Using machine learning approaches to identify imaging markers that - - PowerPoint PPT Presentation

▶
using machine learning approaches to identify imaging
SMART_READER_LITE
LIVE PREVIEW

Using machine learning approaches to identify imaging markers that - - PowerPoint PPT Presentation

Mar 05, 2023 238 likes •482 views

Using machine learning approaches to identify imaging markers that predict antidepressant response Amit Etkin, MD, PhD Associate Professor Department of Psychiatry and Behavioral Sciences Stanford Neurosciences Institute Stanford University

slide-1
SLIDE 1

Using machine learning approaches to identify imaging markers that predict antidepressant response

Amit Etkin, MD, PhD

Associate Professor Department of Psychiatry and Behavioral Sciences Stanford Neurosciences Institute Stanford University Investigator, Sierra-Pacific MIRECC Palo Alto VA etkinlab.stanford.edu

slide-2
SLIDE 2

Disclosures

Equity (SAB): Akili Interactive, Mindstrong Health

slide-3
SLIDE 3

The Guardian 2/21/18 The Guardian 2/26/08

One-size-fits-all? Can the drug define the patient?

slide-4
SLIDE 4

10 20 30 40 50

4 8 12 16 20 2 4 6 8

For whom does an antidepressant work?

EMBARC (PIs: Trivedi, Weissman, McGrath, Parsey): 309 depressed patients -> sertraline vs placebo remission rate (%)

Depression severity (HAMD17)

weeks PBO SER NNT=8.4 All patients taken together:

Etkin, Fonzo, Zhang, under review

d=0.27

slide-5
SLIDE 5

Emotional conflict task

Task: identify facial expression, ignore word Emotional conflict is biologically salient

Etkin, Neuron 2006

Implicit regulation: across-trial adjustments in behavior (RT) Subjects unaware of pattern

+ ... ... +

1s 3-5s 1s 3-5s

slide-6
SLIDE 6

Emotional conflict regulation circuit

reactivity regulation dACC/ dmPFC LPFC insula amygdala vACC/ vmPFC

Etkin, Neuron 2006 Egner, Cer Cort 2008 Etkin, Am J Psych 2010 Etkin, TICS, 2011 Etkin, Am J Psych, 2011

slide-7
SLIDE 7

For whom does an antidepressant work?

Emotion conflict reactivity or regulation as a moderator?

  • Intent-to-treat framework
  • Whole-brain voxelwise FDR correction

Etkin, Fonzo, Zhang, under review

  • Conflict

regulation but not reactivity

  • Unrelated to

baseline clinical severity

slide-8
SLIDE 8

2 4 6 8 4 8 12 16 20 2 4 6 8

For whom does an antidepressant work?

Etkin, Fonzo, Zhang, under review

Remission: Symptoms:

week week below median (better regulation) above median (worse regulation)

Example result:

Depression severity (HAMD17) below median (better regulation) above median (worse regulation) remission rate (%)

NNT=3.4 d=0.76 PBO SER

10 20 30 40 50

slide-9
SLIDE 9

For whom does an antidepressant work?

Etkin, Fonzo, Zhang, under review

Machine learning to define a response phenotype:

  • Whole-brain coverage, impute HAMD17
  • Sparse Bayesian Learning, 10-fold cross-validation
  • 15
  • 5

5 15 25

  • 15

15

  • 15

15

  • 15

15 True HAMD17 change (pre-post)

SER

Predicted HAMD17 change (pre-post)

r=0.49 p<3x10-8

True HAMD17 change (pre-post) Predicted HAMD17 change (pre-post)

r=-0.06 p=0.48

SER train test SER PBO train test

slide-10
SLIDE 10

Predicting outcome with EEG

Widge, AJP, in press

slide-11
SLIDE 11

Challenges of EEG machine learning

volume conduction dimensionality reduction

  • ptimization
slide-12
SLIDE 12

A novel machine learning approach for outcome prediction

EEG

S1 S2 SN

(・)2 (・)2

∫t ∫t

C1 C2 F1 F2 Pz

Spatial Filtering Band Power Feature Extraction Linear Regression Joint Parameter Estimation to Minimize Prediction Error Treatment Outcome

Constructed as a convex optimization problem

Wu et al., in prep

slide-13
SLIDE 13

Spatial filtering

X 1 X 0.7

Spatial filter optimization is the procedure of estimating the unknown weight coefficient of each electrode spatial filter

slide-14
SLIDE 14

Individual-level outcome prediction with rsEEG

SER SER train test SER PBO train test

Wu et al., in prep

slide-15
SLIDE 15

Comparison with other models

SER SER train test

alpha power theta cordance

Wu et al., in prep

PCA ICA

slide-16
SLIDE 16

Spatial patterns contributing to the prediction

SER SER train test

+

  • +
  • scalp spatial pattern

weights cortical spatial pattern weights

latent signal with most positive regression weight latent signal with most negative regression weight

Wu et al., in prep

slide-17
SLIDE 17

Predicting across sites and equipment

Wu et al., in prep

slide-18
SLIDE 18

Do the task fMRI and rsEEG models identify a convergent treatment-responsive phenotype?

Wu et al., in prep

Correlate predictions in an independent depressed sample:

slide-19
SLIDE 19

Is the phenotype associated with differences in neural responsivity?

latent signal with most positive regression weight latent signal with most negative regression weight

Wu et al., in prep

Wu, Keller, HBM, 2018

slide-20
SLIDE 20

Is the phenotype associated with differences in neural responsivity?

latent signal with most positive regression weight latent signal with most negative regression weight

Wu et al., in prep

ቐ ቐ

aMFG pMFG M1 aMFG pMFG M1 V1

left right

✽ ✽ ✽ 0-200ms 200-400ms 400-600ms 0-200ms 200-400ms 400-600ms 0-200ms 200-400ms 400-600ms 0-200ms 200-400ms 400-600ms

𝜄 𝛽 𝛾 𝛿

aMFG pMFG M1 V1

slide-21
SLIDE 21

Study stratification using our signatures

% with 5.5 point change (SER-PBO) 10 20 30 40 50 60

Resting EEG

Overall rate

10 20 30 40 50 60

Overall rate

% with 5.5 point change (SER-PBO)

Emotional conflict fMRI

slide-22
SLIDE 22

Acknowledgements

Funding: NIMH, VA, Dana Foundation, Cohen Veterans Bioscience,

Stanford Neurosciences Institute, private donors

Collaborators Madhukar Trivedi & EMBARC Etkin lab Wei Wu Greg Fonzo Yu Zhang etkinlab.stanford.edu