1

Discriminating Progressive Supranuclear Palsy from Parkinson's Disease usingwearabletechnologyandmachinelearning. Authornamesandaffiliations

Maarten De Vos1, John Prince1, Tim Buchanan PhD2, James J. FitzGerald PhD3,4, and

ChrystalinaA.AntoniadesPhD4*

1 Department of Engineering Science, Institute of Biomedical Engineering, University ofOxford,OldRoadCampusResearchBuilding,OX37DQ,Oxford,UK.2UCBBiopharmaSPRL,Brussels,Belgium

3NuffieldDepartmentofSurgicalSciences,UniversityofOxford,Oxford,OX39DU,UK.

4NuffieldDepartmentofClinicalNeurosciences,NeuroMetrologyLab,UniversityofOxford,

Oxford,OX39DU,UK.

Correspondingauthor

ProfessorChrystalinaA.Antoniades,NuffieldDepartmentofClinicalNeurosciences,Level6,

WestWing,JohnRadcliffeHospital,UniversityofOxford,UnitedKingdom

E-mail:chrystalina.antoniades@ndcn.ox.ac.uk

Wordcount:2849(excludingreferences)

References:21

RUNNINGTITLE:WearingtechnologyforPSP

KEYWORDS:ProgressiveSupranuclearPasly,Parkinson’sdisease,gait,wearables,inertial

sensorarray,machinelearning.

FINANCIALDISCLOSURE/CONFLICTOFINTEREST:

MDV,JP,TB,andCAAhavenoconflictsofinterest.

JJFhasperformedconsultancyunrelatedtothesubjectofthisstudyforcompaniesincludingAbbott,Renishaw,andHerantis,andalsohasaresearchgrantfromAbbott,foraprojectunrelatedtothisstudy.

FUNDINGSOURCESFORTHESTUDY:

ThiswasstudywasfundedbyagrantfromtheUCB-OxfordCollaboration.JJFandMDVweresupportedbytheNationalInstituteforHealthResearch(NIHR)OxfordBiomedicalResearchCentre(BRC).

OxQuip is registered on the clinical trials website and this is the CT Identifier:NCT04139551

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

2

AbstractBackground:Progressivesupranuclearpalsy(PSP),aneurodegenerativeconditionsmaybedifficulttodiscriminateclinicallyfromidiopathicParkinson’sdisease(PD).Itiscriticalthatweareabletodothisaccuratelyandasearlyaspossibleinorderthatfuturediseasemodifying therapies forPSPmaybedeployedata stagewhen theyarelikelytohavemaximalbenefit.Analysisofgaitandrelatedtasksisonepossiblemeansofdiscrimination.Research Question: Here we investigate a wearable sensor array coupled withmachinelearningapproachesasameansofdiseaseclassification.Methods:21participantswithPSP,20withPD,and39healthycontrol(HC)subjectsperformed a two minute walk, static sway test, and timed up-and-go task, whilewearing an array of six inertial measurement units. The data were analysed todeterminewhatfeaturesdiscriminatedPSPfromPDandPSPfromHC.Twomachinelearningalgorithmswereapplied,LogisticRegression(LR)andRandomForest(RF).Results:17 features were identified in the combined dataset that containedindependent information. The RF classifier outperformed the LR classifier, andalloweddiscriminationofPSPfromPDwith86%sensitivityand90%specificity,andPSP fromHCwith 90% sensitivity and 97% specificity. Using data from the singlelumbar sensor only resulted in only amodest reduction in classification accuracy,whichcouldberestoredusing3sensors(lumbar,rightarmandfoot).Howeverformaximumspecificitythefullsixsensorarraywasneeded.Significance:Awearable sensor array coupledwithmachine learningmethods canaccurately discriminate PSP from PD. Choice of array complexity depends oncontext; for diagnostic purposes a high specificity is needed suggesting the morecomplete array is advantageous, while for subsequent disease tracking a simplersystemmaysuffice.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

3

Introduction

Progressivesupranuclearpalsy(PSP)(1,2) isanatypicalparkinsoniandisorderthat

can sometimes be difficult to discriminate clinically from the much commoner

Parkinson’s disease (3). It is characterized by vertical supranuclear gaze palsy,

postural instability andaxial rigidityaswell asmild cognitive impairment (1) (4-6).

Recently,theMovementDisordersSociety(MDS)studygrouphasproposedasetof

diagnosticcriteriaforPSP,withtheaimofimprovingthedetectionofthediseasein

clinicalpracticeandresearch(7).

Comparative studies using accelerometers (8) have shown many shared gait

abnormalities(9)(10-12)includingdecreasesinvelocity,steplength,cadence,and

meanacceleration.Somedifferenceshavealsobeenfound,includinglowervertical

displacementandhigheraccelerationinPSP.Inthispaperweexaminetheabilityof

a wearable array of inertial measurement units (IMUs), coupled with machine

learningalgorithms, todistinguishPSPpatients fromPDpatientsand fromhealthy

controls.

Therearea largeandgrowingnumberofwearable technologies (13,14)available

for characterizing and measuring the motor features of neurodegenerative

diseases(15-17).

When deciding whatmeasurement technology to use there are twomajor issues

thatmust be considered. One is the setting for themeasurement, which can be

broadlydividedintolaboratorymeasurementsorambulatorymeasurementsoutside

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

4

the laboratory (e.g. at home). The former will permit closely controlled and

supervisedconditionsgivingthebestchanceofobtainingahighqualitystandardized

dataset, and is well suited to tasks like detailed measurement of individual

components of the gait cycle. However there is growing interest in longer term

ambulatorymeasurementbecausesnapshot laboratoryrecordingsmaynotgivean

accurateoverallpicture,duetomanyfactorssuchasvariabilityofsymptomsbytime

ofday,inconsistenttimingofmedication,orthestressofbeingtestedinahospital

environment.Laboratorymeasurementsarealsounlikelytocaptureinfrequentbut

important features such as falls or near-falls, and a ‘realworld’ environmentmay

possiblybetterelicitfeaturesthatareofdirectclinicalrelevancetothepatient.

Thesecond issue is thechoiceofdevice. Inparticular, there isaconflictbetween

maximisingdataandminimisingtheeffortinvolvedinobtainingit.Atoneendofthe

spectrum, it is possible to have body worn arrays comprising numerous inertial

measurement units attached to all four limbs and the trunk, streaming tens of

megabytesperminute. At theoppositeextreme, some investigators are trying to

develop systems based on single sensors, even those built into consumer

smartphones,sothatallthatisrequiredisanapprunninginthebackground.There

is no doubt that the complex systems will generate much more complete

information;thequestioniscanthesimplersystemsanswerthesamediagnosticand

measurement questions acceptably well, despite having a comparatively meagre

dataset?

Thetwoissuesareclearlylinked.Applyingacomplexsysteminalaboratorysetting

is straightforward and is logical given the intent of obtaining detailed data in a

concentratedperiodof time. Usinga complex systemathome isanothermatter:

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

5

donninganddoffingmaynotbeeasy,andthemultiplesensorsmaygetinthewayof

everydayactivities.Thereisapremiumonusingassimpleasystemasiscompatible

withobtainingsufficientdatatobeuseful.

Our aim in this study is to answer two questions. Firstly, we will use a complex

laboratory systemwith sensorsoneach limband the trunk, to recorddataduring

commonly applied gait-related tasks. We will explore how well the data can

distinguishPSPpatientsfromPDpatientsandfromhealthycontrol(HC)participants,

and what the most essential distinguishing features within the dataset are.

Secondly,wewillanalysewhichsensors/bodylocationsarenecessarytoacquirethis

data.Theaimistodeterminehowfarthesensorarraycanbesimplifiedwhilestill

yieldingsatisfactoryresults.Itishopedthatthiswillhelpmaximisecompliancewith

future study protocols while avoiding collecting an inadequate dataset due to

oversimplification.

Materials&Methods

Participants

The participants in this study were recruited as part of the Oxford study of

Quantification in Parkinsonism (OxQUIP); a large clinical observational study being

conductedattheJohnRadcliffeHospital,Oxford.Werecruited20participantswith

PD,21participantswithPSP,and39healthycontrolparticipants.Allhealthycontrol

participantswere spousesof thePDor PSPparticipants. Thedemographics of the

participantsaresummarisedinTable1.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

6

Sensorarrayandsoftware

Participantsworeasystemconsistingofsixsynchronizedinertialmeasurementunits

(IMUs) (OpalTM,APDM,Portland,USA)thatwirelessly transmitteddatatosoftware

(MobilityLabTM,APDM)runningonanearbylaptop.TheIMUswerepositionedover

the lumbarspine,sternum, leftandrightwrists,and leftandrightfeet.Allsensors

provided tri-axial accelerometer, tri-axial gyroscope, and tri-axial magnetometer

signals at a frequency of 100Hz. Participants performed three tasks: two minute

walk,swaytest,andtimedup-and-go(TUG).Usingthewaveformsfromallsensors,

theMobilityLabsoftwareautomaticallyextractsarangeofclinicalfeaturesspecific

to the three tasks.A full listof the task featuresareprovided in figure2. The full

featuresetincluded109parametersfromanalysisofthegaittask,33fromthesway

test,and14fromtheTUG.

Anyparticipantstakingmedicationwererecordedinthe‘ONmedication’state.

Tasks

Allparticipantscompletedatwominutewalkonthesamestraightandlevelsurface

torecordtheirgait.Next,tomeasuresway,participantswereaskedtostandupright

andasstillaspossibleforthirtysecondswiththeireyesclosedonafirmsurface.A

woodentemplatewasplacedonthefloorsoastoensureallparticipants’feetwere

thesamedistanceapartforthetest.TheTUGwasthenrepeatedthreetimes.Each

repetition entailed the participant starting in a sitting position for three seconds,

followedby standingupwhen instructed,walking forward3metres, performinga

180degreeturn,walkingbacktothechairandsittingbackdown.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

7

HypothesisTesting

The first aspect of this analysis aims to identify features that demonstrate

statisticallysignificantdifferencesbetweenthediseasegroups.Specifically,weaim

todeterminewhetherthePSPgroupshowsadifferencetoHCandPDparticipantsin

a feature.Eachparticipantcontributesone instance ineachof thethreetasks.For

each task type (gait, sway, and TUG), independent t-tests are used to inspect for

statistical significance between the three disease groups for all features using a

significancelevelof0.05.

DiseaseClassification

Inorder toassess towhatextentPSP canbeautomaticallydiscriminated fromPD

and HC subjects based on the measured features, we performed automated

classification with a machine-learning algorithm. For each test, repeated 10-fold

crossvalidationisperformedbasedonthefullfeaturesetfromeachtestseparately.

Thismeansthatthefulldatasetissplitinrandomlyinto10subsets,where9subsets

areusedfortrainingthemachinelearningmodelandtheremainingsubsetisusedas

an independent validation set. This training and validation is repeated 10 times

whereeachtimeadifferentindependentvalidationsetisused.Withineachrepeat

of cross validation, the cross sectional baseline subset undergoes minority class

balancing prior to being assigned to a fold. Within each fold, the training and

validation setsundergo zero-meanunit-variancenormalisationwith respect to the

meanandstandarddeviationofthetrainingset.Duetothelikelihoodthatmanyof

the features within each test are correlated (e.g. Gait Speed Left and Gait Speed

Right),featureselection(LeastAbsoluteShrinkageandSelectionOperator,LASSO)is

performed on each training set, with the selected features subsequently being

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

8

extracted fromthevalidationset.Twoclassifierswereused. ALogisticRegression

(LR) classifierwasuseddue to its popularity in the clinical field.ARandomForest

(RF)classifierwasalsoused,withoutfeaturereductionasRandomForestclassifiers

areknowntodealwellwithahighdimensionalfeatureset.Wereporttheaccuracy,

sensitivity,andspecificityforeachtaskseparatelyandalltaskscombined.

Reducedsensorset

In order to assess the necessity or not of using a large sensor set, we took the

featuresselectedbyLASSOforall taskscombinedandassessedfromwhichsensor

thosefeaturesarecomputed.Wethenrepeatedtheclassificationprocedureusing

only those features extracted from the lumbar sensor, because this sensor alone

accountedforamajorityofthefeaturesidentifiedbyLASSO.Wefurtherrepeated

the classificationusingdataobtainable from the lumbar sensorplus the right arm

andrightlegsensors,andreportedthesameperformancemetrics.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

9

Results

Participants

Thedemographics,clinicalcharacteristics,andratingscalescoresoftheparticipants

foreachgroupareshownintable1.

Figure1 shows thedistributionsof someof the feature valuesextracted from the

three tasks thatwere found todiffer significantlybetweenPSPand the twoother

groups.Theboxplots showthe rangeofvalues foreachof these features ineach

participant group, with individual data points overlaid in grey. Note that there is

considerable redundancy amongst the large number of features, and not all

measurements that were found to be significant individually emerged from the

LASSOasindependentpredictors.Forexampleintheswaytest,meancoronalsway

velocitywasindividuallysignificant,butthisdidnotfeatureintheLASSOresult.

Table2showsthediseaseclassificationaccuracyunderanumberofconditions.The

upper three rows of the table show the accuracy, sensitivity, and specificity of

discriminationbetweenPSPandPDandbetweenPSPandHC,foreachofthethree

tasks separately as indicated in the left column. It can be seen that of the three

tasks,swayismuchlesseffectivethantheothertwoatdiscriminatingbetweenPSP

andPD.Thefourthrowoftable2givestheresultsofanalysingthecombinedfeature

setfromallthreetasks;thisperformsbetterthananyofthe3tasksindividually.

ItisclearthattheRFclassifierperformsbetterthantheLRclassifieroverall.When

distinguishingPSPfromPDthesensitivity,specificity,andaccuracyofRFwasasgood

asorbetterthanLRineveryconditioninthetable.Thegreatestdifferencewas15

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

10

percentage points in specificity using the combined tasks and full sensor set (75%

withLRversus90%withRF).

Table 3 lists the features that emerged from the LASSO analysis of the combined

tasks as the most important discriminators between the groups, with their

weightingsandwhichsensor inthearraythedataareobtainedfromineachcase.

Notably, the lumbar sensorprovides thedata for tenof the seventeen features in

thislist,comparedtojustonefeaturethatisdependentonthesternalsensor.This

promptsonetoaskwhattheresultswouldbeifthelumbarsensorwereusedalone.

The answer (table 2 row 5) is that using only one sensor in the lumbar position

reducestheaccuracyofbothclassifiersmodestly,from80%to78%withLRandfrom

88%to85%withRF.

We thenexplored theeffectof addingonearmandone leg sensor to the lumbar

sensor, making a three sensor array including lumbar, right arm and right foot.

Where LASSO had identified left sided parameter we substituted the equivalent

parameter from the opposite side, making the assumption that, for example,

cadence measured from either leg would report a similar value. Using this

intermediatesizenetwork,theaccuracywithbothclassifierswasthesameasforthe

fullsixsensornetwork(table2row6).Notablyhowever,thehighspecificity(90%)

observedwith the 6 sensor networkwas reduced to 85%with the lumbar sensor

alone,andthiswasnotrecoveredusing3sensors.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

11

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

12

Discussion

Wehaveshownthatitispossiblewithabody-wornIMUarrayandmachinelearning

methodstodifferentiatePSPfromPDandPSPfromcontrol,withahighdegreeof

accuracy. In order to ensure the robustness of these findings we used separate

trainingandvalidationdatasets fortheanalysis,andtoprovidevalidationweused

two different classification algorithms. Overall, the Random Forest classifier

performed better than the Logistic Regression classifier. LR is probably the most

commonly used classifier in biomedical research and therefore the apparent

superiorityofRFisanimportantfinding.

Arecentreviewfound78publishedstudiesofIMUbasedgaitanalysis,butonly16%

oftheseusedmorethanonesensor(18).Wefoundjusttwopreviousstudieslooking

atgaitinPSP.OneoftheseusedasingleIMUinthelumbarregion(8),anddetected

only a difference in vertical displacement during gait between PSP and PD. The

other used two sensors, one attached to each foot (19), and identified significant

differencesingaitspeedandcadence.Inbothcasestheinformationobtainedisfar

lessdetailedthanthatobtainedfrommultiplesensorsinthepresentstudy.

Asexpectedthere isahighdegreeof redundancy in themeasuredfeatureswithin

andacrossthethreetasksused.Thisiswellillustratedbythefactthat,forexample,

the mean velocity of postural sway in the coronal plane was significantly and

substantially differentbetweenPSP and theother groups, yet didnotoccur as an

independentpredictorinthefeaturesetselectedbytheLASSOanalysis.Indeedonly

oneswayvariablewasfoundtocontainindependentinformation.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

13

From the point of view of overall classification accuracy, it may be possible to

simplifythesensorarray.Ofthe17keyparametersidentifiedbytheLASSOanalysis,

10couldbeobtained from just the lumbar sensor. IndistinguishingPSP fromHC,

theaccuracy,sensitivity,andspecificityusingjustthissensorwasvirtually identical

tothatobtainedusingthefullsensorarray.Usingjustthelumbarsensordiddegrade

performance in termsof test accuracywhendifferentiatingPSP fromPD,butonly

modestly (for the RF classifier there was a test accuracy of 88% with six sensors

versus 85% with the lumbar sensor only). Using an intermediate 3 sensor array

(lumbar, right arm, right foot) gave a test accuracy as good as the full six sensor

array.

In some respects however themost important parameter is specificity, because a

higher specificity increases positive predictive value (PPV), which is critical if PSP

specific treatments are tobe initiatedbasedon the result. The specificityof 90%

usingthe6sensorarrayandtheRFclassifierisremarkablygood.Previousbiomarker

studiesusingdiverseapproacheshavenotyieldedspecificitiesashighas this. For

example,MRImorphometryofthebrainstemgaveaspecificityof85%(20),whilea

meta-analysis of studies of probably the most promising CSF biomarker,

neurofilament light chain, gave a specificity of 81%(21). Because of the low

prevalenceofPSPamongstpatientspresentingwithparkinsonism,eventhehigher

85%figuretranslatedintoaPPVofjust57%(20).TheimprovementinPPVbetween

atestwith85%and90%specificityissubstantial.Inmovingfromthefull6sensors

to just the lumbar sensor, the 90% specificity of our RF classifier fell to 85%, an

important reduction in efficacy. Performance was not restored by adding in two

moresensors.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

14

Itshouldbenotedthatthejobthattheclassifierswereaskedtodointhisstudyis

considerably easier than the applications to which we hope the approach will

ultimatelyextend. It isnosurprisetoseeclassificationaccuraciesinexcessof90%

when discriminating PSP patients from controls. The fact that the classification

accuracy between PSP and PD was not far behind (88% with RF) may seem

impressive,butwebeganwithgroupsofclear-cutPSPandPD,whowouldbereadily

distinguishableclinicallybyanexperiencedmovementdisordersneurologist.Forthe

technique to be most useful, it must do something that such a person finds

challenging. The performance differential between more complete and simpler

sensor sets may well widen with more difficult questions and will need to be

carefullyevaluatedineachsituation.

The most obvious next step in testing this approach would be to evaluate its

effectiveness in the differential diagnosis of these conditions at an earlier stage,

when there is still substantial clinical diagnostic uncertainty. Thiswill require the

captureofalargeincidentcohortofparkinsonianpatients,inorderthatitturnsout

eventuallytocontainsufficientnumbersofcasesthatmanifestasPSP,ratherthan

thefarcommonerPD.Theothertaskthatcouldbenefitfromthemachinelearning

approach is the related one of disease severity quantification and progression

tracking. This is an area of great interest because objective numerical biomarkers

capableofdoingthisinPSP(andPD)arelacking,yetverymuchneededfortrialsof

potentialdiseasemodifyingdrugs.

In conclusion, the answer to the question of whether the sensor network can be

simplifieddependsoncontextandtheexactparametersweareinterestedin.Itmay

be that themore complex sensor array is better suited to the diagnostic setting,

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

15

where specificity is paramount, but then a simpler arrangement can be used for

quantificationandtrackingoncethediagnosisissecure.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

16

ACKNOWLEDGMENTS

WethankthepatientsandtheirfamiliesfortheirendlesssupportduringtheOxQUIPstudy.

AUTHORS’ROLES

Researchproject:Conception,organization,executionCAA,JFF

Statisticalanalysis:Design,execution,review&critiqueMDV,JP,JJFandCAA

Manuscript:Drafting,review&critiqueMDV,JP,TB,JJFandCAA

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

17

Figure1:ExamplesofgaitparametersthatdistinguishPSPfromPDandHC.A:Gaitcadence;B:Meanposturalswayvelocityinthecoronalplaneduringtheswaytest;C:Meantimetakentositfromstandingduringthetimedup-and-go(TUG)task;D:Meantimetakentoturnduringthegaittask;E:MeantimetakentoturnduringtheTUGtask;F:Standarddeviationoftimetakentoturnduringthegaittask.Notethatthe large number of parameters generated by the three tasks have considerableredundancy so that some parameters that are significantly different between theconditionsindividuallyarenotintheLASSOoutput.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

18

List of features that are quantified for each task Gait: 'AnticipatoryPosturalAdjustment-APADuration(s)''AnticipatoryPosturalAdjustment-FirstStepDuration(s)''AnticipatoryPosturalAdjustment-FirstStepRangeofMotion(degrees)''AnticipatoryPosturalAdjustment-ForwardAPAPeak(m/s^2)''AnticipatoryPosturalAdjustment-LateralAPAPeak(m/s^2)''Duration(s)''Gait-LowerLimb-CadenceL(steps/min)''Gait-LowerLimb-CadenceR(steps/min)''Gait-LowerLimb-DoubleSupportL(%GCT)''Gait-LowerLimb-DoubleSupportR(%GCT)''Gait-LowerLimb-ElevationatMidswingL(cm)''Gait-LowerLimb-ElevationatMidswingR(cm)''Gait-LowerLimb-GaitCycleDurationL(s)''Gait-LowerLimb-GaitCycleDurationR(s)''Gait-LowerLimb-GaitSpeedL(m/s)''Gait-LowerLimb-GaitSpeedR(m/s)''Gait-LowerLimb-LateralStepVariabilityL(cm)''Gait-LowerLimb-LateralStepVariabilityR(cm)''Gait-LowerLimb-CircumductionL(cm)''Gait-LowerLimb-CircumductionR(cm)''Gait-LowerLimb-N(#)''Gait-LowerLimb-FootStrikeAngleL(degrees)''Gait-LowerLimb-FootStrikeAngleR(degrees)''Gait-LowerLimb-ToeOffAngleL(degrees)''Gait-LowerLimb-ToeOffAngleR(degrees)''Gait-LowerLimb-SingleLimbSupportL(%GCT)''Gait-LowerLimb-SingleLimbSupportR(%GCT)''Gait-LowerLimb-StanceL(%GCT)''Gait-LowerLimb-StanceR(%GCT)''Gait-LowerLimb-StepDurationL(s)''Gait-LowerLimb-StepDurationR(s)''Gait-LowerLimb-StrideLengthL(m)''Gait-LowerLimb-StrideLengthR(m)''Gait-LowerLimb-SwingL(%GCT)''Gait-LowerLimb-SwingR(%GCT)''Gait-LowerLimb-TerminalDoubleSupportL(%GCT)''Gait-LowerLimb-TerminalDoubleSupportR(%GCT)''Gait-LowerLimb-ToeOutAngleL(degrees)''Gait-LowerLimb-ToeOutAngleR(degrees)''Gait-Lumbar-CoronalRangeofMotion(degrees)''Gait-Lumbar-SagittalRangeofMotion(degrees)''Gait-Lumbar-TransverseRangeofMotion(degrees)''Gait-Trunk-CoronalRangeofMotion(degrees)'

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

19

'Gait-Trunk-SagittalRangeofMotion(degrees)''Gait-Trunk-TransverseRangeofMotion(degrees)''Gait-UpperLimb-ArmSwingVelocityL(degrees/s)''Gait-UpperLimb-ArmSwingVelocityR(degrees/s)''Gait-UpperLimb-ArmRangeofMotionL(degrees)''Gait-UpperLimb-ArmRangeofMotionR(degrees)''Turns-Angle(degrees)''Turns-Duration(s)''Turns-N(#)''Turns-TurnVelocity(degrees/s)''Turns-StepsinTurn(#)''AnticipatoryPosturalAdjustment-APADuration(s)STD''AnticipatoryPosturalAdjustment-FirstStepDuration(s)STD''AnticipatoryPosturalAdjustment-FirstStepRangeofMotion(degrees)STD''AnticipatoryPosturalAdjustment-ForwardAPAPeak(m/s^2)STD''AnticipatoryPosturalAdjustment-LateralAPAPeak(m/s^2)STD''Duration(s)STD''Gait-LowerLimb-CadenceL(steps/min)STD''Gait-LowerLimb-CadenceR(steps/min)STD''Gait-LowerLimb-DoubleSupportL(%GCT)STD''Gait-LowerLimb-DoubleSupportR(%GCT)STD''Gait-LowerLimb-ElevationatMidswingL(cm)STD''Gait-LowerLimb-ElevationatMidswingR(cm)STD''Gait-LowerLimb-GaitCycleDurationL(s)STD''Gait-LowerLimb-GaitCycleDurationR(s)STD''Gait-LowerLimb-GaitSpeedL(m/s)STD''Gait-LowerLimb-GaitSpeedR(m/s)STD''Gait-LowerLimb-LateralStepVariabilityL(cm)STD''Gait-LowerLimb-LateralStepVariabilityR(cm)STD''Gait-LowerLimb-CircumductionL(cm)STD''Gait-LowerLimb-CircumductionR(cm)STD''Gait-LowerLimb-N(#)STD''Gait-LowerLimb-FootStrikeAngleL(degrees)STD''Gait-LowerLimb-FootStrikeAngleR(degrees)STD''Gait-LowerLimb-ToeOffAngleL(degrees)STD''Gait-LowerLimb-ToeOffAngleR(degrees)STD''Gait-LowerLimb-SingleLimbSupportL(%GCT)STD''Gait-LowerLimb-SingleLimbSupportR(%GCT)STD''Gait-LowerLimb-StanceL(%GCT)STD''Gait-LowerLimb-StanceR(%GCT)STD''Gait-LowerLimb-StepDurationL(s)STD''Gait-LowerLimb-StepDurationR(s)STD''Gait-LowerLimb-StrideLengthL(m)STD''Gait-LowerLimb-StrideLengthR(m)STD''Gait-LowerLimb-SwingL(%GCT)STD''Gait-LowerLimb-SwingR(%GCT)STD'

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

20

'Gait-LowerLimb-TerminalDoubleSupportL(%GCT)STD''Gait-LowerLimb-TerminalDoubleSupportR(%GCT)STD''Gait-LowerLimb-ToeOutAngleL(degrees)STD''Gait-LowerLimb-ToeOutAngleR(degrees)STD''Gait-Lumbar-CoronalRangeofMotion(degrees)STD''Gait-Lumbar-SagittalRangeofMotion(degrees)STD''Gait-Lumbar-TransverseRangeofMotion(degrees)STD''Gait-Trunk-CoronalRangeofMotion(degrees)STD''Gait-Trunk-SagittalRangeofMotion(degrees)STD''Gait-Trunk-TransverseRangeofMotion(degrees)STD''Gait-UpperLimb-ArmSwingVelocityL(degrees/s)STD''Gait-UpperLimb-ArmSwingVelocityR(degrees/s)STD''Gait-UpperLimb-ArmRangeofMotionL(degrees)STD''Gait-UpperLimb-ArmRangeofMotionR(degrees)STD''Turns-Angle(degrees)STD''Turns-Duration(s)STD''Turns-N(#)STD''Turns-TurnVelocity(degrees/s)STD''Turns-StepsinTurn(#)STD'

Sway: 'PosturalSway-Acc-95%EllipseAxis1Radius(m/s^2)''PosturalSway-Acc-95%EllipseAxis2Radius(m/s^2)''PosturalSway-Acc-95%EllipseRotation(m/s^2)''PosturalSway-Acc-95%EllipseSwayArea(m^2/s^4)''PosturalSway-Acc-CentroidalFrequency(Hz)''PosturalSway-Acc-CentroidalFrequency(Coronal)(Hz)''PosturalSway-Acc-CentroidalFrequency(Sagittal)(Hz)''PosturalSway-Acc-FrequencyDispersion(AD)''PosturalSway-Acc-FrequencyDispersion(Coronal)(AD)''PosturalSway-Acc-FrequencyDispersion(Sagittal)(AD)''PosturalSway-Acc-Jerk(m^2/s^5)''PosturalSway-Acc-Jerk(Coronal)(m^2/s^5)''PosturalSway-Acc-Jerk(Sagittal)(m^2/s^5)''PosturalSway-Acc-MeanVelocity(m/s)''PosturalSway-Acc-MeanVelocity(Coronal)(m/s)''PosturalSway-Acc-MeanVelocity(Sagittal)(m/s)''PosturalSway-Acc-PathLength(m/s^2)''PosturalSway-Acc-PathLength(Coronal)(m/s^2)''PosturalSway-Acc-PathLength(Sagittal)(m/s^2)''PosturalSway-Acc-RMSSway(m/s^2)''PosturalSway-Acc-RMSSway(Coronal)(m/s^2)''PosturalSway-Acc-RMSSway(Sagittal)(m/s^2)''PosturalSway-Acc-Range(m/s^2)''PosturalSway-Acc-Range(Coronal)(m/s^2)''PosturalSway-Acc-Range(Sagittal)(m/s^2)''PosturalSway-Angles-SwayAreaRadius(Coronal)(degrees)'

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

21

'PosturalSway-Angles-95%EllipseAxis2Radius(degrees)''PosturalSway-Angles-SwayAreaRotation(degrees)''PosturalSway-Angles-SwayArea(degrees^2)''PosturalSway-Angles-Duration(s)''PosturalSway-Angles-RMSSway(degrees)''PosturalSway-Angles-RMSSway(Coronal)(degrees)''PosturalSway-Angles-RMSSway(Sagittal)(degrees)'

TUG: 'Duration(s)''SittoStand-Duration(s)''SittoStand-LeanAngle(degrees)''SittoStand-N(#)''StandtoSit-Duration(s)''StandtoSit-LeanAngle(degrees)''StandtoSit-N(#)''Turns-Angle(degrees)''Turns-Duration(s)''Turns-N(#)''Turns-TurnVelocity(degrees/s)''Turns-Angle(degrees)STD''Turns-Duration(s)STD''Turns-TurnVelocity(degrees/s)STD'

Figure2.Afulllistoftaskfeatures.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

22

References

1. LitvanI,AgidY,CalneD,CampbellG,DuboisB,DuvoisinRC,etal.Clinicalresearchcriteria for the diagnosis of progressive supranuclear palsy (Steele-Richardson-Olszewskisyndrome):reportoftheNINDS-SPSPinternationalworkshop.Neurology.1996Jul;47(1):1-9.PubMedPMID:8710059.eng.2. Kuniyoshi S, Riley DE, Zee DS, Reich SG, Whitney C, Leigh RJ. Distinguishingprogressivesupranuclearpalsy fromother formsofParkinson'sdisease:evaluationofnewsigns. Annals of the New York Academy of Sciences. 2002 Apr;956:484-6. PubMed PMID:11960847.eng.3. Fearnley JM, Lees AJ. Ageing and Parkinson's disease: substantia nigra regionalselectivity.Brain.1991Oct;114(Pt5):2283-301.PubMedPMID:1933245.eng.4. BoxerAL,Yu JT,GolbeLI, Litvan I, LangAE,HoglingerGU.Advances inprogressivesupranuclearpalsy:newdiagnosticcriteria,biomarkers,andtherapeuticapproaches.Lancetneurology. 2017 Jul;16(7):552-63. PubMed PMID: 28653647. Pubmed Central PMCID:5802400.5. Lamb R, Rohrer JD, Lees AJ, Morris HR. Progressive Supranuclear Palsy andCorticobasal Degeneration: Pathophysiology and Treatment Options. Current treatmentoptionsinneurology.2016Sep;18(9):42.PubMedPMID:27526039.PubmedCentralPMCID:4985534.6. Schrag A, Ben-Shlomo Y, Quinn NP. Prevalence of progressive supranuclear palsyandmultiplesystematrophy:across-sectionalstudy.Lancet.1999Nov20;354(9192):1771-5.PubMedPMID:10577638.eng.7. HoglingerGU,RespondekG,StamelouM,KurzC,JosephsKA,LangAE,etal.Clinicaldiagnosis of progressive supranuclear palsy: Themovement disorder society criteria.MovDisord.2017Jun;32(6):853-64.PubMedPMID:28467028.PubmedCentralPMCID:5516529.8. Hatanaka N, Sato K, Hishikawa N, Takemoto M, Ohta Y, Yamashita T, et al.Comparative Gait Analysis in Progressive Supranuclear Palsy and Parkinson's Disease.Europeanneurology.2016;75(5-6):282-9.PubMedPMID:27288001.9. Murray MP, Sepic SB, Gardner GM, Downs WJ. Walking patterns of men withparkinsonism. American journal of physical medicine. 1978 Dec;57(6):278-94. PubMedPMID:742658.10. Morris ME, Martin CL, Schenkman ML. Striding out with Parkinson disease:evidence-basedphysicaltherapyforgaitdisorders.Physicaltherapy.2010Feb;90(2):280-8.PubMedPMID:20022998.PubmedCentralPMCID:2816030.11. Hausdorff JM,CudkowiczME,FirtionR,Wei JY,GoldbergerAL.Gaitvariabilityandbasalgangliadisorders:stride-to-stridevariationsofgaitcycletiminginParkinson'sdiseaseandHuntington'sdisease.MovDisord.1998May;13(3):428-37.PubMedPMID:9613733.12. Egerton T,Williams DR, Iansek R. Comparison of gait in progressive supranuclearpalsy, Parkinson's disease and healthy older adults. BMC neurology. 2012 Oct 2;12:116.PubMedPMID:23031506.PubmedCentralPMCID:3517411.13. Godinho C, Domingos J, Cunha G, Santos AT, Fernandes RM, Abreu D, et al. Asystematic review of the characteristics and validity of monitoring technologies to assessParkinson's disease. Journal of neuroengineering and rehabilitation. 2016 Mar 12;13:24.PubMedPMID:26969628.PubmedCentralPMCID:4788909.14. BrognaraL,PalumboP,GrimmB,PalmeriniL.AssessingGaitinParkinson'sDiseaseUsingWearableMotionSensors:ASystematicReview.Diseases.2019Feb5;7(1).PubMedPMID:30764502.PubmedCentralPMCID:6473911.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

23

15. BastonC,ManciniM,SchoneburgB,HorakF,RocchiL.Posturalstrategiesassessedwithinertialsensorsinhealthyandparkinsoniansubjects.Gait&posture.2014;40(1):70-5.PubMedPMID:24656713.PubmedCentralPMCID:4383136.16. ManciniM,Carlson-KuhtaP,ZampieriC,NuttJG,ChiariL,HorakFB.Posturalswayasamarker of progression in Parkinson's disease: a pilot longitudinal study. Gait& posture.2012Jul;36(3):471-6.PubMedPMID:22750016.PubmedCentralPMCID:3894847.17. ManciniM, Zampieri C, Carlson-Kuhta P, Chiari L, Horak FB. Anticipatory posturaladjustments prior to step initiation are hypometric in untreated Parkinson's disease: anaccelerometer-based approach. Eur J Neurol. 2009 Sep;16(9):1028-34. PubMed PMID:19473350.PubmedCentralPMCID:2840629.18. Vienne A, Barrois RP, Buffat S, Ricard D, Vidal PP. Inertial Sensors to Assess GaitQuality in Patients with Neurological Disorders: A Systematic Review of Technical andAnalyticalChallenges.FrontPsychol.2017;8:817.PubMedPMID:28572784.PubmedCentralPMCID:PMC5435996.19. Raccagni C, Gassner H, Eschlboeck S, Boesch S, Krismer F, Seppi K, et al. Sensor-based gait analysis in atypical parkinsonian disorders. Brain Behav. 2018 Jun;8(6):e00977.PubMedPMID:29733529.PubmedCentralPMCID:PMC5991583.20. Kim YH, Ma HI, Kim YJ. Utility of the Midbrain Tegmentum Diameter in theDifferentialDiagnosisofProgressiveSupranuclearPalsyfromIdiopathicParkinson'sDisease.Journal of clinical neurology. 2015 Jul;11(3):268-74. PubMed PMID: 26174787. PubmedCentralPMCID:4507382.21. Ge F, Ding J, Liu Y, Lin H, Chang T. Cerebrospinal fluid NFL in the differentialdiagnosis of parkinsonian disorders: A meta-analysis. Neuroscience letters. 2018 Oct15;685:35-41.PubMedPMID:30036569.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

24

PD(n=20)

Mean(range)

PSP(n=21)

Mean(range)

HC(n=39)

Mean(range)

Age,yrs 66.4(50-79) 71(63-89) 67.1(51-82)

Gender,Male/Female 11:9 12:9 19:20

Diseaseduration(yrs) 11.4 2.0 NA

UPDRSmotorscore(PartIII) 27.9(9-52) 44.6(21-72) 3.1(0-12)

MMSE 26.6(25-29) 25.8(20-30) 27.6(26-30)

MOCA 26.6(24-30) 22.0(12-29) 28.5(27-30)

Table 1. Demographics, clinical characteristics and cognitive scores in the 3 groups. PD =

Parkinson’sdisease,HC=healthycontrolsandPSP=ProgressiveSupranuclearPalsy.UPDRS=

UnifiedParkinson’sDiseaseRatingScale,MMSE=MiniMentalstateexamination.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

25

Task Sensorsused ComparisonLogisticRegression RandomForest

Acc(%) Sens(%) Spec(%) Acc(%)Sens(%) Spec(%)

GaitLumbar,sternal,both

arms,bothfeetPSPvsHC 87 86 87 91 85 94

PSPvsPD 80 81 75 83 81 85

Sway Lumbar,sternal,botharms,bothfeet

PSPvsHC 68 52 77 82 67 90

PSPvsPD 63 66 60 63 67 60

TUGLumbar,sternal,both

arms,bothfeetPSPvsHC 90 81 97 92 86 95

PSPvsPD 70 71 70 83 86 80

Combined Lumbar,sternal,botharms,bothfeet

PSPvsHC 93 86 97 95 90 97

PSPvsPD 80 85 75 88 86 90

Combined LumbaronlyPSPvsHC 93 85 97 95 90 97

PSPvsPD 78 76 80 85 86 85

Combined Lumbar,rightarm,rightfoot

PSPvsHC 93 89 96 93 90 97

PSPvsPD 80 85 75 88 90 85

Table 2. Classification results when automatically classifying different patients based on their gait,

sway,andTUGsignaturesusingtheentiresix-sensorarray(firstthreerowsoftable).Swayistheleast

informative test. Classification accuracy is improvedwhen the feature sets fromall three tasks are

merged (fourth row). TheRandomForest (RF) classifier performsbetter than the logistic regression

(LR)classifier.Usingdatafromthelumbarsensoralone(row5)reducesaccuracymodestly,butthiscan

berecoveredbyaddingbackinarmandfootsensorsfromonesideofthebody(lastrowoftable).The

high specificityofPSPversusPDclassificationobtainedwithRF (90%) isonly seenwhenusing the6

sensorarray.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint

26

Activity Feature Weighting SensorAll

sensorsLumbaronly

Lumbar,arm,leg

TUG 'Turns-Duration(s)' 0.11328 Lumbar X X X

XX 'Gait-LowerLimb-CadenceL(steps/min)STD' 0.09398 LeftLeg X

X

TUG 'Turns-Angle(degrees)STD' 0.06248 Lumbar X X X

TUG 'SittoStand-Duration(s)' 0.04518 Lumbar X X X

Gait 'Gait-Trunk-SagittalRangeofMotion(degrees)STD' -0.03971 Trunk X

Gait 'Gait-LowerLimb-ToeOffAngleR(degrees)STD' 0.03182 RightLeg X

X

TUG 'Turns-Duration(s)STD' 0.03071 Lumbar X X X

TUG 'StandtoSit-LeanAngle(degrees)' -0.03009 Lumbar X X X

TUG 'Turns-Angle(degrees)' -0.02602 Lumbar X X X

TUG 'Duration(s)' 0.01957 Lumbar X X X

Gait 'Gait-UpperLimb-ArmSwingVelocityR(degrees/s)' -0.01876 RightArm X

X

Gait 'Gait-UpperLimb-ArmRangeofMotionR(degrees)' -0.01631 RightArm X

X

Gait 'Turns-StepsinTurn(#)STD' 0.01618 Lumbar X X X

Sway 'PosturalSway-Acc-FrequencyDispersion(AD)' 0.01395 Lumbar X X X

Gait 'Gait-UpperLimb-ArmRangeofMotionL(degrees)' -0.00735 LeftArm X

X

Gait 'Gait-LowerLimb-ToeOutAngleL(degrees)' 0.00101 LeftLeg X

X

Gait 'Turns-Duration(s)' 0.00025 Lumbar X X X

Table 3. Key parameters emerging from LASSO analysis discriminating PSP from PD and HC subjects,

togetherwiththeirrelativeimportance(asreflectedbytheweights).The‘sensor’columnexplainswhich

ofthesensorsinthearrayisresponsibleforcollectingthedatadescribed,andtheactivitycolumnonthe

leftshowsduringwhichtest the featurehasbeencollected.Thethreecolumnsontherightshowhow

muchofthefeaturesetisavailabledependingontheextentofsensorarrayused.

All rights reserved. No reuse allowed without permission. was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint (whichthis version posted November 6, 2019. .https://doi.org/10.1101/19006866doi: medRxiv preprint