biogears iccai poster
TRANSCRIPT
Predictive Analytics for Patient Monitoring
Methodology
Mean Arterial Pressure (M
AP) Prediction Results
Predicting with
Point of Contact
Detailed Comparison(s) of Test Set Specim
ensAggregated Prediction and Error
InputData.
TheU
SAISRprovided
datafor
35sw
inem
odelswith
severehem
orrhageswithdraw
ing~65%
ofblood
volume
priorto
intervention.The
IntegratedData
Exchangeand
Archival(IDEA)
systemrecorded
131tim
eseries
valueselectronically
from:the
DragerInfinity
Deltapatientm
onitor(D),theEdw
ardsEV1000
(E),and
theAesculon
criticalcare
monitor
(A).W
erandom
lysplit
ourspecim
ensinto
two
poolsfor
training(66%
)andtesting
(33%)ourlearned
model.
Machine
Learning.W
eused
PrincipleCom
ponentAnalysis
/Singular
ValueDecom
positionto
calculateeigenvalues
andeigenvectors
todow
n-selectand
weight27
variablesto
trainfourAutoregressive
LinearM
odelsfor1,5,10,and15
minutes
outpredictions:
Mean % Error
We com
pare Actual MAP m
easures to CARP predictions made 1, 5, 10, and 15 m
inutes before, from start of bleed to nadir.
Forecasting the Trajectory of Blood Loss from Vital Signs Collected at the Bedside
Objective:
We
developedand
evaluatedCARP,a
PrincipleCom
ponentAnalysisdriven
Autoregressivetim
eseries
modelthatforw
ardpredicts
ArterialBloodPressure
duringa
hemorrhage
event.Our
inputdata
(previouslycollected
bythe
US
Army
Instituteof
SurgicalResearch)included
131tim
eseries
valuesrecorded
orderivedduring
atestprocedure
(describedbelow
).We
down
selectedthe
inputforourmodelto
27variables.Data
were
recordedelectronically
usingcustom
software
(IntegratedData
Exchangeand
Archival,IDEA).
Conclusion:W
eshow
edhigh
correlationofM
eanArterialPressure
(MAP)prediction
atshort
term(w
itha
2.37%and
-1.9%prediction
percentdifference
errorfor1and
5m
inrespectively),and
moderate
correlationatlong
termpredictions
(with
a-4.98%
and-6.72%
predictionaverage
percentdifferenceerror
for10and
15m
inrespectively)using
ourCARPalgorithm
.Thisim
pliesthatthe
model,on
average,slightlyunder-predicts
MAP
forlongterm
predictionsThe
incorporationof
suchalgorithm
sin
abedside
monitoring
platformor
ina
medicalprofessional’s
fieldm
onitoringdevice
couldinform
providersof
criticaldecline
inblood
pressureleading
totim
elylive-saving
interventions.
Charles E Fisher 1, Randall J Frank1, Chris Petrovitch
1, Eddie Oxford
1, Chris Argenta1,
William
L Baker Jr 2, Timothy S. Park 2, Daniel S W
endroff 2, Leopoldo C Cancio2, Andriy
I Batchinsky2,3, Jerem
y C Pamplin
2
1Applied Research Associates, Raleigh NC 2U
S Army Institute for Surgical Research, San Antonio TX 3Geneva Foundation
Thedata
collectedfor
thisw
orkw
assupported
bythe
U.S.Arm
yM
edicalResearch
andM
aterielCom
mand
(USAM
RMC)
inFort
Detrick,M
Dunder
ContractN
umber:W
81XWH-13-2-0068.
Charles E FisherApplied Research Associates
8537 Six Forks Road Suite 600Raleigh, N
C 27615(919) 582-3300
Dimensionality reduction and w
eighting for models.
TheR-squared
valuesm
onotonicallydecreased
asthe
forecastingtim
eincreased.
Theydropped
from0.9
(forecasting1
minute
ahead)to
0.7(forecasting
15m
inutesinto
thefuture).
All12ofthe
projectedfeature
vectorsused
inthe
modelsare
significantusingan
αof0.05.
BioGears®is
afullbody
physiologicalmodeling
tool.It
modelsofvarioussystem
sasintegratedRC
circuits.The
BioGears®Cardiovascular
Modelis
aclosed-loop
systemrepresenting
systemic
behavior,pulm
onarycirculation,
andthe
heart’sdynam
icpum
ping.This
modelis
basedon
Guyton'sfour-com
partment(three
vascular,plusoneheart)m
odeloftheCV
System.This
modelhas
beenvalidated
atresting
anda
varietyof
cardiovascularrelatedinsultsand
interventions.See
ww
w.biogearsengine.com
fordetails.W
escaled
ourtestspecimens
tothe
expectedhum
anranges
andgenerated
MAP
valuesfor
asim
ilarhem
orrhageevent
usingthe
fixedm
odelin
BioGears®.This
RCm
odelsim
ulatesexpected
physiology,w
hilethe
CARPm
odelrespondsdirectly
tothe
specificsofthesensorreadingslive.
OriginalData
CleanedData
TrainingSet
TestSet
PCA
~200 Attributes
27 Attributes
12 Eigenvectors
TrainLinear
Regression
ApplyLinear
RegressionP
iβ
i
MAP
Prediction
CARP