Probabilistic modeling of sensor artifacts in critical care Norm - - PowerPoint PPT Presentation

▶
probabilistic modeling of sensor artifacts in critical
SMART_READER_LITE
LIVE PREVIEW

Probabilistic modeling of sensor artifacts in critical care Norm - - PowerPoint PPT Presentation

Oct 04, 2022 227 likes •916 views

Probabilistic modeling of sensor artifacts in critical care Norm Aleks and Stuart J. Russell U.C. Berkeley Computer Science Division Michael G. Madden N.U.I. Galway Dept. of Computer Science Kristan Staudenmayer, Mitchell Cohen, Diane

slide-1
SLIDE 1

Probabilistic modeling of sensor artifacts in critical care

Norm Aleks and Stuart J. Russell U.C. Berkeley Computer Science Division Michael G. Madden N.U.I. Galway Dept. of Computer Science Kristan Staudenmayer, Mitchell Cohen, Diane Morabito, and Geoffrey Manley U.C. San Francisco Dept. of Neurological Surgery and San Francisco General Hospital

slide-2
SLIDE 2
slide-3
SLIDE 3

Raw data

slide-4
SLIDE 4
slide-5
SLIDE 5
slide-6
SLIDE 6
slide-7
SLIDE 7

Remaining in this talk

  • Arterial blood pressure (ABP) and its

measurement artifacts

  • Modeling sub-interval events
  • Our ABP/monitoring artifact model
  • Experimental results
slide-8
SLIDE 8
  • Invasive ABP

measurement apparatus

slide-9
SLIDE 9

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

slide-10
SLIDE 10

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

systolic (maximum)

slide-11
SLIDE 11

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

systolic (maximum) mean

slide-12
SLIDE 12

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

systolic (maximum) mean diastolic (minimum)

slide-13
SLIDE 13

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

slide-14
SLIDE 14

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 3 4 5 6

slide-15
SLIDE 15

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 3 4 5 6

blood draw

slide-16
SLIDE 16

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 3 4 5 6

blood draw line flush

slide-17
SLIDE 17

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 3 4 5 6 3 4 5 6

blood draw line flush

slide-18
SLIDE 18

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 3 4 5 6 3 4 5 6

blood draw line flush zeroing

slide-19
SLIDE 19

ABP measurements, artifacts, and the effects of averaging

0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 3 4 5 6 3 4 5 6

blood draw line flush zeroing

slide-20
SLIDE 20

Physiologic ABP changes

Physiologic hypertension (high blood pressure)

slide-21
SLIDE 21

Physiologic ABP changes

Physiologic hypertension (high blood pressure) Physiologic hypotension (low blood pressure)

slide-22
SLIDE 22

How can we detect and correct for these artifacts?

slide-23
SLIDE 23

How can we detect and correct for these artifacts?

  • By hand?
slide-24
SLIDE 24

How can we detect and correct for these artifacts?

  • By hand?
  • No: too costly, doesn’t scale
slide-25
SLIDE 25

How can we detect and correct for these artifacts?

  • By hand?
  • No: too costly, doesn’t scale
  • Low-pass filter?
slide-26
SLIDE 26

How can we detect and correct for these artifacts?

  • By hand?
  • No: too costly, doesn’t scale
  • Low-pass filter?
  • No: frequencies of physiologic events
  • verlap with those of artifactual events
slide-27
SLIDE 27

How can we detect and correct for these artifacts?

  • By hand?
  • No: too costly, doesn’t scale
  • Low-pass filter?
  • No: frequencies of physiologic events
  • verlap with those of artifactual events
  • Careful system modeling?
slide-28
SLIDE 28

How can we detect and correct for these artifacts?

  • By hand?
  • No: too costly, doesn’t scale
  • Low-pass filter?
  • No: frequencies of physiologic events
  • verlap with those of artifactual events
  • Careful system modeling?
  • Of course!
slide-29
SLIDE 29

What should the model’s timestep be?

slide-30
SLIDE 30

What should the model’s timestep be?

  • One second? (“fast model”)
slide-31
SLIDE 31

What should the model’s timestep be?

  • One second? (“fast model”)

✓artifact occurrence modeling is natural

slide-32
SLIDE 32

What should the model’s timestep be?

  • One second? (“fast model”)

✓artifact occurrence modeling is natural

  • inference runs 59x between evidence
slide-33
SLIDE 33

What should the model’s timestep be?

  • One second? (“fast model”)

✓artifact occurrence modeling is natural

  • inference runs 59x between evidence
  • One minute? (“slow model”)
slide-34
SLIDE 34

What should the model’s timestep be?

  • One second? (“fast model”)

✓artifact occurrence modeling is natural

  • inference runs 59x between evidence
  • One minute? (“slow model”)

✓this is the frequency of the sensor data

slide-35
SLIDE 35

What should the model’s timestep be?

  • One second? (“fast model”)

✓artifact occurrence modeling is natural

  • inference runs 59x between evidence
  • One minute? (“slow model”)

✓this is the frequency of the sensor data

  • it’s difficult to model event occurrence
slide-36
SLIDE 36

f0 f1 f2 f3 f4 f5 f6 f7 f8

Event occurrence: fast model

slide-37
SLIDE 37

f0 f1 f2 f3 f4 f5 f6 f7 f8

fi : 1 if event is in process at timestep i, 0 if it is not.

Event occurrence: fast model

slide-38
SLIDE 38

f0 f1 f2 f3 f4 f5 f6 f7 f8

fi : 1 if event is in process at timestep i, 0 if it is not. p = P( fi=1 | fi–1=1) ... “P(event continues)” q = P( fi=1 | fi–1=0) ... “P(new event starts)”

Event occurrence: fast model

slide-39
SLIDE 39

f0 f1 f2 f3 f4 f5 f6 f7 f8

Relate fast model to slow

slide-40
SLIDE 40

f0 f1 f2 f3 f4 f5 f6 f7 f8

G0 G4 G8

∑ ∑

Relate fast model to slow

slide-41
SLIDE 41

f0 f1 f2 f3 f4 f5 f6 f7 f8

E0 E4 E8 G0 G4 G8

∑ ∑

Relate fast model to slow

slide-42
SLIDE 42

f0 f4

G0 E0 G4 E4

f8

G8 E8

Sum out extra fast nodes

slide-43
SLIDE 43

f0 f1

G0 E0 G1 E1

f2

G2 E2

slide-44
SLIDE 44

f0 f1

G0 E0 G1 E1

slide-45
SLIDE 45

Computational cost

  • Using a dynamic program, computing new

CPTs has time complexity O( j2 N j+1 ), where the fast model has ...

  • j mutually exclusive events to count
  • N timesteps per slow-model timestep
  • Markovian transitions
  • Space complexity is O( j2 N j )
slide-46
SLIDE 46

Parameterizing the fast model for “bag” events

Valve position normal bag normal 1 - q q bag 1 - p p

time t+1 time t

slide-47
SLIDE 47

Parameterizing the fast model for “bag” events

Valve position normal bag normal 1 - q q bag 1 - p p

time t+1 time t q = event count / time = 53 / 557,893 = 0.000129

slide-48
SLIDE 48

Parameterizing the fast model for “bag” events

Valve position normal bag normal 1 - q q bag 1 - p p

time t+1 time t p = 1 - (1 / average event length) = 1-1/26.38 = 0.962 q = event count / time = 53 / 557,893 = 0.000129

slide-49
SLIDE 49

Event durations fit a geometric model well

!" #!" $!" %!" &!" '!" (!" )!" #" %" '" )" *" ##"#%"#'"#)"#*"$#"$%"$'"$)"$*"%#"%%"%'"%)"%*"&#"&%"&'"&)"&*"'#"'%"''"')"'*"(#"(%"('"()"(*")#")%")'"))")*"+#" !"#$%&"'&()*&+,+$%-&./"-+&0#1)2"$&3&4%/+&56)57-&,)8#+&

Bag event durations

slide-50
SLIDE 50

P(G | f0=0) P(G | f0=1) P(f1 | G, f0=1) P(f1 | G, f0=0) f1=0 f1=1 f1=0 f1=1

slide-51
SLIDE 51

f0 f1

G0 E0 G1 E1

slide-52
SLIDE 52

valve- pos1 valve- pos0

slide-53
SLIDE 53

bag- time1 zero- time1 valve- pos1 valve- pos0

slide-54
SLIDE 54

bag- time1 zero- time1 valve- pos1 bag- press1 true- BP1 zero- press1 appa- rent1 valve- pos0

Apparent BP = (1/60) • ( (60 – bag-time – zero-time) • true-BP + bag-time • bag-press + zero-time • zero-press )

slide-55
SLIDE 55

bag- time1 zero- time1 valve- pos1 bag- press1 true- BP1 zero- press1 appa- rent1 meas- ured1 valve- pos0

slide-56
SLIDE 56

bag- time1 zero- time1 valve- pos1 bag- press1 true- BP1 zero- press1 appa- rent1 meas- ured1 bag- time0 zero- time0 valve- pos0 bag- press0 true- BP0 zero- press0 appa- rent0 meas- ured0

slide-57
SLIDE 57

Performance evaluation

slide-58
SLIDE 58
  • ~300 hrs. data for training and evaluation
  • Ground truth from a physician, using

inspection of data at 1-second resolution

  • “bag” and “zero” events
  • hypertension (true systolic > 160mmHg)

Performance evaluation

slide-59
SLIDE 59
  • ~300 hrs. data for training and evaluation
  • Ground truth from a physician, using

inspection of data at 1-second resolution

  • “bag” and “zero” events
  • hypertension (true systolic > 160mmHg)
  • Particle filtering used, with 8000 particles
  • CPU time for filtering about 2 hours

Performance evaluation

slide-60
SLIDE 60

Detection of “bag” events

slide-61
SLIDE 61

Detection of “zero” events

slide-62
SLIDE 62

Detection of hypertension

slide-63
SLIDE 63
slide-64
SLIDE 64
slide-65
SLIDE 65
slide-66
SLIDE 66

Future modeling work

slide-67
SLIDE 67
slide-68
SLIDE 68