fMRI markers for Major
Depressive Disorder (MDD)
- diagnosis and prognosis -
Dr. H.G. (Eric) Ruhé, MD PhD
Program for Mood and Anxiety Disorders
University Medical Center Groningen
H.G.Ruhe@UMCG.NL
Co-authors: M.M. Rive, H. Geugies, L. Schmaal, M.J. van Tol, E.M. Opmeer, A.F. Marquand, N.A. van der Wee, B. Penninx, A.H. Schene and D.J. Veltman
Clinical Problems in MDD
Prediction of long-term course
Distinction MDD and Bipolar Disorder
Clinical Problems in MDD
Prediction of long-term course
– NEMESIS-study
n= 7076
n= 273 MDD cases
– 50% recovery in 3 mth
– 20% chronic MDD
Risk factors: severity
& dysthymia
Distinction MDD and Bipolar Disorder
Indication of best treatment
duration (months) 30 20 10 0 p ro p o rt io n s till in e p is o d e 1,2 1,0 ,8 ,6 ,4 ,2 0,0 censored
Clinical Problems in MDD
Prediction of long-term course
Distinction MDD and Bipolar Disorder
– Depressive episode as 1
st
manifestation
– Recurrences
– Similar symptoms
– Conversion rate 10% in 5 years
Risk factors: Age, Family-history, psychotic
features, hypomanic Syx
Clinical Problems in MDD
Prediction of long-term course
Distinction MDD and Bipolar Disorder
Indication of best treatment
– Psychotherapy vs Pharmacotherpy
– Which therapy / antidepressant
– One-size-fits all / Trial & change
Treatment Resistant Depression (TRD) or
How can we individualize
predictions?
Group vs Individual level
Possibilities of new approaches especially
with (big) neuroimaging data
Big data analysis
Artificial intelligence
– algorithms and techniques to automatically
extract information from big data
(Support Vector) Machine Learning
(SVM) techniques
Big data analysis
Artificial intelligence
– algorithms and techniques to automatically
extract information from big data
(Support Vector) Machine Learning
(SVM) techniques
– Supervised OR unsupervised
Big data analysis
Artificial intelligence
– algorithms and techniques to automatically
extract information from big data
(Support Vector) Machine Learning
(SVM) techniques
– Supervised OR unsupervised
– Training -> Testing
Validation
– Cross-validation
– New datasets
Big data analysis
Artificial intelligence
– algorithms and techniques to automatically
extract information from big data
(Support Vector) Machine Learning
(SVM) techniques
– Supervised OR unsupervised
– Training -> Testing
Validation
– Cross-validation
– New datasets
Big data analysis
Artificial intelligence
– algorithms and techniques to automatically
extract information from big data
(Support Vector) Machine Learning
(SVM) techniques
– Supervised OR unsupervised
– Training -> Testing
Validation
– Cross-validation
– New datasets
Information obtained
- diagnostic testing -
Sensitivity (%)
– a/a+c
Specificity (%)
– d/b+d
Accuracy (%)
– a+d/a+b+c+d
Outcome
Yes
No
Total
Test
Yes
..
..
..
a+b
a b
D
No
..
..
c d
..
... ..
c+d
Total
..
..
…
a+c
b+d
a+b+c+d
Prediction of long-term
course/chronicity
Clinical predictors:
Symptom severity/duration, age, age of onset, comorbidity,
childhood adversity, personality
(Karsten et al 2013, Penninx et al.
2011)
Neuroimaging predictors
(group-level)
:
Reduced hippocampus and ACC volume
(e.g. Woudstra et al. under
review)
Abnormal activation in mPFC regions during emotional
processing
(Farb et al. 2011; Siegle et al. 2012)
Abnormal DLPFC recruitment during visuospatial planning
NESDA 2 year longitudinal course prediction
fMRI
- Design -
MDD-Patients (n= 188; 6 months Dx)
1.
A rapid remission trajectory (REM;
N=59)
2.
Gradual decline of symptoms (DEC;
N=36)
3.
Chronic trajectory (CHR; N=23)
NESDA 2 year longitudinal course prediction
fMRI
- Design -
MDD-Patients (n= 188; 6 months Dx)
1.
A rapid remission trajectory (REM;
N=59)
2.
Gradual decline of symptoms (DEC;
N=36)
3.
Chronic trajectory (CHR; N=23)
NESDA 2 year longitudinal course prediction
fMRI
- Design -
Schmaal et al. Submitted
Neuroimaging
T1 sMRIfMRI: Tower of London, Emotional Facial Expressions
✔ ✔
Clinical information
Depression severity (IDS) Anxiety severity (BAI)
Duration of symptoms prior to baseline (Life Chart) Age of onset
Personality: neuroticism, extraversion, conscientiousness (NEO-FFI) ✔ ✔ ✔ ✔ ✔
Environmental factors
Childhood maltreatment ✔ACCURACIES (Sens/Spec): Modality MDD-CHR (N=23) versus MDD-REM (N=59) MDD-CHR (N=23) versus MDD-DEC (N=36) MDD-DEC (N=36) versus MDD-REM (N=59) Faces Task Angry>baseline Fear>baseline Happy>baseline Sad>baseline Neutral>baseline Overall emotion>baseline 64%* (67/62) 62%* (67/56) 64%* (73/54) 58% (60/56) 53% (47/59) 73%** (80/67) 54% (53/55) 59% (60/58) 69%* (67/71) 49% (47/52) 67%* (67/68) 59% (53/65) 48% (42/54) 40% (35/45) 53% (55/51) 45% (39/51) 37% (32/41) 50% (48/51) Tower of London 51% (53/50) 38% (37/46) 48% (46/50) Grey matter images 43% (35/52)
53% (48/58) 43% (33/53)
Clinical characteristics 69%* (70/68) 61% (61/61) 61% (69/53)
Faces contrast images and clinical characteristics combined
65%* (52/78)
52% (35/69) 54% (14/93)
All modalities combined c 62%* (74/49) 61% (65/57) 44% (43/44)
ACCURACIES (Sens/Spec): Modality MDD-CHR (N=23) versus MDD-REM (N=59) MDD-CHR (N=23) versus MDD-DEC (N=36) MDD-DEC (N=36) versus MDD-REM (N=59) Faces Task Angry>baseline Fear>baseline Happy>baseline Sad>baseline Neutral>baseline Overall emotion>baseline 64%* (67/62) 62%* (67/56) 64%* (73/54) 58% (60/56) 53% (47/59) 73%** (80/67) 54% (53/55) 59% (60/58) 69%* (67/71) 49% (47/52) 67%* (67/68) 59% (53/65) 48% (42/54) 40% (35/45) 53% (55/51) 45% (39/51) 37% (32/41) 50% (48/51) Tower of London 51% (53/50) 38% (37/46) 48% (46/50) Grey matter images 43% (35/52)
53% (48/58) 43% (33/53)
Clinical characteristics 69%* (70/68) 61% (61/61) 61% (69/53)
Faces contrast images and clinical characteristics combined
65%* (52/78)
52% (35/69) 54% (14/93)
All modalities combined c 62%* (74/49) 61% (65/57) 44% (43/44)
ACCURACIES (Sens/Spec): Modality MDD-CHR (N=23) versus MDD-REM (N=59) MDD-CHR (N=23) versus MDD-DEC (N=36) MDD-DEC (N=36) versus MDD-REM (N=59) Faces Task Angry>baseline Fear>baseline Happy>baseline Sad>baseline Neutral>baseline Overall emotion>baseline 64%* (67/62) 62%* (67/56) 64%* (73/54) 58% (60/56) 53% (47/59) 73%** (80/67) 54% (53/55) 59% (60/58) 69%* (67/71) 49% (47/52) 67%* (67/68) 59% (53/65) 48% (42/54) 40% (35/45) 53% (55/51) 45% (39/51) 37% (32/41) 50% (48/51) Tower of London 51% (53/50) 38% (37/46) 48% (46/50) Grey matter images 43% (35/52)
53% (48/58) 43% (33/53)
Clinical characteristics 69%* (70/68) 61% (61/61) 61% (69/53)
Faces contrast images and clinical characteristics combined
65%* (52/78)
52% (35/69) 54% (14/93)
All modalities combined c 62%* (74/49) 61% (65/57) 44% (43/44)
Conclusions:
- Clinical < neuroimaging precdictors
- Prediction by emotion recognition task
- No prediction by visuospatial planning task and structural
neuroimaging
- A distributed network including many regions having low
individual but strong aggregate prognostic value
2 Sites: Pittsburg / Münster
2x3x29 depressed BD / depressed
MDD / HC
Mostly medicated patients
3T sMRI
Univariate analyses & SVM/GPC
– Mask for regions emotion regulation
Distinction MDD and BD
- Redlich et al. -
Redlich et al. JAMA Psychiatry 2014 Online
Distinction MDD and BD
2 Sites: Pittsburg / Münster
2x3x29 depressed BD / depressed
MDD / HC
Mostly medicated patients
3T sMRI
Univariate analyses & SVM/GPC
– Mask for regions emotion regulation
Distinction MDD and BD
- Redlich et al. -
Redlich et al. JAMA Psychiatry 2014 Online
Distinction MDD and BD
2 Sites: Pittsburg / Münster
2x3x29 depressed BD / depressed
MDD / HC
Mostly medicated patients
3T sMRI
Univariate analyses & SVM/GPC
– Mask for regions emotion regulation
Distinction MDD and BD
- Redlich et al. -
Redlich et al. JAMA Psychiatry 2014 Online
Distinction MDD and BD
2 Sites: Pittsburg / Münster
2x3x29 depressed BD / depressed
MDD / HC
Mostly medicated patients
3T sMRI
Univariate analyses & SVM/GPC
– Mask for regions emotion regulation
Distinction MDD and BD
- Redlich et al. -
Redlich et al. JAMA Psychiatry 2014 Online
Distinction MDD and BD
Distinction MDD and BD
- design DIADE -
Drug-free patients: SCID +ve; ≥2 episodes; ≥5 yrs
of illness; age of onset ≤40yrs
– 32 BD (10 depressed (BDd), 22 remitted (BDr))
– 32 MDD (10 MDDd and 22 MDDr)
– matched for age, gender, education and depression severity
32 healthy controls (HC)
Resting state (RS) fMRI (7 min)
Independent components analysis (ICA)
11 RSN -> a priori selection of 3 RSN
– Default Mode Network (DMN), FrontoParietal Network (FPN),
Salience Network (SN)
Distinction MDD and BD
- results -
ACCURACIES (Sens/Spec): Modality MDDd (N=10) versus BDd (N=10) MDDr (N=22) versus BDr (N=22) HC (N=32) versus MDD (N=32) HC (N=32) versus BD (N=32) DMN 85%** (90/80) 52% (46/59) 65%* (70/59) 74%** (76/72) FPN Left Right 45% (50/40) 40% (40/40) 50% (59/41) 55% (55/55) NA NA SN 30% (30/30) 48% (46/50) NA NARive et al. Submitted
* p<0.05; **p<0.01
Distinction MDD and BD
- results -
ACCURACIES (Sens/Spec): Modality MDDd (N=10) versus BDd (N=10) MDDr (N=22) versus BDr (N=22) HC (N=32) versus MDD (N=32) HC (N=32) versus BD (N=32) DMN 85%** (90/80) 52% (46/59) 65%* (70/59) 74%** (76/72) FPN Left Right 45% (50/40) 40% (40/40) 50% (59/41) 55% (55/55) NA NA SN 30% (30/30) 48% (46/50) NA NARive et al. Submitted
* p<0.05; **p<0.01
Distinction MDD and BD
- results -
ACCURACIES (Sens/Spec): Modality MDDd (N=10) versus BDd (N=10) MDDr (N=22) versus BDr (N=22) HC (N=32) versus MDD (N=32) HC (N=32) versus BD (N=32) DMN 85%** (90/80) 52% (46/59) 65%* (70/59) 74%** (76/72) FPN Left Right 45% (50/40) 40% (40/40) 50% (59/41) 55% (55/55) NA NA SN 30% (30/30) 48% (46/50) NA NARive et al. Submitted
* p<0.05; **p<0.01
Conclusions:
- Distinction MDD vs BD (vs HC) based on DMN / sMRI
- Especially in depressed state
- Drug free or drug use in subjects might matter!!
- A distributed network including many regions having low
individual but strong aggregate diagnostic value
Best Treatment?
Prognosis of treatment outcome
Responders vs non-responders
Best Treatment?
Prognosis of treatment outcome
Responders vs non-responders
Best Treatment?
Prognosis of treatment outcome
Responders vs non-responders
Pizzagalli. Neuropsychopharmacol. 2011
Conclusions:
- Group-level responders
vs
non-responders
- rACC hyperactivation associated with better treatment
outcome
- High effect-sizes
- Important role of rACC with
- Emotion regulation
- DMN (intermediating between Task positive and task
negative activities)
NESDA Prediction of TRD
Baseline scan in 112 current MDD
2yr follow-up
– TRD (≥2 antidepressants; n= 17)
– Responders (<2 antidepressants; n=32;
matched age/sex/education)
No differences in depression severity (IDS) anxiety
severity (BAI) and illness duration (in months)
between TRD and nonTRD group
RS fMRI -> ICA-analysis
Group-analyses
HC vs TRD
no difference
HC vs nonTRD no difference
TRD vs nonTRD
↓
connectivity of right insula in the
salience network (
P
FWE
= 0.006)
Second level group differences
HC
TRD
nonTRD
Contrast estimates and 90% C.I.
HC vs TRD
no difference
HC vs nonTRD no difference
TRD vs nonTRD
↓
connectivity of right insula in the
salience network (
P
FWE
= 0.006)
Second level group differences
HC
TRD
nonTRD
Contrast estimates and 90% C.I.
Geugies et al. In preparation
Conclusions:
- Right insula
hypo
connectivity in SN associated with TRD
- Group-level TRD
vs
non-TRD
- Important role of Insula
- DMN (activity preceeding switch Task negative to Task
positive)
Prognosis of treatment outcome
Individualized approaches
Psychotherapy & Pharmacotherapy
Prognosis of treatment outcome
Individualized approaches
Psychotherapy & Pharmacotherapy
sMRI > fMRI ?
Prognosis of treatment outcome
Individualized approaches
Psychotherapy & Pharmacotherapy
sMRI > fMRI ?
Very small samples -> new studies
Individual prediction in TRD
- Outcome of ECT -
N=45 TRD-patients (25 final remitters)
RS fMRI -> ICA
Individual prediction in TRD
- Outcome of ECT -
N=45 TRD-patients (25 final remitters)
RS fMRI -> ICA
Individual prediction in TRD
- Outcome of ECT -
N=45 TRD-patients (25 final remitters)
RS fMRI -> ICA
Accuracies:
– 84% (84 sens /85 spec)
– 77% (80 sens /75 spec)
– sMRI 61% (n.s.)
Individual prediction in TRD
- Outcome of ECT -
N=45 TRD-patients (25 final remitters)
RS fMRI -> ICA
Accuracies:
– 84% (84 sens /85 spec)
– 77% (80 sens /75 spec)
– sMRI 61% (n.s.)
Van Waarde et al. Mol Psychiatry 2014
Conclusions:
- Connectivity in 2 cognitive/affective networks are associated
with remission by ECT
- Prediction fMRI > sMRI
- Role of DLPFC
- Emotion regulation / Dorsal Nexus
- A distributed network including many regions having low
individual but strong aggregate diagnostic value
Summary
Advantages of machine learning
vs
univariate approaches
However
– Small samples
– Need for replications in independent
samples and if so -> worse prediction ???
– Accuracies may/must improve
Combinations are not always better
Early change in neural processing of
fear predicts antidepressant response
Acknowledgement
Maaike Rive
Hanneke Geugies
Marie Jose van Tol
Lianne Schmaal
Esther Opmeer
Andre Marquand
Nic van der Wee
Brenda Penninx
Aart Schene
Dick Veltman