Machine Learning Algorithms for Neuroimaging-based Clinical Trials - - PowerPoint PPT Presentation

▶
machine learning algorithms for neuroimaging based
SMART_READER_LITE
LIVE PREVIEW

Machine Learning Algorithms for Neuroimaging-based Clinical Trials - - PowerPoint PPT Presentation

Mar 16, 2024 42 likes •605 views

Machine Learning Algorithms for Neuroimaging-based Clinical Trials in Preclinical Alzheimers Disease Vamsi K. Ithapu Wisconsin Alzheimers Disease Research Center University of Wisconsin-Madison April 2, 2017 (BRAIN Initiative Symposium

slide-1
SLIDE 1

Machine Learning Algorithms for Neuroimaging-based Clinical Trials in Preclinical Alzheimer’s Disease

Vamsi K. Ithapu

Wisconsin Alzheimer’s Disease Research Center University of Wisconsin-Madison April 2, 2017

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 1 / 18

slide-2
SLIDE 2

A Clinical Trial – The work flow

Randomized Controlled Trial

Study Population (BRAIN Initiative Symposium 2017) Learning methods for enrichment 2 / 18

slide-3
SLIDE 3

A Clinical Trial – The work flow

Randomized Controlled Trial

Study Population Controls Intervention At enrollment (BRAIN Initiative Symposium 2017) Learning methods for enrichment 2 / 18

slide-4
SLIDE 4

A Clinical Trial – The work flow

Randomized Controlled Trial

Study Population Controls Intervention At enrollment Trial ends Controls Intervention (BRAIN Initiative Symposium 2017) Learning methods for enrichment 2 / 18

slide-5
SLIDE 5

A Clinical Trial – The work flow

Randomized Controlled Trial

Study Population Controls Intervention At enrollment Trial ends Controls Intervention Do Controls differ from interneved (BRAIN Initiative Symposium 2017) Learning methods for enrichment 2 / 18

slide-6
SLIDE 6

Setting up a clinical trial – My work

Who is participating in the trial? Clinical Trial Enrichment How to differentiate control from intervened? Trial Outcome Design

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 3 / 18

slide-7
SLIDE 7

Setting up a clinical trial – My work

Who is participating in the trial? Clinical Trial Enrichment How to differentiate control from intervened? Trial Outcome Design

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 3 / 18

slide-8
SLIDE 8

Setting up a clinical trial – My work

Who is participating in the trial? Clinical Trial Enrichment How to differentiate control from intervened? Trial Outcome Design trials aimed for Alzheimer’s Disease

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 3 / 18

slide-9
SLIDE 9

Alzheimer’s Disease

Destroys memory and cognition

  • Irreversible. Strongest risk factor is age

Diagnosis ← { Age, Family History, Cognitive/Neuropsych/Physical Exams, Brain Scans }

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 4 / 18

slide-10
SLIDE 10

Alzheimer’s Disease

Destroys memory and cognition

  • Irreversible. Strongest risk factor is age

Diagnosis ← { Age, Family History, Cognitive/Neuropsych/Physical Exams, Brain Scans }

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 4 / 18

slide-11
SLIDE 11

Alzheimer’s Disease

Destroys memory and cognition

  • Irreversible. Strongest risk factor is age

Diagnosis ← { Age, Family History, Cognitive/Neuropsych/Physical Exams, Brain Scans }

Normal/Healthy Dementia Mild Cognitive Impairment (MCI) Preclinical

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 4 / 18

slide-12
SLIDE 12

Landscape of AD Clinical Trials

Clinicaltrials.gov lists 485 recruiting studies 225 in US; 147 in Europe; 68 are in Phase III and IV

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 5 / 18

slide-13
SLIDE 13

Landscape of AD Clinical Trials

Clinicaltrials.gov lists 485 recruiting studies 225 in US; 147 in Europe; 68 are in Phase III and IV Very little success . . . more than 550 trials since 2002 (Cummings 2014)

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 5 / 18

slide-14
SLIDE 14

Landscape of AD Clinical Trials

Clinicaltrials.gov lists 485 recruiting studies 225 in US; 147 in Europe; 68 are in Phase III and IV Very little success . . . more than 550 trials since 2002 (Cummings 2014) AD diagnosis itself is messy → Early diagnosis is much harder → CN vs. MCI ≈ 70%

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 5 / 18

slide-15
SLIDE 15

Landscape of AD Clinical Trials

Clinicaltrials.gov lists 485 recruiting studies 225 in US; 147 in Europe; 68 are in Phase III and IV Very little success . . . more than 550 trials since 2002 (Cummings 2014) AD diagnosis itself is messy → Early diagnosis is much harder → CN vs. MCI ≈ 70% < 20% of MCIs convert to AD = ⇒ 8 out of 10 trial subjects are not-eligible!!

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 5 / 18

slide-16
SLIDE 16

. . . but there is light

Imaging to the rescue Cognitive decline follows atypical brain scans

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 6 / 18

slide-17
SLIDE 17

. . . but there is light

Imaging to the rescue Cognitive decline follows atypical brain scans

Alzheimer’s Machine Learning High Dimensional Imaging

Risk Factors

Disease

Markers Clinical

Enricher Design Optimal

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 6 / 18

slide-18
SLIDE 18

Population enrichment

Worsening Disease Healthy AD (Enrichment Criterion)

Enrichment Cut-Off

Included Discarded

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 7 / 18

slide-19
SLIDE 19

Population enrichment

Worsening Disease Healthy AD (Enrichment Criterion)

Enrichment Cut-Off

Included Discarded

Good enrichment criterion ⇐ ⇒ High correlation with disease Practical enrichment criterion ⇐ ⇒ High predictive power

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 7 / 18

slide-20
SLIDE 20

Designing a good enricher

Given some marker δ : Longitudinal change σ : Pooled Variance

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 8 / 18

slide-21
SLIDE 21

Designing a good enricher

Given some marker δ : Longitudinal change σ : Pooled Variance

Small σ Large δ + Optimal Enricher

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 8 / 18

slide-22
SLIDE 22

Designing a good enricher

Given some marker δ : Longitudinal change σ : Pooled Variance

Low-Variance Un-Biased + An Ensemble Neural Networks + Optimal Enricher

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 8 / 18

slide-23
SLIDE 23

Randomized deep networks for enrichment

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 9 / 18

slide-24
SLIDE 24

Randomized deep networks for enrichment

L Layered NN L Layered NN L Layered NN

Block 1

T networks

T outputs

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 9 / 18

slide-25
SLIDE 25

Randomized deep networks for enrichment

L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN

Block 1 Block 2

T networks T networks

T outputs T outputs

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 9 / 18

slide-26
SLIDE 26

Randomized deep networks for enrichment

L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN

Block 1 Block 2 Block B

T networks T networks T networks

T outputs T outputs T outputs

f1 f1 f2f4 f5 f3 f6 fB f2 fB

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 9 / 18

slide-27
SLIDE 27

Randomized deep networks for enrichment

Linear Combination

Final Output

L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN L Layered NN

Block 1 Block 2 Block B

T networks T networks T networks

T outputs T outputs T outputs

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 9 / 18

slide-28
SLIDE 28

Randomized deep network Markers – rDm

Training baseline rDm Inputs → MRI and PET Images Labels → AD – 0, healthy – 1

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 10 / 18

slide-29
SLIDE 29

Randomized deep network Markers – rDm

Training baseline rDm Inputs → MRI and PET Images Labels → AD – 0, healthy – 1 rDm at test time Predict on MCI

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 10 / 18

slide-30
SLIDE 30

Randomized deep network Markers – rDm

Training baseline rDm Inputs → MRI and PET Images Labels → AD – 0, healthy – 1 rDm at test time Predict on MCI Choose a cut-off t ∈ [0, 1] & filter out subjects with rDm prediction > t

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 10 / 18

slide-31
SLIDE 31

Predictive power of baseline rDm

Baseline rDm versus change (12 and 24 months) in outcomes Spearman correlation coefficient (and p-value) Marker 12m 24m MMSE 0.2123, p = 0.0008 0.3311, p = 0.0003 ADAS 0.2139, p = 0.0007 –0.5300, p < 10−4 MOCA 0.0568, p > 0.1 0.5952, p = 10−4 RAVLT 0.1285, p = 0.04 0.5702, p = 0.0008 PsyMEM 0.2811, p < 10−4 0.4207, p = 0.001 HippoVol 0.3262, p ≪ 10−4 0.4744, p ≪ 10−4 CDR-SB –0.3643, p ≪ 10−4 –0.5344, p ≪ 10−4 DXConv1 > 20, p ≪ 10−4 > 20, p ≪ 10−4

1ANOVA test results are reported since this variable is categorical (BRAIN Initiative Symposium 2017) Learning methods for enrichment 11 / 18

slide-32
SLIDE 32

Predictive power of baseline rDm

Baseline rDm versus change (12 and 24 months) in outcomes Spearman correlation coefficient (and p-value) Marker 12m 24m MMSE 0.2123, p = 0.0008 0.3311, p = 0.0003 ADAS 0.2139, p = 0.0007 –0.5300, p < 10−4 MOCA 0.0568, p > 0.1 0.5952, p = 10−4 RAVLT 0.1285, p = 0.04 0.5702, p = 0.0008 PsyMEM 0.2811, p < 10−4 0.4207, p = 0.001 HippoVol 0.3262, p ≪ 10−4 0.4744, p ≪ 10−4 CDR-SB –0.3643, p ≪ 10−4 –0.5344, p ≪ 10−4 DXConv1 > 20, p ≪ 10−4 > 20, p ≪ 10−4

1ANOVA test results are reported since this variable is categorical (BRAIN Initiative Symposium 2017) Learning methods for enrichment 11 / 18

slide-33
SLIDE 33

Predictive power of baseline rDm

Baseline rDm versus change (12 and 24 months) in outcomes Spearman correlation coefficient (and p-value) Marker 12m 24m MMSE 0.2123, p = 0.0008 0.3311, p = 0.0003 ADAS 0.2139, p = 0.0007 –0.5300, p < 10−4 MOCA 0.0568, p > 0.1 0.5952, p = 10−4 RAVLT 0.1285, p = 0.04 0.5702, p = 0.0008 PsyMEM 0.2811, p < 10−4 0.4207, p = 0.001 HippoVol 0.3262, p ≪ 10−4 0.4744, p ≪ 10−4 CDR-SB –0.3643, p ≪ 10−4 –0.5344, p ≪ 10−4 DXConv1 > 20, p ≪ 10−4 > 20, p ≪ 10−4

1ANOVA test results are reported since this variable is categorical (BRAIN Initiative Symposium 2017) Learning methods for enrichment 11 / 18

slide-34
SLIDE 34

Predictive power of baseline rDm

Baseline rDm versus change (12 and 24 months) in outcomes Spearman correlation coefficient (and p-value) Marker 12m 24m MMSE 0.2123, p = 0.0008 0.3311, p = 0.0003 ADAS 0.2139, p = 0.0007 –0.5300, p < 10−4 MOCA 0.0568, p > 0.1 0.5952, p = 10−4 RAVLT 0.1285, p = 0.04 0.5702, p = 0.0008 PsyMEM 0.2811, p < 10−4 0.4207, p = 0.001 HippoVol 0.3262, p ≪ 10−4 0.4744, p ≪ 10−4 CDR-SB –0.3643, p ≪ 10−4 –0.5344, p ≪ 10−4 DXConv1 > 20, p ≪ 10−4 > 20, p ≪ 10−4

1ANOVA test results are reported since this variable is categorical (BRAIN Initiative Symposium 2017) Learning methods for enrichment 11 / 18

slide-35
SLIDE 35

Predictive power of baseline rDm

Baseline rDm versus change (12 and 24 months) in outcomes Spearman correlation coefficient (and p-value) Marker 12m 24m MMSE 0.2123, p = 0.0008 0.3311, p = 0.0003 ADAS 0.2139, p = 0.0007 –0.5300, p < 10−4 MOCA 0.0568, p > 0.1 0.5952, p = 10−4 RAVLT 0.1285, p = 0.04 0.5702, p = 0.0008 PsyMEM 0.2811, p < 10−4 0.4207, p = 0.001 HippoVol 0.3262, p ≪ 10−4 0.4744, p ≪ 10−4 CDR-SB –0.3643, p ≪ 10−4 –0.5344, p ≪ 10−4 DXConv1 > 20, p ≪ 10−4 > 20, p ≪ 10−4 Very strong correlations across all markers

1ANOVA test results are reported since this variable is categorical (BRAIN Initiative Symposium 2017) Learning methods for enrichment 11 / 18

slide-36
SLIDE 36

Predictive power of baseline rDm

Mean longitudinal change in MMSE & CDR Important trial outcomes

0.2 0.4 0.6 0.8 1 −4 −3.5 −3 −2.5 −2 −1.5 −1 −0.5 rDAm Cut−Off MMSE Change 12m 24m 0.2 0.4 0.6 0.8 1 1 2 3 4 5 rDAm Cut−Off CDR Change 12m 24m (BRAIN Initiative Symposium 2017) Learning methods for enrichment 12 / 18

slide-37
SLIDE 37

Baseline rDm vs. alternate enrichers

Sample sizes per arm 80% power, 25% improvement from treatment

Sample Outcome measure enricher MMSE ADAS MOCA RAVLT PsyMEM HipVol CDR-SB DxConv HipVol 500 >2000 1005 1606 1009 >2000 389 420 FDG 384 1954 579 >2000 832 752 415 371 AV45 224 >2000 875 >2000 826 698 382 443 FAH 296 >2000 705 >2000 826 722 397 402 MKLm 2 228 874 827 896 487 877 295 284 rDm 200 775 449 591 420 543 281 230

2MKLm is the current state-of-the-art based on SVMs (BRAIN Initiative Symposium 2017) Learning methods for enrichment 13 / 18

slide-38
SLIDE 38

Baseline rDm vs. alternate enrichers

Sample sizes per arm 80% power, 25% improvement from treatment

Sample Outcome measure enricher MMSE ADAS MOCA RAVLT PsyMEM HipVol CDR-SB DxConv HipVol 500 >2000 1005 1606 1009 >2000 389 420 FDG 384 1954 579 >2000 832 752 415 371 AV45 224 >2000 875 >2000 826 698 382 443 FAH 296 >2000 705 >2000 826 722 397 402 MKLm 2 228 874 827 896 487 877 295 284 rDm 200 775 449 591 420 543 281 230

2MKLm is the current state-of-the-art based on SVMs (BRAIN Initiative Symposium 2017) Learning methods for enrichment 13 / 18

slide-39
SLIDE 39

Baseline rDm vs. alternate enrichers

Sample sizes per arm 80% power, 25% improvement from treatment

Sample Outcome measure enricher MMSE ADAS MOCA RAVLT PsyMEM HipVol CDR-SB DxConv HipVol 500 >2000 1005 1606 1009 >2000 389 420 FDG 384 1954 579 >2000 832 752 415 371 AV45 224 >2000 875 >2000 826 698 382 443 FAH 296 >2000 705 >2000 826 722 397 402 MKLm 3 228 874 827 896 487 877 295 284 rDm 200 775 449 591 420 543 281 230

rDm has smallest estimates across all outcomes

3MKLm is the current state-of-the-art based on SVMs (BRAIN Initiative Symposium 2017) Learning methods for enrichment 14 / 18

slide-40
SLIDE 40

Baseline rDm vs. alternate enrichers

Sample sizes per arm 80% power, 25% improvement from treatment

Sample Outcome measure enricher MMSE ADAS MOCA RAVLT PsyMEM HipVol CDR-SB DxConv HipVol 500 >2000 1005 1606 1009 >2000 389 420 FDG 384 1954 579 >2000 832 752 415 371 AV45 224 >2000 875 >2000 826 698 382 443 FAH 296 >2000 705 >2000 826 722 397 402 MKLm 3 228 874 827 896 487 877 295 284 rDm 200 775 449 591 420 543 281 230

Baseline rDm can detect weak treatment effects

3MKLm is the current state-of-the-art based on SVMs (BRAIN Initiative Symposium 2017) Learning methods for enrichment 14 / 18

slide-41
SLIDE 41

The proposed enricher – AD Spectrum

Normal/Healthy Dementia Mild Cognitive Impairment (MCI) Preclinical

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 15 / 18

slide-42
SLIDE 42

The proposed enricher – AD Spectrum

Normal/Healthy Dementia Mild Cognitive Impairment (MCI) Preclinical Label 1 Label 0 Training Regimes

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 15 / 18

slide-43
SLIDE 43

The proposed enricher – AD Spectrum

Normal/Healthy Dementia Mild Cognitive Impairment (MCI) Preclinical Label 1 Label 0 Training Regimes Testing Regime

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 15 / 18

slide-44
SLIDE 44

The proposed enricher – AD Spectrum

Normal/Healthy Dementia Mild Cognitive Impairment (MCI) Preclinical Label 1 Label 0 Training Regimes Testing Regime

Two Issues

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 15 / 18

slide-45
SLIDE 45

The proposed enricher – AD Spectrum

Normal/Healthy Dementia Mild Cognitive Impairment (MCI) Preclinical Label 1 Label 0 Training Regimes Testing Regime

Two Issues Disease spectrum is continuous → Labels somewhat artificial – Supervised models are sensitive

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 15 / 18

slide-46
SLIDE 46

The proposed enricher – AD Spectrum

Normal/Healthy Dementia Mild Cognitive Impairment (MCI) Preclinical Label 1 Label 0 Training Regimes Testing Regime

Two Issues Disease spectrum is continuous → Labels somewhat artificial – Supervised models are sensitive Bio-markers interact differently in preclinical vs. AD

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 15 / 18

slide-47
SLIDE 47

The proposed enricher – AD Spectrum

Normal/Healthy Dementia Mild Cognitive Impairment (MCI) Preclinical Label 1 Label 0 Training Regimes Testing Regime

Two Issues Disease spectrum is continuous → Labels somewhat artificial – Supervised models are sensitive Bio-markers interact differently in preclinical vs. AD Can we instead select subjects without information transfer from AD stage?

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 15 / 18

slide-48
SLIDE 48

An alternate View – Sampling

Select atypical subjects The more unique a subject is → . . . the more information they contribute to trial Some typical points also needed

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 16 / 18

slide-49
SLIDE 49

An alternate View – Sampling

Select atypical subjects The more unique a subject is → . . . the more information they contribute to trial Some typical points also needed

AD Imaging Features AD Clinical/Neuropsych Scores

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 16 / 18

slide-50
SLIDE 50

An alternate View – Sampling

Select atypical subjects The more unique a subject is → . . . the more information they contribute to trial Some typical points also needed

AD Imaging Features AD Clinical/Neuropsych Scores

Very Rich Block (Hierarchical) Structure

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 16 / 18

slide-51
SLIDE 51

Multiresolution Matrix Factorization

Chutney Salsa Ketchup Salad Strawberry Shortcake Bannock Chapati Pita Margarine Saute Limpa

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 17 / 18

slide-52
SLIDE 52

Multiresolution Matrix Factorization

Chutney Salsa Ketchup Salad Strawberry Shortcake Bannock Chapati Pita Margarine Saute Limpa

ℓ = 1 Chapati Chutney Bannock Limpa Pita (BRAIN Initiative Symposium 2017) Learning methods for enrichment 17 / 18

slide-53
SLIDE 53

Multiresolution Matrix Factorization

Chutney Salsa Ketchup Salad Strawberry Shortcake Bannock Chapati Pita Margarine Saute Limpa

ℓ = 1 ℓ = 2 Chapati Salad Chutney Bannock Saute Limpa Pita Salsa (BRAIN Initiative Symposium 2017) Learning methods for enrichment 17 / 18

slide-54
SLIDE 54

Multiresolution Matrix Factorization

Chutney Salsa Ketchup Salad Strawberry Shortcake Bannock Chapati Pita Margarine Saute Limpa

ℓ = 1 ℓ = 2 ℓ = 3 ℓ = 4 ℓ = 5 Chapati Salad Ketchup Chutney Bannock Saute Limpa Pita Salsa Margarine Strawberry Shortcake (BRAIN Initiative Symposium 2017) Learning methods for enrichment 17 / 18

slide-55
SLIDE 55

Multiresolution Matrix Factorization

Chutney Salsa Ketchup Salad Strawberry Shortcake Bannock Chapati Pita Margarine Saute Limpa

Chapati Salad Ketchup Chutney Bannock Saute Limpa Pita Salsa Margarine Strawberry Shortcake

More Outlier-ish

Maximally different from the rest

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 17 / 18

slide-56
SLIDE 56

Thank you . . . Questions?

I., V. Singh, O. C. Okonkwo, S. C. Johnson, A predictive multi-modal imaging marker for designing efficient and robust AD clinical trials, Clinical Trials on Alzheimer’s Disease (CTAD), 2014 I., V. Singh, S. C. Johnson, Randomized deep learning methods for clinical trial enrichment and design in Alzheimer’s disease, Deep Learning for Medical Image Analysis (1st Edition) ISBN: 9780128104088; Chapter 15 I., V. Singh, O. C. Okonkwo, R. J. Chappell, N. M. Dowling, S. C. Johnson, Imaging based enrichment criteria using deep learning algorithms for efficient clinical trials in MCI, Alzheimer’s and Dementia, 2015 I., R. Kondor, S. C. Johnson, V. Singh, The Incremental Multiresolution Matrix Factorization Algorithm, Computer Vision and Pattern Recognition (CVPR), 2017 http://pages.cs.wisc.edu/~vamsi/publications.html Acknowledgements: Vikas Singh, Sterling Johnson, Chris Hinrichs, Risi Kondor, Barbara Bendlin, Ozioma Okonkwa NIH AG040396, NSF CAREER 1252725, NSF CCF 1320755, UW ADRC AG033514

(BRAIN Initiative Symposium 2017) Learning methods for enrichment 18 / 18