. 11 shows example strain datacollected while the subject was walking for patches 1 and 9. Resultsin Fig. 11 show that consistent strain data can be collected whilethe subject is walking.

Negative strains can be seen between the strain peaks, whichindicate the swing phase of the gait. Negative strains are seen inthis phase due to the momentum of the prosthesis as it swings,generating a reaction between the limb and prosthesis. The reactionis in the opposite direction to that generated in the stance phase of

gait and is also significantly smaller as there is no axial load throughthe prosthesis thus generating a small negative strain.

The strains were presented to the ANN to predict the pressure oneach patch on the prosthesis. Fig. 12 shows the predicted pressures

ocket adduction, (b) ideal and (c) extreme socket abduction.

38 P. Sewell et al. / Artificial Intelligence in Medicine 54 (2012) 29– 41

Fig. 11. Strain data collected from the socket surface during walking.

fpsloitwmto

4

ws

4w

s

Table 6Comparison of ANN and reported pressure (kPa) distribution results.

Study Patch number

5 6 7 8

ANN results (half weight) 65.7 46.0 17.4 0ANN results (full weight) 107.0 181.5 94.4 0Sanders et al. [27] 42.8–86.1 41.3–91.3 3.2–289.8 3.2–289.8Sanders et al. [28] 2.5–13.9 0.3–38 2.2–18.2 2.2–18.2Sanders and Daly [29] 48.9–98.2 23.7–117.9 78.3–120 78.3–120Dumbleton et al. [4] 59–70 41–62 52–59 88–100

Fig. 12. Predicted pressures during walking.

or patches 2 and 5. The results for patch 5 clearly show increasedressure peaks after every third step (circled on Fig. 12) where theubject has turned around to walk in the opposite direction. This isess evident for patch 2 as the predicted pressures on this patch aref a much lower magnitude. The consistent time between the peaksndicate that the subject walked at a fairly constant pace despitehe unavailability of a treadmill. Studying the predicted pressurealking peaks, between the turning peaks, for patch 5 indicates aaximum variance in the data of less than 50 kPa. Some variance in

he predicted pressure for each step is expected as the repeatabilityf gait of a lower-limb amputee is affected by their disability.

. Analysis of results

Numerical analysis of the interfacial pressure data collectedhile assessing the developed system is detailed in the following

ections.

.1. Case Study 1 – interfacial pressure changes due to varied

eight applied to the prosthesis

Table 4 presents the predicted pressure data and average pres-ure data for each patch and for each weight variation. These results

Fig. 13. Pressure intensity mapped to the socket (a) anterior and (b) posterior.

show that the predicted pressures increase as the weight appliedto the prosthesis increases, as was expected. Patches 1, 3, 8, 12, 13,15 and 16 do not seem to be subjected to significant pressure whilepatches 2, 9, 10, 11 and 14 are subjected to high pressures. Thepercentage increase from a quarter weight to full weight throughthe prosthesis for these high pressure patches are 198.1%, 187.4%,455.2%, 290.7% and 301.3% respectively. Pressure data was still pre-dicted under the ‘no weight’ condition as this data was collectedwith the subject’s foot raised; therefore, the weight of the prosthe-sis would still generate pressure. These results show the system wasable to predict the change in loading distribution due to changes inweight applied to the prosthesis.

Mapping the pressure data (darker shading represents higherpressure) predicted at the centre of each patch to the patchlocations (Fig. 13) indicates that the majority of the pressure isdistributed around the anterior-lateral and posterior-lateral por-tions of the socket while the minimum pressure is seen on theposterior-medial portion.

To further investigate the accuracy of the pressure results,a comparison between the ANN pressures at half weight (nor-mal stance) and full weight with those reported by the otherresearchers, under the same loading conditions, in the approxi-mate areas represented by the specified patches, was performed(Table 6). The pressure results found in this study were within thesame range as those from other reported results.

4.2. Case Study 2 – interfacial pressure changes due to variedalignment of the prosthesis

Table 5 presents the predicted pressure data for each patch andfor each alignment. The results in Table 5 show that there is anoticeable difference in pressure predicted on each patch due to

the different alignments. It can be seen that the pressure increaseson patches 2, 4, 9 and 15. The pressure decreases on patches 5, 6,7, 11 and 14. There is little or no pressure on patches 1, 3, 8, 12, 13and 16 as was found in case study one.

P. Sewell et al. / Artificial Intelligence

Fig. 14. Expected change in pressure distribution from extreme adduction toextreme abduction.

Fg

eodps

4w

rptgfrt

5

dtem

ig. 15. (a) Pressure variation on patch five during one step and (b) normal verticalround reaction force expected during gait.

These results are as expected; changing the alignment fromxtreme adduction to abduction would cause a shift in the balancef pressure. It would be expected that the pressure on the medial-istal and lateral-proximal areas of the socket would increase. Theressure on the medial-proximal and lateral-distal areas of theocket would decrease (Fig. 14).

.3. Case Study 3 – determination of interfacial pressure duringalking

The walking pressure data presented in Fig. 12 was further inter-ogated by analysing the variation in pressure on a representativeatch (patch 5) during one step (Fig. 15(a)). The result shows thathe pattern of each step follows the typical variation in verticalround force expected between the heel-strike and toe-off phaseor normal gait (Fig. 15(b)). The force is very small at heel-strike (a),ises to a maximum soon afterwards (b), then falls a little duringhe middle of stance (c), rises again (d), then falls before toe-off (e).

. Discussion

The use of high-tech analysis methods (e.g. mechanical trans-

ucers in the socket wall and FEA) have their place in increasinghe understanding of the limb/socket biomechanics in a laboratorynvironment. However, they are impractical for everyday assess-ent in a clinical environment, which would aid the prosthetist

in Medicine 54 (2012) 29– 41 39

in assessing the fit of a prosthesis. Qualitative photoelastic analy-sis has been shown to be able to quickly provide an understandingof the pressure distribution between the limb and the socket [30].For a more detailed analysis of the pressure distribution the devel-oped quantitative inverse problem methodology, presented in thispaper, has been shown to be able to accurately predict the levels ofpressure within the socket.

The methodology has been shown to remove the need for dataon tissue properties of the limb, which is of significant importance.It also minimises the interference of transducers with both thesocket and the tissue as these are low weight and are placed onthe socket surface. The interfacial pressures can also be predictedon large surfaces, depending on how the socket is divided. It also hasthe ability to give an accurate full-field study of interfacial pressuredistribution.

The case studies have shown the success of the artificial intelli-gence technique for static and dynamic testing. This methodologyallows the prosthetist to study the distribution of the pressures atevery phase of the gait cycle which is currently difficult to achieve ina clinical environment using traditional data acquisition methods.

The methodology is based on the utilisation of an ANN approachwhich has its own limitations. These limitations can be summarisedas follows:

• There is no standard method for selecting the most efficientANN parameters such as learning weight, momentum parame-ter, transfer function or even the number of training and testingpatterns.

• Training an ANN requires producing sufficient amounts of train-ing and testing patterns, which can be a time consuming andexpensive procedure.

• Training an ANN requires consideration of the local and globalminima as well as over-training. Improper consideration of thenumber of training or testing patterns may result in the ANNbecoming stuck in a local minima and/or over-training.

• Any modification to the socket being analysed requires retrainingof the ANN before a solution can be found.

In understanding the limitations steps have been taken in themethodology to overcome them. Investigation has been conductedto find the optimum ANN parameters, number of training and test-ing patterns, and training approach. The design of the removableand reusable transducers, the loading device to speed up the collec-tion of training data, and utilisation of the superposition techniquealso ensures the socket can be quickly retrained if the geometry ismodified.

The methodology provides the detailed levels of pressure withinthe socket. However, there currently is no data specifying the typ-ical pressures for a well fitted or poorly fitted socket. Therefore,the prosthetist would currently have to use their own judgementof how these levels relate to the fit of the prosthesis. It is envis-aged that the methodology could transform every prosthetic socketinto a full-field pressure transducer, or “Intelligent Socket”, at min-imal cost providing a large amount of information regarding therelationship between pressure and socket fit. This will allow theprosthetist to quantitatively analyse the distribution of pressureswithin the prosthetic socket in a clinical setting. It will allow thepressure profile to be captured and stored for comparison duringdifferent fitting visits to aid in determining where issues with thesocket fit could have arisen. This methodology could also have animpact in the way prosthetists are educated. Lastly, the method-ology has the potential to be adapted for remote monitoring of

the pressures. This could enable prosthetists to remotely diagnoseissues with socket fit through an Internet application, for example.

The overall aim of this research is to enable the real-time assess-ment and analysis of socket fit for a trans-tibial amputee during

40 P. Sewell et al. / Artificial Intelligence in Medicine 54 (2012) 29– 41

) prop

sont“ts

atad

6

tprtia

Fig. 16. Comparison of (a) current and (b

tance and walking. The tool is expected to accelerate the post-perative and rehabilitation phase for amputees, by reducing theumber of design fitting iterations required to achieve and main-ain optimal fit (Fig. 16). This will facilitate the move towards aright first time” approach to socket fitting by providing the pros-hetist with a clinical tool for assessment of socket fit at the fittingtage.

This methodology also provides a means of gathering largemounts of pressure data to provide further insights into the rela-ionship between pressure and socket fit. Future research by theuthors will investigate this relationship utilising the methodologyetailed in this article.

. Conclusion

A methodology has been developed that enables a prosthetisto quantitatively analyse the distribution of pressures within therosthetic socket during stance and walking in the clinical envi-

onment. This has the potential to aid in facilitating the “right firstime” approach to socket fitting which will benefit both the patientn terms of comfort, and the prosthetist by reducing the time andssociated costs of providing a high level of socket fit.

osed socket fitting assessment strategy.

Acknowledgements

The authors would like to thank the charity REMEDI for provid-ing funding for this research. REMEDI have not had any involvementin the study design, in the collection, analysis and interpretation ofdata, in writing the manuscript, or in the decision to submit themanuscript for publication.

References

[1] Torres MM. Incidence and causes of limb amputations. Archives of PhysicalMedicine and Rehabilitation: State of the Art Review 1994;8(1):1–8.

[2] Fergason J, Smith DG. Socket considerations for the patient with a transtibialamputation. Clinical Orthopedics and Related Research 1999;361:76–84.

[3] Mak FT, Zhang M, Boone DA. State-of-the-art research in lower-limb pros-thetic biomechanics–socket interface. Journal of Rehabilitation Research andDevelopment 2001;38(2):161–74.

[4] Dumbleton T, Buis AWP, McFadyen A, McHugh BF, McKay G, Murray KD,et al. Dynamic interface pressure distributions of two transtibial pros-

thetic socket concepts. Journal of Rehabilitation Research & Development2009;46(3):405–16.

[5] Audit Commission. Fully equipped – the provision of equipment to older ordisabled people by the NHS and Social Services in England and Wales. AuditCommission; 2000. ISBN: 1 86240 213 2.

igence

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

[

P. Sewell et al. / Artificial Intell

[6] Zachariah SG, Sanders JE. Standing interface stresses as a predictor of walk-ing interface stresses in the trans-tibial prosthesis. Prosthetics and OrthoticsInternational 2001;25(1):34–40.

[7] Sanders JE, Zachariah SG, Baker AB, Grieve JM, Clinton C. Effects of changesin cadence prosthetic componentry, and time on interface pressures and shearstresses of three trans-tibial amputees. Clinical Biomechanics 2000;15:684–94.

[8] Zhang M, Mak AF. A finite element analysis of the load transfer between anabove-knee residual limb and its prosthetic socket-roles of interface frictionand distal-end boundary conditions. IEEE Transactions on Rehabilitation Engi-neering 1996;4(4):337–46.

[9] Portnoy S, Yarnitzky G, Yizhar Z, Kristal A, Oppenheim U, Siev-Ner I, et al.Real-time patient-specific finite element analysis of internal stresses in the softtissues of a residual limb: a new tool for prosthetic fitting. Annals of BiomedicalEngineering 2007;35(1):120–35.

10] Lee WCC, Zhang M. Using computational simulation to aid in the prediction ofsocket fit: a preliminary study. Medical Engineering & Physics 2007;29:923–9.

11] Commean PK, Smith KE, Vannier MW, Szabo BA, Actis RL. Finite element mod-elling and experimental verification of lower extremity shape change underload. Journal of Biomechanics 1997;30(5):531–6.

12] Torres-Moreno R, Jones D, Solomonidis SE, Mackie H. Magnetic resonanceimaging of residual soft tissues for computer-aided technology applica-tions in prosthetics—a case study. Journal of Prosthetics and Orthotics1999;11(1):6–11.

13] Sanders JE, Fergason JR, Zachariah SG, Jacobsen AK. Interface pressure and shearstress changes with amputee weight loss: case studies from two trans-tibialamputee subjects. Prosthetics and Orthotics International 2002;26(3):243–50.

14] Schmalz T, Blumentritt S, Jarasch R. Energy expenditure and biome-chanical characteristics of lower limb amputee gait: the influence ofprosthetic alignment and different prosthetic components. Gait and Posture2002;16(3):255–63.

15] Twiste M, Rithalia S. Transverse rotation and longitudinal translation duringprosthetic gait—a literature review. Journal of Rehabilitation Research andDevelopment 2003;40(1):9–18.

16] Selles R, Bussmann J, Van Soest AJ, Stam H. The effect of prosthetic mass prop-erties on the gait of transtibial amputees—a mathematical model. Disability &Rehabilitation 2004;26(12):694–704.

17] Sewell P, Noroozi S, Vinney J, Andrews S. Developments in the trans-tibial pros-

thetic socket fitting process: a review of past and present research. Prostheticsand Orthotics International 2000;24:97–107.

18] Schrock L. The effect of surface curvature and a gel liner interface on perfor-mance properties of the Tekscan socket system. In: Special seminar: Master ofScience in prosthetics and orthotics student presentations. 2007.

[

in Medicine 54 (2012) 29– 41 41

19] Brimacombe JM, Wilson DR, Hodgson AJ, Ho KC, Anglin C. Effect of calibra-tion method on Tekscan sensor accuracy. Journal of Biomechanical Engineering2009;131(3):034503.

20] Sanders JE, Reed RD, Marks RJ. Computer-aided prosthetic alignment forlower-limb amputees. In: Rangayyan RM, editor. Proceedings of the annualinternational conference of the IEEE engineering in medicine and biology soci-ety: San Diego, California, USA, October 28–31. 1993. p. 1282–3.

21] Holzreiter SH, Kohle ME. Assessment of gait patterns using neural networks.Journal of Biomechanics 1993;26(6):645–51.

22] Kalanovic VD, Popovic D, Skaug NT. Feedback error learning neural networkfor trans-femoral prosthesis. IEEE Transactions on Rehabilitation Engineering2000;8(1):71–80.

23] Chahande AI, Billakanti SR. Identification of load bearing areas for prostheticlimbs in a below-the-knee amputee using neural networks. In: Rangayyan RM,editor. Proceedings of the annual international conference of the IEEE engi-neering in medicine and biology society: San Diego, California, USA, October28–31. 1993. p. 1284–5.

24] Sewell P, Noroozi S, Vinney J, Amali R, Andrews S. Improvements in the accuracyof an inverse problem engine’s output for the prediction of below-knee pros-thetic socket interfacial loads. Engineering Applications of Artificial Intelligence2010;23(6):1000–11.

25] Amali R, Noroozi S, Vinney J, Sewell P, Andrews S. An artificial intel-ligence approach for measurement and monitoring of pressure at theresidual limb/socket interface—a clinical study. INSIGHT Journal 2008;50(7):373–83.

26] Amali R, Noroozi S, Vinney J, Sewell P, Andrews S. Predicting interfacial loadsbetween the prosthetic socket and the residual limb for below-knee amputees– a case study. Strain 2006;42(1):3–10.

27] Sanders JE, Lam D, Dralle AJ, Okumura RM. Interface pressures and shearstresses at thirteen socket sites on two persons with transtibial ampu-tation. Journal of Rehabilitation Research & Development 1997;34(1):19–43.

28] Sanders JE, Bell DM, Okumura RM, Dralle AJ. Effects of alignment changeson stance phase pressures and shear stresses on transtibial amputees: mea-surements from 13 transducer sites. IEEE Transactions on RehabilitationEngineering 1998;6(1):21–31.

29] Sanders JE, Daly CH. Interface pressure and shear stresses: sagittal plane angular

alignment effects in three trans-tibial amputee case studies. Prosthetics andOrthotics International 1999;23:21–9.

30] Sewell P, Vinney J, Noroozi S, Amali R, Andrews S. A photoelastic clinical study ofthe static load distribution at the residual limb/socket interface of PTB sockets.Prosthetics and Orthotics International 2005;29(3):291–301.