hic

A

pchwirweah©

K

1

Fhssttsoi

bhitaE

0d

Accident Analysis and Prevention 40 (2008) 1135–1148

Head injury prediction capability of the HIC, HIP, SIMon and ULP criteria

Daniel Marjoux ∗, Daniel Baumgartner, Caroline Deck, Remy WillingerUniversite Louis Pasteur, IMFS, 2 rue Boussingault, F-6700 Strasbourg, France

Received 5 October 2006; received in revised form 22 October 2007; accepted 8 December 2007

bstract

The objective of the present study is to synthesize and investigate using the same set of sixty-one real-world accidents the human head injuryrediction capability of the head injury criterion (HIC) and the head impact power (HIP) based criterion as well as the injury mechanisms relatedriteria provided by the simulated injury monitor (SIMon) and the Louis Pasteur University (ULP) finite element head models. Each accidentas been classified according to whether neurological injuries, subdural haematoma and skull fractures were reported. Furthermore, the accidentsere reconstructed experimentally or numerically in order to provide loading conditions such as acceleration fields of the head or initial head

mpact conditions. Finally, thanks to this large statistical population of head trauma cases, injury risk curves were computed and the correspondingegression quality estimators permitted to check the correlation of the injury criteria with the injury occurrences. As different kinds of accidentsere used, i.e. footballer, motorcyclist and pedestrian cases, the case-independency could also be checked. As a result, FE head modeling provides

ssential information on the intracranial mechanical behavior and, therefore, better injury criteria can be computed. It is clearly shown that moderatend severe neurological injuries can only be distinguished with a criterion that is computed using intracranial variables and not with the sole globalead acceleration.

2007 Elsevier Ltd. All rights reserved.

tdWtaoaNpcutoenp

eywords: Finite element head model; Injury criteria; Impact biomechanics

. Introduction

The brain is among the most vital organs of the human body.rom a mechanical point of view, the natural evolution of theead has lead to a number of integrated protection devices. Thecalp and the skull, but also to a certain extent the pressurizedubarachnoidal space and the dura matter, are natural protec-ions for the brain. However, these structures are not adaptedo the dynamic loading conditions involved in modern road andports accidents. The consequence of this extreme loading isften moderate-to-severe injuries. Thus, preventing these headnjuries is a high priority.

Over the past 40 years, an emphasis has been placed byiomechanical research on understanding the mechanism ofead injuries. One of the main difficulties in this research fields that functional deficiency is not necessarily directly linked
o a visibly damaged tissue. Nevertheless, an injury is alwaysconsequence of exceeded tissue tolerance to a specific load.ven if local tissue tolerance has been investigated for decades,
∗ Corresponding author.E-mail address: [email protected] (D. Marjoux).

(oal

it

001-4575/$ – see front matter © 2007 Elsevier Ltd. All rights reserved.oi:10.1016/j.aap.2007.12.006

he global acceleration of the impacted head as well as impacturation is usually utilized as impact severity descriptors. Theayne State University Tolerance Curve has been used since

he early 1960s; according to Lissner et al. (1960) and Gurdjiannd Webster (1958) the curve shows the link between the impactf the head, described by head acceleration, the impact durationnd risk of head injury. Based on the work of Gadd (1966), theational Highway Traffic Safety Administration (NHTSA) pro-osed the head injury criterion (HIC) in 1972. This is the toolurrently used in safety standards for head protection systemssing headforms. Since it is based solely on global linear resul-ant acceleration of a one-mass head model, some limitationsf this empiric criterion are well known (Newman, 1986). Forxample, the HIC is not specific to direction of impact and iteglects the angular accelerations. This is why Newman pro-osed the GAMBIT and more recently the head impact powerHIP) at the end of the 1990s (Newman et al., 2000). A method-logy was formulated to assess brain injuries based on multipleccident reconstructions of American football players’ head col-
isions during recorded games.
In the computation of the HIC and HIP criteria, the heads modeled as a rigid mass without any deformation. Usinghe finite element method and, as a consequence of comput-

mailto:[email protected]

dx.doi.org/10.1016/j.aap.2007.12.006

1136 D. Marjoux et al. / Accident Analysis and Prevention 40 (2008) 1135–1148

Table 1Correspondences between the conscience state evaluation and the neurological injuries occurrence (Levin et al., 1979; Teasdale and Jennet, 1974)

Anterograde amnesia Glasgow scale AIS

MM hS

itatnteaa(eicidastr

bTbc(FtprscacTItd3

daU

2

oHc

pr

-

-

-

miph

heoaAicouhaamessary for the pedestrian cases. Finally, all the 61 cases could be

ild neurological injuries Length < 20–30 minoderate neurological injuries 30 min < length < 24

evere neurological injuries Length > 24 h

ng capacitity improvement, the deformation of the skull andhe internal components can now be simulated. This methodllows for the addition of mechanical observations that begino match the description of known injury mechanisms. Hence,ew injury criteria are proposed. In the past few decades, morehan ten different three-dimensional finite element head mod-ls (FEHM) have been reported in the literature by Ward etl. (1980), Shugar (1977), Hosey and Liu (1980), Dimasi etl. (1991), Mendis (1992), Ruan et al. (1991), Bandak et al.1994), Zhou et al. (1995), Al-Bsharat et al. (1999), Willingert al. (1999) and Zhang et al. (2001). Fully documented headmpact cases can be simulated in order to compute mechani-al loading sustained by brain tissue as well as other tissuesn the surrounding areas and we compare it to the real injuriesescribed in the medical reports. Zhou et al. (1996), Kang etl. (1997), and more recently King et al. (2003) show that brainhear stress and strain rates predicted by their FEHM agree withhe location and severity of axonal injuries described in medicaleports.

Due to finite element head models, new injury prediction toolsased on computed intracranial loading are becoming available.he FEHM developed at Wayne State University has been usedy Zhou et al. (1995) to develop these tools. Thirteen motorcy-list accidents were reconstructed by Willinger and Baumgartner2001) at Strasbourg Louis Pasteur University (ULP) using theEHM presented in Willinger et al. (1999) and described in Sec-

ion 3 of this paper. This study establishes that computed brainressure is not correlated with the occurrence of brain hemor-hages, whereas brain Von Mises stress is. In order to undertake atatistical approach to injury mechanisms, more football, motor-ycle and pedestrian accident cases were introduced in Willingernd Baumgartner (2003). This was the first proposal of injuryriteria to specific mechanisms. Another FEHM, developed byakhounts et al. (2003) is the simulated injury monitor or SIMon.t is appropriate for this kind of study due to its short computingime. In that study, a number of scaled animal model loading con-itions lead the authors to provide criteria as reported in Section.

The objective of this study is to use a set of real-world acci-ents to assess the injury prediction capability of the HIC, HIPs well as the criteria provided by the SIMon FEHM and theLP FEHM.

. Methodology

A database of sixty-one real-world accident cases is used inrder to compare the injury prediction capability of the HIC, theIP, the SIMon and the ULP FEHM criteria. Football, motor-

ycle and pedestrian accidents are described in Section 3 of this

cwsS

13 ≤ GCS ≤ 15 1 ≤ AIS ≤ 29 ≤ GCS ≤ 12 3 ≤ AIS ≤ 43 ≤ GCS ≤ 8 AIS 5

aper. Each accident case is classified according to its medicaleport as follows:

Cases of neurological injuries are the cases in which a con-cussion, unconsciousness, a coma or diffuse axonal injuries(DAI) have been reported. These injuries are limited to thebrain region and stem from the neurological system rather thanthe vascular. MRI examination is the only way (with autop-sies) to examine directly such injuries but is rarely available.Therefore, in most cases, the only way to know the occur-rence of neurological injuries is to measure the consciencestate. Some tools exits such as the Glasgow scale (Teasdaleand Jennet, 1974) but a very confident way to evaluate theextent of DAI is to measure the duration of the anterogradeamnesia, as shown by Levin et al. (1979). Table 1 showsthe correspondences. Hence, we shall define moderate neu-rological injuries when unconsciousness lasts less than 24 h(at least 30 min). Severe neurological injuries will refer tounconsciousness lasting more than 24 h.Cases of subdural haematoma (SDH) when vascular injurieswith bleeding are observed between the brain and the skull.Cases of skull fracture which can be linear or depressive.Among these cases, there is not any case where the onlyreported skull fracture is a basilar fracture.

No special classification is used for subjects with injuries inore than one of the three categories: they are assigned multiple

njury classifications. Besides, deaths (seven cases, only amongedestrian cases) are treated as other cases. Twenty-four subjectsad no injury at all.

Moreover, each accident provides loading condition of theead. These loading conditions can be described in terms of lin-ar and angular acceleration curves of the head center of gravityr in terms of relative position and velocity between the headnd the impacted surface at the time just prior to the impact.lthough the ULP FEHM can be driven for both kinds of load-

ng conditions, the HIC, HIP and SIMon criteria can only beomputed using 3D acceleration fields. These accelerations arebtained from experimental or numerical accident replicationssing a Hybrid III dummy head. Since experimental replicationsad already been achieved for the footballer and motorcyclistccident cases, the 3D acceleration fields were already avail-ble. Thus, numerical accident replications using finite elementodels of the Hybrid III head and the windscreen were only nec-

onsidered for HIC and HIP computation and could be simulatedith both the SIMon and the ULP FEHM. This methodology

ynthesized in Fig. 1 allows the computation of the HIC, HIP,IMon and ULP injury criteria for the whole set of accident data.

D. Marjoux et al. / Accident Analysis and Prevention 40 (2008) 1135–1148 1137

SIMo

cabwcl

3

ihic

3

2st

-

-

-

Fig. 1. Methodology permitting the computation of HIC, HIP,

As a last step and in order to evaluate the injury predictionapability of the different criteria, an injury mechanism relatedpproach was adapted. For each kind of injury, the correlationetween the injury parameter values and the injury occurrencesas reported and illustrated through histograms. Injury risk

urves could then be computed for each injury mechanism fol-owing the method described in Nakahira et al. (2000).

. Data sources

This section describes the real-world accidents that were usedn the present study and details how the initial conditions areandled to drive the head models in order to compute the relatednjury criteria. Furthermore, head models and details on criteriaomputation are also synthesized.

.1. The real-world accidents used in the study

Twelve motorcyclist accidents, 22 footballer accidents and7 pedestrian accidents have been used in this study. The recon-truction of these accident cases was performed previously tohe present analysis in works that are listed below.

The motorcyclist accidents are those described in Chinn et al.(1999). They were experimentally reconstructed in collabo-ration between the ULP, the Transport Research Laboratory(TRL) and the Glasgow Southern General Hospital. Theaccident scenario was analyzed by accidentologists and thevictim’s helmet was collected on the scene. The accelerationfield sustained by the head during the impact was then inferredexperimentally by using an instrumented Hybrid III dummyhead, which was fitted inside four or five new helmets sim-ilar to the one worn by the victim. Dummy head and eachhelmet were dropped from heights corresponding to variousimpact velocities against different kinds of anvils. The goalwas to reproduce on the new helmet the same damages as thoseobserved on the victim’s one. Obviously, a strong assumptionwas made with this methodology since the influence of thebody inertia was not taken into account. Therefore, only the
very beginning of the head impact must be considered: thefirst acceleration peak, which is clearly larger than the rest ofthe curve amplitude. In this study, the motorcyclist accidentsare referenced with a letter “M”.
n and ULP criteria for all the 61 real-world head trauma cases.

The footballer accidents are those described in Newman et al.(2000) who used the following methodology: during Ameri-can football games, two dedicated cameras were used in orderto provide a 3D reconstruction of the scene. The relative posi-tion, orientation and velocities between the helmeted heads oftwo players colliding together could be extracted from thisreconstruction. Then, the scene was replicated experimen-tally using two helmeted Hybrid III dummy heads, whichwere instrumented with accelerometers. The validation of thismethod was based on the rebound of the full body dummies,which was compared to the rebound of the true football play-ers’ bodies shown on the movie. In this methodology, theinfluence of the body was taken into account. However, the firstpeak corresponding to the head impact is here again clearlylarger than the rest of the curve amplitude. In this study, thefootballer accidents are referenced with a letter “S”.Finally, the pedestrian accidents are those reconstructed fromthe database of the Accident Research Unit of the MedicalUniversity of Hanover. The traces collected on the implicatedvehicles and around the accident scene had been reported byaccidentologists. These data provided the exact location of thehead impact as well as an evaluation of the impact velocities.Only accidents in which the victim’s head hit the middle of awindscreen were extracted from this database. An analyticalstudy was then handled in order to infer the kinematics of thepedestrian body until the impact of the head. A pedestrian rigidbody model was used and parameterized with body memberlengths, mass and inertias. The impact velocity between thelegs and the car bumper that had previously been evaluatedby accidentologists was used as the input. Each reconstruc-tion was validated when the head of the pedestrian model wasfound to impact the middle of the windscreen. The results werealso confronted to the reported declarations of accident wit-ness. Finally, a numerical replication using a finite elementmodel of a windscreen and of a Hybrid III head was per-formed in order to obtain the acceleration curves undergoneby the head. The windscreen model was previously describedin Willinger and Baumgartner (2001). It consists on three lay-ers of composite shell elements with a mechanical behavior
based on the experimental data presented by Harward (1975).The finite element Hybrid III dummy head model was mod-eled with a viscoelastic skin and a rigid mass. The mechanicalparameters of the skin model were determined from the mate-


Table 2Injuries sustained by accident victims, ranges of accelerations and durations � of the accelerations

Accident origin Acc. (g) τ (ms) Skull fractures SDH Mod. neuro. inj. Sev. neuro. inj.

Motorcyclists (12 cases) 90–270 5.5–16 0 1 6 1Footballers (22 cases) 30–130 2–7.5 0 0 9 0P

T

p33ta

3

3

m

H

wt(

3

oad

H

-

-

edestrians (27 cases) 50–300 9.5–14.5 18

otal (61 cases) 30–300 2–16 18

rial used for the skin of the physical head dummy. Since thehead dummy is rigid, the mechanical behavior of this par-ticular skin is calibrated so that the energy, which should beabsorbed by a real deformable skull is compensated. Thus,the inertias of the head model are close to those measuredon a real dummy. In this methodology, the secondary impactis neglected. Hence, a strong assumption is made when con-sidering the head impact as the supposed injurious one. Inthis study, the pedestrian accidents are referenced with aletter “P”.

The age and gender of the victims were not available foredestrian cases. The footballers were all males between 20 and0 years old. Motorcyclists were all males but one woman, with1 average age (between 21 and 38). Besides, the injuries sus-ained by the victims, as well as the ranges of head accelerationsnd acceleration durations, are summarized in Table 2.

.2. Head injury criteria description

.2.1. HIC criterionProposed by the NHTSA in 1972, the head is seen as a one-

ass structure. It is computed using the following formula:

IC = max(t1,t2)

⎧⎪⎨⎪⎩(t2 − t1)

⎡⎣ 1

t2 − t1

t2∫t1

a(t) dt

⎤⎦

2.5⎫⎪⎬⎪⎭

-

Fig. 2. SIMon finite element head mod

5 8 8

6 23 9

here a (m s−2) is the resultant linear acceleration measured athe center of gravity of the Hybrid III dummy head. t1 and t2ms) are chosen in order to maximize the HIC value.

.2.2. HIP criterionProposed by Newman et al. (2000), the head is also seen as a

ne-mass structure. It is computed using both linear and angularccelerations measured at the center of gravity of a Hybrid IIIummy head as shown in the following formula:

IP = C1ax

∫axdt + C2ay

∫aydt + C3az

∫azdt

︸︷︷︸Linear contribution

+C4αx

∫αxdt + C5αy

∫αydt + C6αz

∫αzdt

︸︷︷︸Angular contribution

The Ci coefficients are set as the mass and appro-priate moments of inertia for the human head (50thpercentile): C1 = C2 = C3 = 4.5 kg, C4 = 0.016 N m s−2,C5 = 0.024 N m s−2, C6 = 0.022 N m s−2.ax, ay and az (m s−2) are the linear acceleration componentsalong the three axes of the inertial reference space attached to
the dummy head.αx, αy and αz (rad s−2) are the angular acceleration compo-nents around the three axes of the inertial reference spaceattached to the dummy head.
el (from Takhounts et al. (2003)).


elem

abitHfc

3

bi

Fig. 3. ULP finite

Since the HIP is a time-dependant function, the value takens an injury predictor candidate is the maximum value reachedy this function. The algorithm has been implemented and val-dated using the results provided by Newman et al. (2000) on
he same footballer cases as the ones used in the present study.IP was designed only for brain injury and not for SDH or skull
racture. It seemed, nevertheless, interesting to test its predictionapability for these injuries too.

isse

Fig. 4. Histograms of the four injury c

ent head model.

.2.3. SIMon criteriaThese criteria are computed using the intracranial mechanical

ehavior simulated by the finite element head model describedn Bandak et al. (1994), Takhounts et al. (2003) and illustrated
n Fig. 2. The advantage of its simple geometry is of course thehort computing duration, which makes the statistical approachimpler. A limitation of this model is the skull, which is consid-red as rigid, and the FEM can only be driven by acceleration
riteria for neurological injuries.


ry cr

fim

s

-

-

-

ars

(

3

odarsn

Fig. 5. Injury risk curves for the four inju

elds. Direct impacts can therefore not be simulated with thisodel as explained in the next section.Three injury criteria detailed in Takhounts et al. (2003) are

pecific to injury mechanisms as follows:

Cumulative strain damage measure (CSDM), which is sup-posed to be correlated with neurological injury occurrences.It measures the cumulative portion of the brain tissue experi-encing tensile strains over a predefined critical level. Severalsuch critical levels are proposed in the software and a level of15% is chosen as it seems to show the best correlation withinjuries after scaled animal test simulations.Dilatation damage measure (DDM), which is supposed to be acorrelate with contusions. Since there are very few cases withreported contusions among the 61 cases, the relevance of thiscriterion will unfortunately not be investigated in this study.
Relative motion damage measure (RMDM) is supposed to bea correlate with acute subdural haematoma. It is based on thebrain motion computation relative to the interior surface of thecranium.
e

ab

iteria for moderate neurological injuries.

Even if no skull fracture criterion can be calculated fromloading descriptor computed by the SIMon itself, a crite-

ion named the skull fracture criterion (SFC) is available in theoftware:

AHIC = �VHIC�THIC

with VHIC =∫

THIC

a dt where THIC is the time

t2 − t1) derived from the HIC calculation.

.2.4. ULP criteriaThe ULP three-dimensional FEHM used in this study is the

ne detailed in Willinger et al. (1999). This model, which isescribed more in details in the literature, includes the skin,deformable skull, the face, the dura matter (falx and tento-

ium), the subarachnoidal space, the brain and the cerebellum ashown on the Fig. 3. The model is constituted by about 12,000odes, 10,500 brick elements and 2800 shells. The characteristic
lement size is about 1 cm.
The ULP FEHM can be both driven by acceleration fieldspplied to a skull supposed to be rigid (motorcyclist and foot-aller cases) or through a direct impact with a deformable


jury c

sc

i

-

-

-

3e

t

Fig. 6. Injury risk curves for the four in

kull and using the windscreen finite element model (pedestrianases).

As described in Willinger and Baumgartner (2001), threenjury criteria are computed with this model:

The maximal Von Mises stress value reached by a signifi-cant volume of at least 10 contiguous elements (representingabout 3 cm3 of brain volume) from the brain is proposedas a correlate to neurological injury occurrences. Von Misesstress was preferred to strain for empirical reasons: it wasshown to be better correlated to neurological injuries inprevious studies. Other authors like Anderson (2000) madethe same conclusions. Another reason is mathematical: VonMises stress is a frame independent scalar (such as pres-sure) whereas strain depends on the orientation of the
frame.The maximum value reached by the global internal strainenergy of the elements modeling the space between the brainand the skull is proposed as a correlate to subdural haematoma
tpd

riteria for severe neurological injuries.

occurrences. This value represents the integral of the σ × ε

product among the whole space between the brain and theskull. It is a way to quantify the energy absorbed by this space.The maximum value reached by the global internal strainenergy of the deformable skull is proposed as a correlate toskull fracture occurrences. This criterion is only computedfor the pedestrian cases where the deformable skull FEHM isdriven with a direct impact. This value aims to quantify theenergy absorbed by the skull.

.3. Specificities in the selection of the accident samples forach category of injury

The SDH and neurological injuries prediction capability ofhese criteria is assessed using the whole set of accidents.

The skull fracture prediction capability is assessed using onlyhe pedestrian cases. In these cases, HIC, HIP and SFC are com-uted with 3D acceleration fields obtained from the previouslyescribed numerical reconstructions whereas the internal defor-


jury c

mtfr

4

ctotTfiatvhd

vttamWvi

ts

P

wxemt(

ott“pa

tTb

Fig. 7. Histograms of the four in

ation energy of the ULP FEHM deformable skull is computedhrough direct impact simulation. The consequences of this dif-erence in the inputs are discussed at the end of the followingesults section.

. Results

The determination of the head injury risk curves for spe-ific injury mechanisms is based on a correlation study betweenhe values of the proposed candidate criteria and the injuryccurrences. A histogram is built for each specific injury andhe value taken by a given criterion for each case is plotted.hese accident cases are sorted according to the injury classi-cation as explained in the methodology section, i.e. moderatend severe neurological injuries, SDH and skull fractures. Whenhe injury predictor candidate is adequate, a clear distinction isisible between the low values of the uninjured cases and theigh values of the injured cases and a threshold can thereby beetermined.

This threshold can accurately be calculated since it is thealue leading to a 50% risk of an injury risk curve. In this work,he modified maximum likelihood method is chosen. It is a logis-ic regression method developed and described by Nakahira etl. (2000), which shows better results than the classical maxi-
um likelihood method and the method described by Mertz andeber (1982). On the obtained curves, the circles represent the
ictims with mention to their injury statement (uninjured: 0 andnjured: 1) in y-coordinate and to their considered injury predic-

i9ae

riteria for subdural haematoma.

or candidate value in x-coordinate. The injury risk curve is aigmoid with the following formula:

(x) = 1

1 + e−(a+bx)

here P is the probability of injury for the given valueof the injury predictor candidate. The a and b param-

ters are determined using maximum likelihood method toaximize the function’s fit to the data. The estimator of

he goodness of fit has been called EB by Nakahira et al.2000) and is defined as equal to the log likelihood:EB =

1n

log

⎧⎨⎩

∏i

P(xi) ×∏j

(1 − P(xj))

⎫⎬⎭where n is the total number

f accident cases, xi are the predictors of the injured cases and xj

he predictors of the uninjured cases. In addition, another estima-or called EA by Nakahira et al. (2000) evaluates the assumptionthe injury probability approaches zero when the injury relatedarameter approaches zero”. An EA = 5% level has been useds proposed by Nakahira et al. (2000).

The quality of the regression is thereby given by the nega-ive estimator EB, which should be as close to zero as possible.hus, the 95% confidence limits of each injury risk curve haseen calculated and plotted. It notably gives the 95% confidence
ntervals of the deducted thresholds for risks of 5%, 50% and5%. These thresholds are indicated on the figures as well as theand b regression parameters and EA and EB corresponding
stimators.


the f

r

-

-

-

bcnbevpb

HaiU

cTFtrav−te

Fig. 8. Injury risk curves for

For the four injury mechanisms and the four injury criteriaesults are reported as follows:

Figs. 4–6 show the results for the prediction of neurologicalinjuries, both moderate and severe.Figs. 7 and 8 show the results for the prediction of subduralhaematoma.Figs. 9 and 10 show the results for the prediction of skullfractures.

In the four histograms in Fig. 4, cases have been separatedetween three categories: cases without neurological injuries,ases with moderate neurological injuries and cases with severeeurological injuries. Thus, within each category, the cases haveeen sorted according to the increasing values of the consid-
red criterion. For the HIC, the HIP and the ULP criteria, thealues for uninjured cases are globally lower than the ones com-uted for injured cases. Such a distinction can also be doneetween cases with moderate and severe neurological injuries.
Eo

b

our injury criteria for SDH.

owever, no correlation between neurological injury severitynd CSDM0.15 value can be observed. Therefore, tolerance lim-ts are only underlined by these histograms for HIC, HIP andLP criteria.Moreover, these graphical observations are quantitatively

onfirmed by the injury risk curves provided in Figs. 5 and 6.he moderate neurological injury tolerance limit is estimated inig. 5 and the severe neurological one in Fig. 6. The EB parame-

ers are very close to zero for the ULP criterion (−0.6 and −0.5,espectively for moderate and severe neurological injury toler-nce limit), indicating a good correlation between the Von Misesalue and the neurological injury gravity. For the HIC (−1.3 and0.8) and HIP (−1.2 and −0.8), the EB values are also close

o zero but further than the ULP ones, especially for the mod-rate neurological injury tolerance limit. Finally, the very low
B values computed for the CSDM0.15 confirm what has beenbserved on the corresponding histogram.
The histograms in Fig. 7 underline a poor statistical distri-ution between cases without subdural haematoma and cases


r inju

wsmvScwcbt

wltt

cm

ptbftm

e3D acceleration fields can be sufficient for the computation ofthe other ULP criteria (subdural haematoma and neurologicalinjuries). Hence, these specific ULP criteria were also computedusing 3D accelerations. The new results obtained for pedestrian

Table 3Summary of the main results of the injury risk curves

Injury type Proposed injury criterion EB 50% risk

Skull fracture HIC −1.0 667HIP (kW) −1.0 38 kWSFC (g) −0.7 73 gULP skull IE (mJ) −0.6 833 mJ

SDH HIC −0.9 1429HIP (kW) −1.4 55 kWSIMon RMDM −2.0 2.5ULP CSF IE (mJ) −0.9 4211 mJ

Moderate neurological injury HIC −1.3 533HIP (kW) −1.2 24 kWSIMon CSDM0.15 (%) −3.7 25%ULP VM (kPa) −0.6 27 kPa

Fig. 9. Histograms of the fou

here such injuries were revealed. Therefore, no conclusionhould be drawn from these histograms. Nevertheless, the largeajority of uninjured cases lead to HIC, HIP or ULP criterion

alues lower than the values reached by all but one injured cases.uch a tendency is less obvious for the RMDM criterion. Theorresponding injury risk curves (Fig. 8) lead to EB values,hich are farther to zero (especially due to the poor statisti-

al distribution) for all the criteria, even if the −1.0 value foroth HIC and ULP criterion confirms the previously observedendencies.

Histograms in Fig. 9 show that for all the criteria, the casesithout skull fractures globally reach values that are clearly

ower to the ones reached by cases with fracture. Here again,his is confirmed by the injury risk curves (Fig. 10), which leadso EB values close to zero, especially for SFC and ULP criteria.

Finally, a synthesis of the prediction capability of each injuryriterion in terms of EB value is reported for the different injuryechanisms in Fig. 11 and results are synthesized in Table 3.Pedestrian cases were evaluated by two ways as explained

reviously (direct impact for ULP criteria and using 3D accelera-ion fields for the other metrics). This difference in the inputs can
e arguable but was needed in order to compute the ULP skullracture criterion: simulating the direct impact is the only wayo evaluate the skull deformations. Thus, the skull deformation
ust have an influence on the intracranial mechanical behavior,

S

ry criteria for skull fractures.

specially on the zones closed to the deformed skull. However,

evere neurological injury HIC −0.8 1032HIP (kW) −0.9 48 kWSIMon CSDM0.15 (%) −4.4 44%ULP VM (kPa) −0.5 39 kPa


four

ccwph

Fio

i

Fig. 10. Injury risk curves for the

ases were added to the ones obtained with footballer and motor-
yclist cases. The resulting histograms and injury risk curvesere very similar to the ones presented here. Moreover, the EBarameter was even better: −0.8 instead of −0.9 for subduralaematoma; −0.5 instead of −0.6 for moderate neurological
ig. 11. EB regression quality estimator for each injury type with the associatednjury criteria. The closer to zero this negative parameter is, the best is the qualityf the regression.

ic

5

rswtamtm

tmtl

injury criteria for skull fractures.

njuries; and −0.43 instead of −0.5 for severe neurologicalnjuries. However, these new results do not change the followingonclusions of the present study.

. Discussion

The logistic regression analysis has been made on a ratherelevant statistical population of 61 accident cases when con-idering neurological injuries or SDH and of 27 accident caseshen considering skull fractures. The estimator EB of the logis-

ic regression takes the quality of the statistical populations intoccount as well as the correlation between the proposed injuryetric and the injury occurrences. It is also important to note

hat there are different kinds of accidents so that the injuryechanisms should not be case-dependent.In order to sharpen the injury thresholds inferred by the logis-

ic regression, an important alternative point was to select asuch non-extreme accidents as possible, i.e. neither too mild nor

oo violent. This selection should nevertheless explain the over-ap in the histograms between some non-injured cases whose

1 is and

cc

mnccatda

sceroisrimh

csbmrneticctbtpTi

dTorsarsrpHen

s

sa

bmhts

Etcnstifiat“cw

nidnFoeriaTfiaai

imcaeTuFvMcaan

146 D. Marjoux et al. / Accident Analys

onsidered injury mechanism value is high and some injured-ases for which this value is low though.

Concerning the quality of the statistical population, a com-ent should be made about the proportion between injured and

on-injured cases. This proportion is acceptable for neurologi-al injuries and for skull fractures since the number of injuredases is comparable to the number of uninjured cases and bothre statistically consistent. It may be more arguable concerninghe SDH since there are very few injured cases. However, thisisproportion should explain the low quality of the regressions indicated by the EB regression quality estimator.

Since the injury criteria have been computed on the sameet of accident cases, the comparison of their injury predictionapability is thereby possible. In terms of EB regression qualitystimator as reported in Table 3, the ULP FEHM based crite-ia seem to have the best prediction capability for each typef injury. This is particularly true concerning the neurologicalnjuries since the injury criterion based on the peaks of Von Misestress keeps its accuracy even when predicting the moderate neu-ological injuries. An injury mechanism based on the computedntracranial mechanical behavior of the brain was obviously the

ain motivation for building a finite element model of the humanead.

The rather bad results obtained with the SIMon based criteriaoncerning neurological injuries are thereby surprising. It is noture that the simplicity of this model is the only explication sinceoth SIMon and ULP FEHM have a comparable number of ele-ents. However, the geometry of ULP model seems closer to the

eal anatomy of the head and, therefore, the computed intracra-ial mechanical parameters could be more realistic. This mayxplain why the CSDM0.15 that seems to be an interesting wayo exploit the whole computed intracranial mechanical behav-or, does not lead to the expected results. This hypothesis isonfirmed in Franklyn et al. (2003) who showed that the CDSMomputed with the WSU model is much more accurate than withhe SIMon model. Furthermore, the 15% critical parameter maye too rigid. The CSDM0.15 should actually be calculated withhe strain fields computed by the ULP FEHM and Von Miseseaks could also be computed from the SIMon FEHM results.his could indicate whether the model or the way the CSDM0.15

s computed is responsible for these unexpected results.Although HIP was designed only for brain injuries, its pre-

iction capability was also tested for SDH and skull factures.he HIP results for skull fractures are moreover as good as thenes with HIC. When considering moderate brain injuries, theesults for HIP are slightly better than HIC. This was expectedince the HIP calculation takes rotational acceleration fields intoccount and neurological injuries are supposed to be more cor-elated with angular accelerations than linear accelerations asuggested in King et al. (2003). For more violent cases, theotational accelerations were found negligible in this study com-ared to the linear ones. This could explain why the results of theIC, which provides a more elaborate way to take linear accel-

rations into account, become better than the HIP for severeeurological injuries.

As explained previously, the evaluation of the prediction ofubdural haematoma is clearly less accurate and no conclusion

acot

Prevention 40 (2008) 1135–1148

hould be drawn at this stage as not enough injuries of this typere reported in the present study.

As regards to skull fracture, even if the results are slightlyetter with the ULP criterion, the use of a finite element modelay be less justified. The (SFC), which is based on a single-mass

ead model, leads indeed to comparable results at this stage ofhe study. Further investigations are needed concerning basilarkull fracture prediction.

While the injury prediction capability is assessed using theB estimator, the accuracy of the injury thresholds inferred by

he regression analysis can be evaluated with confidence limitsurves. In this study, like in most biomechanical studies, theumber of data is limited. Thus, the data are usually censoredince they are biased in one direction or another. The sign ofhe bias is known but not the magnitude. This explains the quitemportant width of the 95% confidence limits plotted on thegures. However, the slopes underlying risk function are steepnd according to Di Domenico and Nusholtz (2003), the steeperhese slopes, the smaller the sample size that is needed to obtaingood” risk estimation are. Given the censored nature of the data,onsistence threshold (CT) methods may be used in a futureork as presented in Kent and Funk (2004) for instance.Another limitation of this study is the hypothesis that there is

o correlation between the different categories of injuries. Fornstance, the energy absorbed by a skull fracture could allowecreasing the loading of the brain and therefore prevent fromeurological injuries. This is taken into account by the ULPEHM with a deformable skull (pedestrian cases) but not by thether criteria. The loading of the brain might thereby be over-valuated in cases with fractures and the resulting tolerance limitelative to brain injuries could be affected. Besides, skull fractures often accompanied by extradural haematoma, but there is notny case with this kind of injury in the database used in this study.olerance limits of a second impact might also be affected after arst impact. This is not taken into account by any injury criterionnd it is obviously a strong limitation. Furthermore, the victims’ge, gender and head morphology are not taken into account evenf tolerance limits might depend on such parameters as well.

The overall main limitation for such a study is the reliabil-ty of the replication of the accidents that are used. The authors

ust trust the reconstructions, which have been made by spe-ialists. The footballer cases are well known and have been usednd discussed in several studies such as the one by Newmant al. (2000). The motorcyclist cases have been made by theRL using reliable experimental techniques. The TRL eval-ates the uncertainty on the acceleration field to about 10%.inally, an uncertainty of about 20% on the resulting initialelocities is proposed by the Accident Research Unit of theedical University of Hanover for the pedestrian cases. More

omplex accident reconstructions are also described in the liter-ture such as the ones handled by Franklyn et al. (2005). Theseccidents involve car occupants, a category of victims which wasot investigated in the present study. Nevertheless, the important
mount of work and means needed by such reconstructions is notompatible with the statistical approach of the present method-logy. Besides, this kind of very complex accidents might leado uncertainties in the victim’s head kinematics too.

is and

tnvwvlhs1ifikc

6

iiUcmcaTcbgitum

toecuc

aFuw(iei

R

A

A

B

C

D

D

F

F

G

GH

H

K

K

K

L

L

M

M

N

N

N

R

S

D. Marjoux et al. / Accident Analys

A sensibility analysis was handled in order to assess howhe uncertainty in the inputs affects EB for each metric: theumerical Hybrid III model was dropped with various impactelocities and the resulting different criteria were analyzed. Itas shown that all criteria are not affected the same way. For lowelocities (around 5 m/s), a 10% variation in the impact velocityeads to a comparable variation for all the criteria. However, forigher impact speeds (10 or 20 m s−1), while the ULP Von Misestress criterion and the SFC remain stable (15% variation for a0% variation in the impact velocity), the other criteria are verynstable. Therefore, a range of EB should have been providedor each criterion in order to evaluate how the uncertainty in thenputs would affect the conclusions. To this point, one must justeep in mind that the uncertainty is lower for Von Mises stressriterion and SFC than for the other criteria.

. Conclusion and perspectives

Sixty-one real-world accident cases have been reconstructedn order to provide head acceleration fields and head initialmpact conditions so that the HIC, the HIP, the SIMon and theLP criteria could be computed. New tolerance limits to spe-

ific injury mechanisms were deduced for the ULP head FEodel and the relevance of their capability to predict injuries

ould therefore be investigated comparatively with HIC, HIPnd SIMON criteria, using histograms and injury risk curves.he advantage of this methodology is that this injury predictionapability is not deduced from ex vivo or animal experimentsut on real-world accidents. The main result of this study is theood capability in predicting moderate and severe neurologicalnjuries of criteria based on a finite element head model such ashe ULP model. This was expected since a single-mass modelsed by criteria such as the HIC or the HIP is not able to correctlyodel the intracranial mechanical behavior.Although the quality and the accuracy of the accident replica-

ions and reconstructions are obviously arguable, the relevancef this study should be found in the high number of consid-red accidents. This statistical approach should decrease theonsequences of possible errors. However the statistical pop-lation of cases with subdural haematoma must imperatively beonsolidated.

The justification of these neurological injury criteria, whichre inferred empirically, is obviously still opened to discussion.or this matter, pilot studies with more detailed head models arender progress at the Wayne State University foe example, asell as in vivo studies such as the ones handled by Anderson

2000) with sheep. However, the purpose of this study was tonvestigate the injury prediction capability of existing criteriaven if further understanding of the living tissue thresholds withn vivo experiments is needed.

eferences

l-Bsharat, A., Hardy, W., Yang, K., Khalil, T., Tashman, S., King, A., 1999.Brain/skull relative displacement magnitude due to blunt head impact: newexperimental data and model, Proc. of the 43rd STAPP Car Crash Conf., pp.321–332.

T

T

Prevention 40 (2008) 1135–1148 1147

nderson, R., 2000. A study of the biomechanics of axonal injury. PhD thesis,University of Adelaide.

andak, F.A., Van Der Vorst, M.J., Stuhmiller, L.M., Mlakar, P.F., Chilton, W.E.,Stuhmiller, J.H, 1994. An imaging based computational and experimentalstudy of skull fracture: finite element model development. In: Proceedingsof the Head Injury Symposium, Washington DC.

hinn, B.P., Doyle, D., Otte, D., Schuller, E., 1999. Motorcyclists head injuries:mechanisms identified from accident reconstruction and helmet damagereplication, Proc. IRCOBI Conf., pp. 53–72.

i Domenico, L., Nusholtz, G., 2003. Comparison of parametric and non-parametric methods for determining injury risk. Paper 2003-01-1362,Society of Automotive Engineers.

imasi, F., Marcus, J., Eppinger, R., 1991. 3D anatomic brain model for relatingcortical strains to automobile crash loading. In: Proceedings of the Interna-tional Technical Conference on Experimental Safety Vehicles, NHTSA, vol.2, pp. 916–923.

ranklyn, M., Fildes, B., Dwarampudi, R., Zhang, L., Yang, K., Sparke, L.,Eppinger, R., 2003. Analysis of computer models for head injury investi-gation. In: Proceedings of the 18th International Technical Conference onEnhanced Safety Vehicles.

ranklyn, M., Fildes, B., Zhang, L., Yang, K., 2005. Analysis of finiteelements models for head injury investigation: reconstruction of four real-world impacts. In: Proc. STAPP Car Crash Conf., Paper 2005-22-0001,pp. 1–32.

add, C.W., 1966. Use of a weighted – impulse criterion for estimating injuryhazard, Proc. of the 10th STAPP Car Crash Conf., pp. 164–174.

urdjian, E.S., Webster, A., 1958. Head Injury. Little Brown Company, Boston.arward, R.N., 1975. Strength of Plastics and Glass. Cleaver Hume Press Ltd.,

New York.osey, R.R., Liu, Y.K., 1980. A homeomorphic finite element model of impact

head and neck injury. I.C.P. Finite Elements Biomech. 2, 379–401.ang, H.S., Willinger, R., Diaw, B.M., Chinn, B., 1997. Validation of a 3D

human head model and replication of head impact in motorcycle accidentby finite element modelling. In: Proceedings of the 41th Stapp Car CrashConference Lake Buena Vista USA, pp. 329–338.

ent, R.W., Funk, J.R., 2004. data censoring and parametric distribution assign-ment in the development of injury risk functions from biomechanical data,SAE 2004. World Congress.

ing, A., Yang, K., Zhang, L., Hardy, W., 2003. Is head injury caused by linearor angular acceleration? IRCOBI Conference, pp. 1–12.

evin, H.S., O’Donnel, V.M., Grossman, R.G., 1979. The Galveston Orientationand Amnesia Test: a practical scale to assess cognition after severe injury. J.Nerv. Ment. Dis. 167, 674–684.

issner, H.R., Lebow, M., Evans, F.G., 1960. Experimental studies on the rela-tion between acceleration and intracranial pressure changes in man. Surg.Gynecol. Obstet. 111.

endis, K., Finite element modelling of the brain to establish diffuse axonalinjury criteria, PhD dissertation, Ohio State University, 1992.

ertz, H.J., Weber, D.A., 1982. Interpretations of the impact responses of a3-year-old child dummy relative to child injury potential. In: Proceedings ofthe 9th International Technical Conference on Experimental Safety Vehicles.

akahira, Y., Furukawa, K., Niimi, H., Ishihara, T., Miki, K., Matsuoka, F., 2000.A combined evaluation method and modified maximum likelihood methodfor injury risk curves. Proc. IRCOBI Conf., pp. 147–156.

ewman, J.A., 1986. Generalized acceleration model for brain injury threshold(GAMBIT). Proc. IRCOBI Conf., pp. 121–131.

ewman, J.A., Shewchenko, N., Welbourne, E., 2000. A new biomechanicalhead injury assessment function: the maximum power index. In: Proc. 44thSTAPP Car Crash Conf.

uan, J.S., Kahlil, T., King, A.I., 1991. Human head dynamic response to sideimpact by finite element modelling. J. Biomech. Eng. 113, 276–283.

hugar, T.A., 1977. A finite element head injury model, vol. 1. Report no. DOTHS 289-3-550-TA.

akhounts, E., Eppinger, R., Campbell, J.Q., Tannouns, R.E., Power, E.D.,Shook, L.S., 2003. On the development of the simon finite element headmodel. Stapp Car Crash J. 47, 107–133.

easdale, G., Jennet, B., 1974. Assessment of coma and impaired consciousness– a practical scale. Lancet 2, 81–84.

1 is and

W

W

W

W

Z

Zhou, C., Khalil, T.B., King, A.I., 1995. A 3D human finite element head

148 D. Marjoux et al. / Accident Analys

ard, C.C., Chan, M., Nahum, A.M., 1980. Intracranial pressure: a brain injurycriterion. SAE.

illinger, R., Baumgartner, D., 2001. Numerical and physical modelling of thehuman head under impact – toward new injury criterion. Int. J. Vehicle Des.32 (1–2), 94–115.

illinger, R., Baumgartner, D., 2003. Human head tolerance limits to specificinjury mechanisms. Int. J. Crashworthiness 6–8, 605–617.

illinger, R., Kang, H.S., Diaw, B.M., 1999. 3D human head finite elementmodel validation against two experimental impacts. Ann. Biomed. Eng. 27(3), 403–410.

Z

Prevention 40 (2008) 1135–1148

hang, L., Yang, K., Dwarampudi, R., Omori, K., Li, T., Chang, K., Hardy,W., Kalil, T., King, A., 2001. Recent advances in brain injury research: anew human head model development and validation. Stapp Car Crash J.45.

for impact injury analyses. In: Symposium Proceedings Prevention throughBiomechanics, pp. 137–148.

hou, C., Kahlil, T.B., Dragovic, L.J., 1996. Head injury assessment of a realworld crash by finite element modelling, Proc. AGARD Conf.

Documents

hic