<p>Machine learning techniques in a structural and functional MRI diagnostic approach in schizophrenia: a systematic review</p>

(1)

R E V I E W

Machine learning techniques in a structural and

functional MRI diagnostic approach in

schizophrenia: a systematic review

This article was published in the following Dove Press journal: Neuropsychiatric Disease and Treatment

Renato de Filippis1,*

Elvira Anna Carbone1,*

Raffaele Gaetano1

Antonella Bruni1

Valentina Pugliese1

Cristina Segura-Garcia2

Pasquale De Fazio1

1_{Department of Health Sciences,} University Magna Graecia of Catanzaro, Catanzaro 88100, Italy;2_{Department of} Medical and Surgical Sciences, University Magna Graecia of Catanzaro, Catanzaro 88100, Italy

*These authors contributed equally to this work

Background: Diagnosis of schizophrenia (SCZ) is made exclusively clinically, since spe-ciﬁc biomarkers that can predict the disease accurately remain unknown. Machine learning (ML) represents a promising approach that could support clinicians in the diagnosis of mental disorders.

Objectives: A systematic review, according to the PRISMA statement, was conducted to evaluate its accuracy to distinguish SCZ patients from healthy controls.

Methods: We systematically searched PubMed, Embase, MEDLINE, PsychINFO and the Cochrane Library through December 2018 using generic terms for ML techniques and SCZ without language or time restriction. Thirty-ﬁve studies were included in this review: eight of them used structural neuroimaging, twenty-six used functional neuroimaging and one both, with a minimum accuracy >60% (most of them 75–90%). Sensitivity, Speciﬁcity and accuracy were extracted from each publication or obtained directly from authors.

Results:Support vector machine, the most frequent technique, if associated with other ML techniques achieved accuracy close to 100%. The prefrontal and temporal cortices appeared to be the most useful brain regions for the diagnosis of SCZ. ML analysis can efﬁciently detect signiﬁcantly altered brain connectivity in patients with SCZ (eg, default mode net-work, visual netnet-work, sensorimotor netnet-work, frontoparietal network and salience network).

Conclusion: The greater accuracy demonstrated by these predictive models and the new models resulting from the integration of multiple ML techniques will be increasingly decisive for early diagnosis and evaluation of the treatment response and to establish the prognosis of patients with SCZ. To achieve a real beneﬁt for patients, the future challenge is to reach an accurate diagnosis not only through clinical evaluation but also with the aid of ML algorithms.

Keywords: machine learning, schizophrenia, support vector machine, resting-state fMRI, sMRI, multivariate pattern analysis

Introduction

Schizophrenia (SCZ) is a major psychiatric disorder characterized by delusions and hallucinations with a loss of contact with reality (so-called positive symptoms),ﬂat affect, anhedonia, loss of motivation (avolition), poor speech (alogia), social with-drawal (so-called negative symptoms) and cognitive impairment, which have

a strong impact on the patient and society.1 The average incidence is 15.2 per

100,000 people, the lifetime prevalence is 0.40% and the pooled male–female rate

ratio is 1.4.2A recent review estimated that approximately 1 in 200 individuals will

Correspondence: Pasquale De Fazio Psychiatry Unit, Department of Health Sciences, University Magna Græcia of Catanzaro, Viale Europa, Italy Via Broussard 13, Catanzaro 88100, Italy Tel/Fax +39 096 171 2393

Email [email protected]

Neuropsychiatric Disease and Treatment

Dove

press

open access to scientiﬁc and medical research

Open Access Full Text Article

Neuropsychiatric Disease and Treatment downloaded from https://www.dovepress.com/ by 118.70.13.36 on 25-Aug-2020

(2)

be diagnosed with SCZ at some point during their lifetime,

with variations based on the geographic region.3Diagnosis

is based on the Diagnostic and Statistical Manual of

Mental Disorders 5 (DSM-5)4 or on the ICD5 and is

exclusively clinical, since speciﬁc biomarkers that can

predict the illness accurately remain unknown.6 Pattern

recognition methods applied to neuroimaging data repre-sent a new and promising approach that could support

clinicians in the diagnosis of mental disorders.7Magnetic

resonance imaging (MRI) is today one of the most used techniques for the study of the neurobiological bases of SCZ. Computerized methods are necessary to analyze large amounts of imaging data with great precision. The most relevant information is extracted from the images,

and then a model classiﬁcation method is able to process

this information in order to determine the probability of

disease.8The new approach is exempliﬁed by the

applica-tion of machine learning (ML) techniques such as support vector machines (SVM), a high-dimensional, pattern recognition, supervised learning algorithm9or multivariate pattern analysis (MVPA) or the random forest (RF). These techniques are most useful when analyzing highly com-plex datasets to enable the discovery of relationships that are not evident from simple statistics and thus to make accurate predictions based on these statistics.10

This research seeks to evaluate the evidence about the role of ML techniques in making diagnostic discrimination in SCZ during the last 30 years. To the best of our knowl-edge, there are no systematic reviews investigating these features. We aimed to provide a comprehensive systematic

review according to the PRISMA statement11 of

different ML techniques used to distinguish SCZ patients from healthy controls (HC). It was not our goal to catalog all methods of ML, to report on applications of ML in psychiatry nor to explain or illustrate any particular method in detail.

Materials and methods

Search strategy

We searched PubMed, Embase, MEDLINE, PsychINFO and the Cochrane Library for articles published up to December 31, 2018, without language and time restriction, by using the following keywords: (“Big data”OR“Artiﬁcial

Intelligence” OR“Machine Learning” OR“Gaussian

pro-cess” OR “Cross-validation” OR “Cross validation” OR

“Crossvalidation” OR “Regularized logistic” OR “Linear

discriminant analysis” OR “LDA” OR “Random forest”

OR“Naïve Bayes”OR“Least Absolute selection shrinkage

operator”OR“elastic net”OR“LASSO”OR“RVM”OR

“relevance vector machine” OR “pattern recognition” OR

“Computational Intelligence” OR “Machine Intelligence”

OR “Knowledge Representation” OR “Knowledge

Representations” OR “support vector” OR “SVM” OR

“Pattern classiﬁcation” OR “Deep learning”) AND

(“Schizophrenia”). Two authors reviewed all the selected studies independently. The reference lists were screened to

ﬁnd additional data. The eligible publications have been

included and cited in this review. This systematic review

was developed according to the PRISMA statement.11

Assessment of study quality

The Jadad rating system12was applied to check the

meth-odological quality of studies in this systematic review. Jadad’s procedure enables qualiﬁcation of a study accord-ing to the clarity in describaccord-ing the randomization, the double-blinding procedure and the withdrawal and dropout reports. Score ranges from 0 to 5. For this systematic review, the inclusion criteria was a Jadad score >3.

Selection criteria

We included studies using ML techniques with schizophre-nic patients diagnosed according to the DSM-IV, DSM IV-TR, DSM-5 or ICD-10 criteria. These patients could have

chronic SCZ or be diagnosed with aﬁrst episode of

schizo-phrenia (FES) and could be drug-naïve or under antipsycho-tic treatment. We excluded studies with patients who exhibited general medical, neurological or psychiatric comorbidity, substance abuse or alcohol dependence, trau-matic brain injuries with loss of consciousness or unclear or

unveriﬁed diagnoses according to the DSM or ICD criteria.

Studies without control groups were also excluded.

Data collection, extraction and statistical

analysis

Two researchers (EAC and RdF) independently screened the titles and abstracts of the identiﬁed articles and read the full texts of articles that met the eligibility criteria,

supervised by RG, who made theﬁnal decision in cases of

disagreement. Article data included year of

publication, ML model and algorithm (eg, SVM, MVPA, RF), sample size, diagnoses assessed in the study and statistical data (ie, accuracy, sensitivity, speciﬁcity, area under the ROC curve [AUC]).

(3)

Results

Initially, 2,386 items were identified, of which 2,215 articles were excluded because they did not fulfill inclusion criteria. The abstracts of the remaining 171 articles were reviewed. In all, 60/171 articles were deleted because they were edi-torials, letters to editors, reviews, meta-analyses, case reports or different interventions. Then, 76 manuscripts of 111 papers were deleted because they did not fulfill the criteria for inclusion, while the remaining 35 were included. Of the latter, 8 used structural neuroimaging and 26 used functional neuroimaging and 1 both (Figure 1).

Eight studies used structural magnetic resonance ima-ging (sMRI) in order to distinguish patients with SCZ or adolescent-onset schizophrenia (AOS) or FES from patients with other psychiatric disorders (eg, bipolar disorder (BD),

major depressive disorder (MDD)) or HC (Table 1). In

particular, many of these studies focused on differences between grey matter (GM) and white matter (WM).

Twenty-six studies used functional magnetic resonance

imaging (fMRI) to achieve the same goal (Table 2). They

evaluated abnormalities in brain functional connectivity (FC) parameters (regional homogeneity (ReHo), density), networks or the entire functional connectome.

Only one study used both structural and fMRI.

Structural neuroimaging

Salvador et al combined and compared the most common ML methods (ridge, Least Absolute Shrinkage

and Selection Operator (LASSO), elastic net and L0norm

regularized logistic regression, a support vector classiﬁer, regularized discriminant analysis, RF and a Gaussian

pro-cess classiﬁer) in a sample of 128 SCZ, 128 BD and 127

HC. They used different features: GM and WM voxel-based morphometry (VBM), vertex-voxel-based cortical thick-ness and volume, set as region of interest (ROI), and volumetric measures and wavelet-based morphometry as inputs for the techniques. VBM is the feature of choice for sMRI and it shows the most accuracy in distinguishing patients from HC, whilst the other ML methods (eg, ridge,

LASSO, a support vector classiﬁer, RFs and a Gaussian

Number of records identified via database search:

Number of records identified through other sources:

Number of records after duplicates removed:

Number of abstracts or full-text articles assessed for eligibility:

Number of studies included in review:

Number of abstracts or full-text articles excluded, with reasons:

Number of records excluded based on title irrelevance:

(Duplicates, editorials, letters, review, meta-analyses, guidelines, different interventions)

Number of records screened:

N=2386 N=0

N=2386

N=171

N=60

N=111

N:14 studies not using ML techniques

N:27 studies considering different diagnosis of SCZ

N:8 studies evaluating clinical features of SCZ

N:7 studies considering ML in therapy

N:5 studies considering ML in outcome and costs

N:15 studies not using MRI N=35

Included

Eligibility

Screening

Identification

Figure 1PRISMAﬂow diagram of included studies.

(4)

T

able

1

Summar

y

of

included

studies

of

machine

learning

techniques

classifying

schizophrenia

using

structural

neur

oimaging

First

author

,

Y

ear

Data

utiliz

ed

Sample

siz

e

and

dia

gnosis

Machine learning model Best accuracy (%) Other measur

es

(sensibility

,

speci

ﬁ

city

,

A

UC

,

pr

e-cision, err

or

rate)

Data

featur

es

Brain

regions

and

netw

orks

in

v

olv

ed

Commentar

y

Gr

eenstein,

2012

17

sMRI

1.5T

197

subjects:

-98

Childhood

onset

SCZ

-99

HC

RF

73.7%

N.A.

Cortical

thickness

for

68

fr

ontal,

temporal,

parie-tal,

and

occipital

lobe

regions,

and

bilateral

lat-eral

ventricle,

thalamus,

and

hippocampus

volumes

to

yield

the

74

variables

authors

used

as

featur

es

Left

temporal

lobes,

bilateral

dorsolateral

pr

efr

ontal

regions

and

left

medial

parietal

lobes.

Authors

ha

ve

chosen

to

use

regional

brain

mea-sur

es

that

pr

ovide

low

er

resolution

compar

ed

to

the

higher

resolution

vox

el-wise

measur

es

Castellani,

2012

15

sMRI

1.5T

108

subjects:

-54

SCZ

-54

HC

SVM

66.4

–

84.1%

N.A.

Landmark

point

detec-tion

of

DLPFC

and

Featur

e

vocabular

y

Dorsolateral

pr

efr

ontal

cortex

Findings

ma

y

ha

ve

par

-tially

been

limited

by

administration

of

AP

or

by

length

of

illness

Iwabuchi,

2013

20

sMRI

3T

and

7T

39

subjects:

-19

SCZ

-20

HC

SVM

GM

3T

66.6% 7T

77%

WM

3T

63.9% 7T

69.1%

Sn

63.2%

Sp

70.0%

Sn

78.9%

Sp

75%

Sn

57.9%

Sp

70.0%

Sn

63.2%

Sp

75.0%

The

normalized,

modu-lated,

unsmoothed

WM,

and

GM

images

for

the

3-and

7-T

datasets

w

er

e

then

used

as

inputs

to

the

separate

linear

SVM

classi

ﬁ

ers

Insula,

cortex,

thalamus,

super

-ior

temporal

cortex,

parahippocampal

gyr

us

First

study

compares

head-to-head

GM

and

WM

in

SVM

analysis

Lu,

2016

14

sMRI

3T

83

subjects:

-41

SCZ

-42

HC

SVM RFE SVM-RFE, 88.4%

Sn

91.9%

Sp

84.4%

First,

RGMV

and

R

WMV

in

the

VBM

analysis;

second,

the

sig-ni

ﬁ

cant

betw

een-gr

oup

differ

ences

for

both

RGMV

and

R

WMV

by

the

R

OI

ana-lysis

as

the

input

featur

es

for

the

classi

ﬁ

er

Middle

occipital

gyr

us,

calcarine,

cuneus,

fusi-form

gyrus,

lingual

gyrus,

hippocampus,

parahip-pocampal

gyr

us

Chr

onic

SCZ

patients

brain

structur

es

might

be

affected

by

antipsy-chotic

use

(

Continued

)

(5)

T

able

1

(Continued).

First

author

,

Y

ear

Data

utiliz

ed

Sample

siz

e

and

dia

gnosis

Machine learning model Best accuracy (%) Other measur

es

(sensibility

,

speci

ﬁ

city

,

A

UC

,

pr

e-cision, err

or

rate)

Data

featur

es

Brain

regions

and

netw

orks

in

v

olv

ed

Commentar

y

Pina

ya,

2016

18

sMRI

1.5T

258

subjects:

-143

SCZ

-32

FEP

-83

HC

DBN-DNN SVM DBN-DNN, 73.6%

±6.84

Sn

76.4%

±0.1

Sp

70.7%

±12.2

Err

or

rate

26.1

±6.5

Sixty-eight

brain

regions

(thirty-four

in

each

hemispher

e)

and

volumes

of

the

forty-ﬁ

ve

anatomical

structur

es

Fr

ontal,

temporal,

parie-tal

and

insular

cortices,

corpus

callosum,

puta-men,

cer

ebellum.

The

y

classi

ﬁ

ed

HC

and

SCZ;

FEP

as

thir

d

classi-ﬁ

cation

on

a

continuum

betw

een

HC

and

SCZ

Xiao

,

2017

16

sMRI

3T

326

subjects:

-163

Drug-naïv

e

FES -163

HC

SVM

surface

ar

ea

85,0% cortical thickness 81,8%

Sn

83,0%

Sp

87,0%

Sn

76.9%

Sp

85.0%

The

cortical

thickness

and

surface

ar

ea

featur

es

of

68

cortical

regions

fr

om

the

Desikan

atlas

w

er

e

extracted:

fr

ontal,

tem-poral,

parietal,

and

occi-pital

lobe

re

gions

Left

fusiform,

lingual,

posterior

cingulate

,

supramarginal,

insula,

inferior

parietal,

rostral

,

ros-tral

middle

fr

ontal

AND

right

isthmus

cingulate,

lateral

occipital,

lingual,

caudal

middle

fr

ontal,

inferior

parietal,

fr

ontal

and

temporal

pole

cortex

Most

of

inter

ested

ar

eas

belong

to

the

DMN,

sal-ience

netw

ork

or

visual

system

Salvador

,

2017

13

sMRI

1.5T

383

subjects:

-128

SCZ

-128

BD

-127

HC

Ridge LASSO SVC RF GPC Elastic

Net

L0

74

–

77%

(SCZ

vs

HC)

N.A.

Cortical

thickness

of

left

and

right

hemispher

es;

Cortical

volume

of

left

and

right

hemispher

es;

Gr

ey

and

White

matter

vo

x

el

based

morphome-tr

y

(VBM)

images;

Gr

e

y

and

White

matter

wa

ve-let

based

morphometr

y

(WBM)

images

Cortical

thickness,

Cortical

vol

ume

,

Gr

e

y

and

White

matter

Multi-class

classi

ﬁ

cations

considering

all

gr

oups

at

the

same

time

ha

ve

made

high

pr

edictiv

e

pow

er

for

the

SCZ

gr

oup

(

Continued

)

(6)

T able 1 (Continued). First author , Y ear Data utiliz ed Sample siz e and dia gnosis

Machine learning model Best accuracy (%) Other measur es (sensibility , speci fi city , A UC , pr e-cision, err or rate) Data featur es Brain regions and netw orks in v olv ed Commentar y Amin, 2018 47 sMRI 3T T1-w eighted 298 subjects: -144 SCZ -154 HC T ranslation- _based multi-modal fusion appr oach N.A. N.A. dFNC as the functional featur es and IC A-based sour ces fr om gra y mat-ter densities as the structural featur es Putamen, insular , pr ecu-neus, posterior cingulate cortex and temporal cortex The deep learning appr oach has a potential for learning dynamic fea-tur es fr om the fMRI data, and thus can offer a fa vorable frame w ork for multimodal fusion in the brain imaging resear ch. Pina ya, 2019 19 sMRI 75 subjects: -40 HC -35 SCZ Deep auto-encoder N.A. Mean de viation metric of 1,14 ±0.28 Cortical cortical thick-ness of the 68 cortical subr egions (34 per hemispher e) and the anatomical structural volume of 36 structur es; total number : 104 Left ve ntral diencepha-lon, left chor oid plexus, left lateral ventricle, right cuneus, right superior temporal, left putamen, right lateral ventricle, left cer ebellum cortex, left pr ecentral. The model was also able to detect distinct pat-terns of neur oanatomical de viations in SCZ. The deep autoencoder can be used to measur e the ov erall de viation metric of an individual and elu-cidate which re gions ar e the most differe nt com-par ed to health y gr oup . Note: Fo r this systematic re vie w the inclusion criteria was a Jadad scor e >3. Abbre viations: A OS, adolescent-onset schizophr enia; ASD , autism spectrum disor der ; A UC , ar ea under R OC cur ve; DBN, deep belief network; dFNC , dynamic functional connectivity; DLPFC , dorsolateral prefr ontal cortex; DNN, deep neural network; eMIC , extended maximal information coef fi cient; FCs, functional connectivities; FEP , fi rst episode psychosis; fMRI, functional magnetic resonance imaging; GM, gr e y matter ; GPC , Gaussian pr ocess classi fi ers; GSM, generalized sparse model; HC , health y contr ol; LD A, linear discriminant analysis; LIBSVM, lea ve-one-out SVM; LOOCV , lea ve-one-out cr oss validat ion; MDD , major depr essiv e disor der ; ML, machine learning; MVP A, multivariate pattern analysis; L0, L0 norm regularization; LASSO , Least Absolute Shrinkage and Selection Operator ; N.A., not applicable; PCC , P earson corre lation coef fi cient; RF , random for est; RGMV , re gional gr e y matter volume; RFE, re cursiv e featur e elimination; R OC, receiv er -operating characteristic cur ve analysis; R OI, region of intere st; rsMRI, resting state magnetic re sonance imaging; R WMV , re gional white matter volume; SCZ, schizophr enia; sMRI, structural magnetic re sonance imaging; Sn, sensitivity; SNPs, single-nucleotide polymorphisms; Sp , speci fi city; SR VS, sparse-r epr esentation-based variable selection; SVC , support vector classi fi er ; SVM, support vector machine; T , tesla; VBM, vo x el-based morphometr y; v-ELM, voting-ELM; ν -MKL, multiple k ernel learning; VMHC, vo x el-mirr or ed homotopic connectivity; WBM, wa velet-based morphometry ; WM, white matter .

(7)

T able 2 Summar y of included studies of Machine Learning techniques classifying Schizophr enia using functional neur oimaging First author , Y ear Data utiliz ed Sample siz e and dia gnosis

Machine learning model

Best accu-racy (%) Other measur es (sensibility , speci ﬁ city , AU C , pr e-cision, err or rate) Data featur es Brain regions and net-w orks in v olv ed Commentar y Y oon, 2008 24 fMRI 1.5T 34 subjects: − 19 SCZ -15 HC MVP A 93% N.A. Images w er e acquire d with a echo-planar imaging (EPI) in the anterior commissure- pos-terior commissure (A C-PC) aligned axial plane. T w enty-se ven interlea ved slices for whole-brain cov erage w er e collected with a 22-cm ﬁ eld of vie w Pr efr ontal Cortex, sensor y cortex, visual cortex Accuracy of MVP A was higher for HC gr oup Y ang, 2010 33 fMRI 3T

SNPs Combined fMRI-SNPs

40 subjects: − 20 SCZ -20 HC SVM-E/C Hybrid ML method SVM-C 67, 5% Hybrid ML 87, 3% SVM-C Sn 65% Sp 70% Hybrid ML Sn 85, 8% Sp 88.8% 150 SNPs fr om a database; auditor y stimuli w er e used and found to be effectiv e in eliciting fMRI BOLD patterns differ en-tiating HC fr om SZ subjects P ost-/pr e-central gyrus, para-central lobule, cingulate gyr us, superior and inferior parietal lobule, pr ecuneus Hybrid ML method using fMRI and SNP data Su, 2013 39 fMRI 1.5T 64 subjects: -32 SCZ -32 HC SVM MIC 76, 6% PCC 81, 2% eMIC 82, 8% MIC Sn 71, 9% Sp 81, 2% PCC Sn 81, 2% Sp 81, 2% eMIC Sn 81, 2% Sp 84, 4% 180 re gistere d fMRI volu mes with the MNI template than furthe r divided into 116 regions accor ding to the automatic anatomical labelling atlas. The atlas divides the cer ebrum into 90 regions (45 in each hemi-sphere ) and divides the cer e-bellum into 26 regions (nine in each cer ebellar hemispher e and eight in the ve rmis) Default mode netw ork, cer e-bellum, visual netw ork, sensor -imotor netw ork, fr onto-parietal netw ork, cingulo-oper cular netw ork The study does not include de ﬁ niton of R OIs and it is con ﬁ ned to connectivity analyses ( Continued )

(8)

T

able

2

(Continued).

First author

,

Y

ear

Data utiliz

ed

Sample siz

e

and

dia

gnosis

Best

accu-racy

(%)

Other measur

es

(sensibility

,

speci

ﬁ

city

,

AU

C

,

pr

e-cision, err

or

rate)

Data

featur

es

Brain

regions

and

net-w

orks

in

v

olv

ed

Commentar

y

Zhu,

2014

31

rs-fMRI

3T

45

subjects:

-27

SCZ

-28

HC

LOOCV

83.6%

Sn

81.5%

Sp

85.7%

Blood

o

xygen

lev

el

–

dependent

(BOLD)

time

courses

w

er

e

generated

for

23

regions

of

intere

st

(R

OIs)

W

ernick

e

’

s

ar

ea,

inferior

parie-tal,

Br

oca

’

s

ar

ea,pars

triangu-laris,

middle

fr

ontal,

pars

oper

cularis,

orbitalis,

inferior

temporal,

superior

fr

ontal,

caudate,

putamen,

ventral

tha-lamus,

cer

ebellum

crus,

striate,

extrastriate,

posterior

,

superior

parietal,

superior

temporale,

cingulate.

Authors

pr

opose

an

adaptiv

e

learning

algorithm

to

distin-guish

HC

and

SCZ

patients

using

resting-state

functional

language

netw

ork

Castr

o

,

2014

43

fMRI

52

subjects:

-31

SCZ

-21

HC

ν

-MKL

85%

Sn

90%

Sp

76%

Authors

used

the

simulation

toolbox

for

fMRI

data

(SimTB)

mimics

the

BOLD

re

sponse

of

subjects.

The

experimental

design

is

characterized

by

the

absence

of

task

e

vents.

Among

the

30

components

available

by

default

on

SimTB,

the

y

did

not

include

in

the

simulation

those

associated

with

the

visual

cor

-tex,

the

pr

ecentral

and

post-central

gyri,

the

subcor

tical

nuclei

and

the

hippocampus

Right

Caudate

Nucleus,

pr

ecu-neus,

cingulated

gyrus,

occipi-tal,

gyr

us,

parietal

lobe

and

left

pr

ecuneus,

temporal

gyrus,

parahippocampal

gyr

us,

para-central

lobule

All

patients

w

er

e

on

stable

medication

prior

to

the

scan

session

Arbabshirani, 2014

34

fMRI

3T

370

sub-jects:

−

195

SCZ -175

HC

SVM

88,

2%

Sn

86,

7%

Sp

89,

5%

47

ICNs

w

er

e

selected,

re

sult-ing

in

1081

FNC

featur

es

for

each

subjects:

in

total

w

e

extracted

1128

featur

es

for

each

subject

Subcortical,

auditor

y,

visual,

somatomotor

regions

and

con-tr

ol

pr

ocesses,

default-mode,

cer

ebellar

netw

orks

The

study

e

valuates

functional

netw

ork

connectivity

and

autoconnectivity

impr

oving

the

classi

ﬁ

cation

performance

signi

ﬁ

cantly

(

Continued

)

(9)

T

able

2

(Continued).

First author

,

Y

ear

Data utiliz

ed

Sample siz

e

and

dia

gnosis

Best

accu-racy

(%)

Other measur

es

(sensibility

,

speci

ﬁ

city

,

AU

C

,

pr

e-cision, err

or

rate)

Data

featur

es

Brain

regions

and

net-w

orks

in

v

olv

ed

Commentar

y

W

atanabe,

2014

38

rs-fMRI

91

subjects:

-54

SCZ

-67

HC

SVM LASSO Graph

Net

Elastic

Net

RO

C

77

–

88.2%

N.A.

Authors

pr

oduced

a

whole-brain

resting

state

functional

connectome.

347

non-ov

erlapping

spherical

nodes

ar

e

placed

thr

oughout

the

entire

brain

in

a

regularly-spaced

grid

pattern;

each

of

these

nodes

re

pr

esents

a

pseudo-spherical

R

OI.

F

or

each

of

these

nodes,

a

single

repr

esentative

time-series

is

assigned

by

spatially

av

eraging

the

BOLD

signals

falling

within

the

R

OI

Intra-fr

ontoparietal,

fr

ontopar

-ietal-default,

intracere

bellum

netw

orks,

lateral

pr

efr

ontal

cortex

Authors

intr

oduced

empirical

risk

minimization

frame

w

ork

that

accounts

for

6-D

spatial

structur

e

in

the

connectome

via

the

fused

LASSOand

Graph

Net

regularizer

Cao

,

2014

32

fMRI 1.5-3T SNPs

208

sub-jects: -92

SCZ

-116

HC

SR

VS

GSM

SNPs

83.1%

±1.3 fMRI

63.1%

±0.7

N.A.

SNPs/fMRI

vo

x

els

(n=759,075/

153,594.

F

or

each

method,

authors

carried

out

100

runs

and

the

av

erage

of

the

classi

ﬁ

-cation

ratios

was

used

as

the

ﬁ

nal

identi

ﬁ

cation

accuracy

T

emporal

lobe,

lateral

fr

ontal

lobe,

occipital

lobe

,

motor

cortex.

Differ

ent

natur

e

of

the

data

used

for

the

ML

analysis

Matsuraba, 2015

36

rs-fMRI

211

sub-jects: -48

SCZ

-46

BD

-117

HC

DGM

76.6%

Sn

58.5%

Sp

84.9%

Authors

par

cellated

each

fMRI

image

into

116

R

OIs

using

an

automated

anatomical

labeling

template. The

data

of

subjects

whose

fMRI

images

did

not

match

the

template

ev

en

after

the

spatial

normalization

w

er

e

also

discarded

Left

and

right

thalamus,

left

fusiform,

right

re

ctus,

right

middle

cingulum,

right

supra-marginal,

left

mild

temporal,

left

superior

orbital

fr

ontal,

left

superior

temporal,

right

caudate

The

DGM

was

a

generativ

e

model

implemented

using

deep

neural

netw

orks

(

Continued

)

(10)

T

able

2

(Continued).

First author

,

Y

ear

Data utiliz

ed

Sample siz

e

and

dia

gnosis

Best

accu-racy

(%)

Other measur

es

(sensibility

,

speci

ﬁ

city

,

AU

C

,

pr

e-cision, err

or

rate)

Data

featur

es

Brain

regions

and

net-w

orks

in

v

olv

ed

Commentar

y

K

och,

2015

23

fMRI

1.5T

98

subjects:

-44

SCZ

-54

HC

MVP

A

SVM RO

C

LIBSVM

69.3

–

93.2%

Sn

70.5

–

100%

Sp

40.9

–

93.2%

Gradient-echo

echo-planar

imaging

was

used

to

pr

oduce

eighteen

slices

appr

ox

imately

parallel

to

the

A

C-PC

plane,

cov

ering

the

inferior

part

of

the

fr

ontal

lobe

(superior

bor

der

abov

e

the

caudate

nucleus),

the

entire

temporal

lobe

,

and

large

parts

of

the

occipital

region.

Six

fMRI

volumes

w

er

e

acquir

ed

per

trial,

resulting

in

450

vo

lumes

per

run

V

entral

striatum,

right

pallidum,

putamen,

right

inferior

fr

ontal

gyrus,

nucleus

accumbens,

am

ygdala,

insula,

thalamus,

inferior

temporal

gyr

us

Se

ve

n

patients

taking

atypical

AP

.

21

typical

AP

and

16

not

receiving

an

y

medications

Ch

yzh

yk,

2015

30

rs-fMRI

147

sub-jects: -72

SCZ

-75

HC

V

-ELM

SVM RF

VMHC

≈

90%

N.A.

rs-fMRI

data

was

collected

with

single-shot

full

k-space

EPI

with

ramp

sampling

corr

ection

using

the

A

C-PC

as

a

re

fere

nce

.

The

ﬁ

rst

6

volumes

w

er

e

discar

ded

for

scanner

calibration

lea

ving

144

time

volumes

Inferior

temporal

gyrus,

para-hippocampalgyrus,

planumpo-lar

e,

temporal

fusiform

cortex,

thalamus

Application

of

SVM

and

RF

to

the

featur

es

extracted

fr

om

cr

oss-validation

as

the

ensem-bles

of

ELM

(

Continued

)

(11)

Best accu-racy (%) Other measur es (sensibility , speci fi city , AU C , pr e-cision, err or rate) Data featur es Brain regions and net-w orks in v olv ed Commentar y Cabral, 2016 21 rs-fMRI sMRI 145 sub-jects: -71 SCZ -74 HC MVP A rs-fMRI 70.5% sMRI 69.7% s/rs-fMRI 75% GM, Sn 82.4% Sp 63.4% acc . 69.7% RS, Sn 71.2% Sp 69.7% rs-fMRI, Sn 100% Sp 62.5 – 87.5% acc . 76.9 – 92.3% MRI, Sn 75% Sp 78.8% acc . 71.2% sMRI data w er e pr epr ocessed to segment the brain into WM, GM, and cer ebral spinal fl uid. Blood o xygenation le vel dependent images of the whole brain using an EPI sequence w er e acquire d in 32 axial slices using the A C-PC plane as a refer ence. RS scans resulted in 304 seconds duration (152 vo lumes). After discar ding the fi rst 10 images and 2 dumm y scans, the remaining 140 images w er e unwarped Fr onto-occipital, fr onto-parietal, fr onto .temporal, cor -tico-thalamic regions, left infer -ior temporal gyrus, parahippocampal gyr us The AP medication at MRI scan was con verted to chlorpr oma-zine and olanzapine equiva-lents; chr onic effects of AP could contribute to alterate FC with longer duration of illness Kim, 2016 37 rs-fMRI 3T 100 sub-jects: -50 SCZ -50 HC DNN Err or rate of DNN: 14.2% vs err or rate of SVM: 22.3%. Authors acquire d 150 vol umes. The fi rst 5 volumes w er e re mov ed to allow equilibration of the T1-r elated signal. Each of the 116 AAL re gions was re adily available in each vox el of the normalized EPI vol umes Whole-brain FC patterns The study successfully demon-strated the feasibility of the DNN classi fi er towar d the automated diagnosis of SZ patients Liu, 2017 29 rs-fMRI 3T 79 subjects: -48 Drug-naïv e SCZ -31 HC SVM 89.9% Sn 91.67% Sp 87.10% F or each subject, the fMRI scan lasted for 480s, and 240 vo lumes w er e obtained. The fi rst 10 volumes of each subject w er e discar ded to cer -tain steady-state conditions Left postcentral gyrus, left superior temporal gyrus, left paracentral lobule, right pr e-central gyr us, right inferior parietal lobule, right middle fr ontal gyr us, bilateral pr ecuneus Study includes only A OS with-out SCZ patients ( Continued )

(12)

T

able

2

(Continued).

First author

,

Y

ear

Data utiliz

ed

Sample siz

e

and

dia

gnosis

Best

accu-racy

(%)

Other measur

es

(sensibility

,

speci

ﬁ

city

,

AU

C

,

pr

e-cision, err

or

rate)

Data

featur

es

Brain

regions

and

net-w

orks

in

v

olv

ed

Commentar

y

Guo

,

2017

27

fMRI short/long- range

FCs

96

subjects:

-28

SCZ

-40

HC

-28 relativ

es

SVM RO

C

94.6%

Sn

92.9%

Sp

96.4%

Long-range

and

short-range

FCs.

rs-fMRI

images

w

er

e

obtained

with

a

gradient-echo

echo-planar

imaging

sequence

using

250

volumes

Default-mode

netw

ork,

sen-sorimotor

cir

cuits,

right

super

-ior

parietal

lobule,

left

fusiform

gyrus,

cer

ebellum

The

results

ma

y

be

confounded

by

acute

positiv

e

symptoms

since

the

patients

sample

was

unmedicated

and

re

cent

onset

Orban,

2017

45

fMRI

382

sub-jects: -191

SCZ

-191

HC

SVM

84%

N.A.

Functional

brain

connectomes

included

2016

functional

con-nections

betw

een

64

brain

par

cels

Whole

brain

connectivity

Brain

imaging

data

fr

om

6

independent

studies

and

datasets

W

ang,

2017

28

rs-fMRI

3T

79

subjects:

-48

A

OS

-31

HC

SVM

90.1%

Sn

88.2%

Sp

91.9%

Brain

regions

with

signi

ﬁ

cantly

differe

nt

ReHo

values

betwe

en

patients

with

A

OS

and

health

y

contr

ols

Bilateral

superior

medial

pr

e-fr

ontal

cortex,

left

superior

temporal

gyr

us,

right

pr

ecen-tral

lobule,

right

inferior

parie-tal

lobule,

left

paracentral

lobule

Differ

ent

number

of

partici-pants

in

the

tw

o

gr

oups.

with

A

OS

(with

less

confounding

factors)

instead

of

SCZ

Chen,

2017

22

rs-fMRI

3T

60

subjects:

-20

SCZ

-20

MDD

-20

HC

MVP

A

85%

SCZ

vs

MDD (average

79%)

Sn

98%

Sp

60%

The

scan

used

a

gradient-echo

echo-planar

imaging

sequence.

A

total

of

255

volu

mes

w

er

e

collected

for

each

subject.

Each

functional

volume

contained

35

slices.

To

minimize

the

effects

of

scanner

signal

stabilization,

the

ﬁ

rst

ﬁ

ve

volumes

of

each

subject

we

re

excluded

fr

om

all

analyses

Orbitofr

ontal

cortex

The

systematically

differ

ent

medications

for

SCZ

and

MDD

ma

y

ha

ve

differe

nt

effects

on

functional

connectivity;

the

medicated

patients

w

er

e

in

stable

condition

and

this

ma

y

ha

ve

an

impact

on

the

func-tional

connectivity

of

cortical

netw

orks

(

Continued

)

(13)

T

able

2

(Continued).

First author

,

Y

ear

Data utiliz

ed

Sample siz

e

and

dia

gnosis

Best

accu-racy

(%)

Other measur

es

(sensibility

,

speci

ﬁ

city

,

AU

C

,

pr

e-cision, err

or

rate)

Data

featur

es

Brain

regions

and

net-w

orks

in

v

olv

ed

Commentar

y

W

ang,

2017

26

rs-fMRI

3T

79

subjects:

-48

drug-na

ïv

e

A

OS

-31

HC

SVM FC

analysis

LIBSVM RO

C

92.4%

Sn

89.6%

Sp

96.8%

Authors

used

regional

homo-geneity

(ReHo),

a

measur

ement

that

re

ﬂ

ects

brain

local

func-tional

connectivity

or

synchr

o-nization

and

indicates

regional

integration

of

information

pr

ocessing

Right

middle

fr

ontal

gyrus,

right

superior

medial

pr

efr

ontal

cor

-tex,

left

superior

temporal

gyrus

SVM

analysis

applied

to

an

independent

dataset

Bae,

2017

40

fMRI

3T

75

subjects:

-21

SCZ

-54

HC

SVM

92.1%±10.5

Sn

92.0±15.8%

Sp

92.1±8.1%

Pre

cision

94

±6.3%

Authors

cr

eated

anatomic

labels

for

90

R

OIs

fr

om

the

image

database.

An

y

subject

with

fe

w

er

than

85

R

OIs

auto-matically

labeled

was

excluded

cingulate

cortex,

superior

right

temporal

region,

inferior

left

parietal

region

P

ossible

in

ﬂ

uence

of

pharma-cological

tre

atment

and

disease

stage

on

the

in

vestigated

func-tional

connections,

the

y

used

only

n-back

tests

without

rs-fMRI

Qur

eshi,

2017

44

rs-fMRI

3T

144

sub-jects: -72

SCZ

-72

HC

ELM SVM LD

A

RF

99.3%

Sn

100%

Sp

98,

6%

Authors

used

measur

es

includ-ing

the

mean

cortical

thickness,

cortical

thickness

standar

d

de

viation,

surface

ar

ea,

vo

lume

,

mean

cur

vatur

e,

white

matter

vo

lume

,

subcor

tical

segment

vo

lume

,

subcor

tical

intensity

,

and

ove

rall

brain

volume

and

intensity

as

the

structural

featur

es

Cortical

thickness,

Surface

ar

ea,

WM/subcortical/overall

volume,

cur

vatur

e,

global

av

er

-age

functional

connectivity

Authors

de

ve

loped

an

ELM,

whose

effectiv

eness

was

com-par

ed

with

that

of

mor

e

known

ML

methods

(

Continued

)

(14)

T

able

2

(Continued).

First author

,

Y

ear

Data utiliz

ed

Sample siz

e

and

dia

gnosis

Best

accu-racy

(%)

Other measur

es

(sensibility

,

speci

ﬁ

city

,

AU

C

,

pr

e-cision, err

or

rate)

Data

featur

es

Brain

regions

and

net-w

orks

in

v

olv

ed

Commentar

y

Rea

vis,

2017

25

fMRI

3T

148

sub-jects: -50

SCZ

-51

BD

-47

HC

MVP

A

41%

N.A.

Structural

data

w

er

e

pr

ocessed

and

par

cellated

into

anatomical

re

gions

used

to

constrain

R

OI

de

ﬁ

nitions

Lateral

occipital

lobe

The

paper

shows

MVP

A

can

be

used

successfully

to

classify

individual

per

ceptual

stimuli

in

SCZ

and

BD

.

The

re

sults

do

not

pr

ovide

gr

oups

differ

ences

with

utilizeted

stimuli

Chen,

2017

35

rs-fMRI

109

sub-jects: -35

SCZ

-31

HC

-22

ASD

-21

HC

MVP

A

LOOCV

Schizophr

enia,

83% ASD

,

80.4%

SCZ,

Sn

80%

Sp

87.1%

A

UC

0.83

ASD

,

Sn

77.27% Sp

83.23%

A

UC

0.76

A

total

of

255

volu

mes

w

er

e

collected

for

each

subject.

Each

functional

volume

contained

35

slices.

The

ﬁ

rst

ﬁ

ve

vol

umes

of

each

subject

w

er

e

excluded

fr

om

all

analyses.

Only

the

brain

ar

eas

within

the

ar

eas

identi

ﬁ

ed

in

the

MVP

A

w

er

e

included

in

the

subsequent

analyses.

The

y

w

er

e

regarde

d

as

meaningful

brain

ar

eas

and

used

as

R

OIs

in

the

functional

connectivity

analysis

Default-mode

netw

ork,

sal-ience

netw

ork

The

tw

o

dataset

w

er

e

fr

om

tw

o

differe

nt

imaging

platforms

Pläschk

e,

2017

41

rs-fMRI

170

sub-jects: -86

SCZ

-84

HC

SVM

61

–

72%

Sn

65

–

77%

Sp

46

–

69%

AU

C

0.61

–

0.79

Authors

in

vestigated

12

func-tional

netw

orks.

Only

meta-analytic

netw

orks

with

a

minimum

of

10

nodes

w

er

e

included,

since

a

low

er

number

of

featur

es

ar

e

uninformative

for

robust

classi

ﬁ

cation

Emotion-pr

ocessing,

empath

y

and

cognitive

action

contr

ol

netw

orks

Y

oung-old

classi

ﬁ

cation

was

based

on

all

netw

orks

and

outperformed

clinical

classi

ﬁ

cation

(

Continued

)

(15)

Best accu-racy (%) Other measur es (sensibility , speci fi city , AU C , pr e-cision, err or rate) Data featur es Brain regions and net-w orks in v olv ed Commentar y Liu, 2018 42 rs-fMRI 79 subjects: − 48 Drug-Naïv e FES AO S -31 HC SVM VMHC 94.93% Sn 100% Sp 87.09% Fo r e ac h su b jec t, th e fM R I sca n la ste d fo r 4 8 0 s, and 2 40 vo lu m e s w e re ob ta in e d. The fi rst 1 0 vol ume s of each subj e ct w er e d is ca rd e d to cer -ta in ste ad y-st ate con di ti on s an d for p arti ci pa nt s to ac cl im ati ze to a sc an n in g e n vir on m en t du rin g th e an al yz e d p ort io n of the d ata Fusiform gyr us, superior tem-poral gyrus, insula, pr ecentral gyrus and pr ecuneus authors also u se d a batt er y o f n e ur ocognitiv e tests and th ey de monstrated de fi cit s in multi p le cognitiv e funct ions in patients Zeng, 2018 46 rs-fMRI 1.5-3T 734 sub-jects: -357 SCZ -377 HC RFE-SVM RFE-LD A

SAN DANS

81 – 85% Sn 75 – 83% Sp 81 – 86% Authors used multi-atlas based whole-brain fcMRI in the MVP A, which measur es func-tional connectivity of the same image in differ ent spaces. The thr ee atlases used included 176, 160 and 116 R OIs re spective ly Cortical-striatal-cerebellar cir -cuit (default, salience , fr onto-parietal contr ol, ve ntral attention, dorsal attention and somatomotor , visual). This paper pr ovides for a multi-variate based whole-brain fcMRI pattern analysis to ensur e the optimal use of the w ealth of information pr esent in fcMRI scans. Amin, 2018 47 fMRI 3T

T2- weighted

298 sub-jects: -144 SCZ -154 HC T ranslation- _based

mul-timodal fusion appr

oach N.A. N.A. dFNC as the functional featur es and IC A-based sour ces fr om gr e y matter densities as the structural featur es Putamen, insular , pr ecuneus, posterior cingulate cortex and temporal cortex The d ee p learn in g appr oach has a p otential for lea rning d ynamic features fr om the fM R I d at a, an d th us ca n o ff er a fa vorab le fra m e-w ork for m ultimodal fusion in th e b rai n ima gi ng re se ar ch. Note: Fo r this systematic re vie w the inclusion criteria was a Jadad scor e >3. Abbre viations: A C-PC plane, bicommissural plane; A OS, adolescent-onset schizophr enia; ASD , autism spectrum disorder ; A UC , area under R OC cur ve, Cohe-ReHo , regi onal homogeneity based on cohere nce; D ANS, discriminant autoencoder netw ork with sparsity constraint; DBN, Deep Belief Netw ork; DNN, deep neural network; dFNC , dynamic functional connectivity; DGM, dee p neural generativ e model; eMIC , extended maximal information coef fi cient; EPI, echo-planar imaging; FC , functional connectivity; FEP , fi rst episode psychosis; fMRI, functional magnetic re sonance imaging; GM, gr e y matter ; GPC , Gaussian pr ocess classi fi ers; GSM, generalized sparse model; HC , health y contr ol; ICNs, intrinsic connectivity networks ; LD A, linear discriminant analysis; LIBSVM, lea ve-one-out SVM; LOOCV , lea ve-one-out cr oss validation; MD D , major depr essiv e disor der ; ML, machine learning; MVP A, multivariate pattern analysis; L0, L0 norm regularization; LASSO , Least Absolute Shrinkage and Selection Operator ; N.A., not applicable; PCC , P earson corre lation coef fi cient; ReHo , re gional homogeneity; RF , random fore st; RFE, recursiv e featur e elimination; R OC , re ceiv er -operating characteristic cur ve analysis; R OI, re gion of intere st; rsMRI, resting state magnetic resonance imaging; SAN, sparse autoenc oder network; SCZ, schizophr enia; sMRI, structural magnetic re sonance imaging; Sn, sensitivity; SimTB, simulation toolbo x for fMRI data; SNP , single-nucleotide polymorphisms; Sp , speci fi city; SR VS, sparse-r epr esentation-based variable selection; SVC , support vector classi fi er ; SVM, support vector machine; T , tesla; v-ELM, voting-ELM; ν -MKL, multiple k ernel learning; VMHC, vo x el-mirr or ed homotopic connectivity; WM, white matter .

(16)

process classiﬁer) have similar and lower performance levels.13

The grey matter volume (GMV) and WM volume of 41 SCZ patients and 42 HC were compared through a ML method providing both SVM and recursive feature elim-ination. Important abnormalities in brain structures

primar-ily involved in cognitive functions (eg, memory,

emotionality) were identiﬁed with 88.4% accuracy

(sensi-tivity 91.9%, speciﬁcity 84.4%). The authors concluded

with the possibility of identifying speciﬁc brain

neuroima-ging proﬁles associated with SCZ as a potential biomarker

for disease diagnosis.14

The dorsolateral prefrontal cortex (DLPFC) was cho-sen as the ROI in another study. The ROI-SVM approach was used to classify 54 SCZ patients and 54 HC with accuracy (left side 75%; right side 66.38%). The technique showed better results among females and older subjects, probably due to differences in GM distribution (left side 84.09%, right side 77.27% for females and left side 81.25%, right side 70.83% for seniors). According to the authors, DLPFC can be used as a brain structural marker for the disease, especially in older female chronic patients.15

Patterns of illness-related grey matter changes could be potential biomarker for identifying structural brain altera-tions in individuals with SCZ. Xiao et al used SVM to analyze cortical thickness and surface area measurements in the differentiation of 163 FES patients and 163 HC.

Regions contributing to classiﬁcation accuracy included

the grey matter in default mode network (DMN), central executive network, salience network (SN) and visual net-work: left fusiform, lingual, posterior cingulate, supramar-ginal, insula and right isthmus cingulate, lateral occipital, lingual and frontal pole cortex. The accuracy of correct

classiﬁcation of patients and HC was 85.0% for surface

area and 81.8% for cortical thickness.16

The RF method with 74 anatomic brain MRI subre-gions was used to classify 98 childhood-onset SCZ patients and 99 HC. The association of copy number variations and these 74 subregions in the RF produced

74% accuracy. The brain areas speciﬁcally involved were

the left temporal lobe, bilateral dorsolateral prefrontal region and left medial parietal lobe.17

Pinaya et al developed a new deep learning model, named the Deep Belief Network, used for extrapolating and interpreting features from neuromorphometry data on 83 HC and 143 patients with SCZ. Deep Belief Network accuracy, utilized as SVM was 68.1%, whereas as

a classiﬁer it was 73.6%, detecting large differences

between classes, especially for frontal, temporal, parietal and insular cortices and the corpus callosum, putamen and cerebellum.18

In a later study, it was proposed an alternative concep-tual and practical approach for investigating brain-based disorders that not requires a large number of cases. They

used an artiﬁcial neural network known as “deep

autoen-coder”to create a normative model using sMRI data. The

model was able to generate different values of total neu-roanatomical deviation and distinct patterns of neuroana-tomical deviations.19

Finally, pathologic changes both in WM and GM (in bilateral insula, anterior cingulate cortex, thalamus, super-ior temporal cortex and parahippocampal gyrus) were demonstrated though SVM. Changes were more evident at 7T RMI rather than at 3T RMI; the accuracy was higher with 7T in discriminating patients from controls (77% vs

66%) (20 HC vs 19 SCZ).20

Functional neuroimaging

Cabral et al applied MVPA to both structural and func-tional magnetic resonance neuroimaging methods on 21 SCZ patients and 74 HC. The resting state fMRI (rs-fMRI) reported similar accuracy to the structural classiﬁer (70.5% vs 69.7%, respectively). The combination of sMRI and rs-fMRI outperformed single MRI modalities by reaching 75% accuracy. Subcortical short-range, in particular inter-hemispheric connections, were the main aberrant

connec-tions identiﬁed in SCZ patients. Abnormalities found in

the frontotemporal area could explain cognitive

impair-ment (eg, reduced verbalﬂuency, negative symptoms).21

Chen et al used rs-fMRI to compare three groups (20 SCZ, 20 MDD and 20 HC) through local FC density analysis combined with MVPA. Based on this method, they could differentiate MDD patients from SCZ patients according to the local FC density value in the orbitofrontal cortex; in fact,

the connection in the prefrontal cortex was signiﬁcantly

lower in patients with SCZ compared with other groups

(accuracy 85%). Theseﬁndings emphasize the great

impor-tance the prefrontal cortex may have in future ML studies attempting to improve the diagnosis of SCZ by differentiat-ing it from other pathological conditions such as MDD. A limitation of the study could be the enrolment of patients who were in a chronic phase and receiving pharmacological treatment, which could alter FC.22

Koch et al reached 93% accuracy in identifying SCZ patients through the MPVA analysis of patterns of

(17)

anticipation of a monetary reward. They studied the fron-tal, temporal, occipital and midbrain regions, with a peak in the right pallidum and claimed to predict the severity of the negative symptoms of patients based on ventricular striatal activation patterns. Therefore, they conclude that

such ML methods could bring signiﬁcant beneﬁts to

diag-nostic classiﬁcation based on fMRI signal patterns.23 Yoon et al enrolled a sample of 19 SCZ and 15 HC and

tried to decode the speciﬁc neural networks activated in

the recognition and classiﬁcation of stimuli. Two cycles of stimuli including faces, scenes, objects and scrambled images were presented to participants. Functional MRI

data were analyzed through MVPA. The authors’

hypoth-esis about an alteration in brain connections in SCZ

patients was conﬁrmed: patients were less accurate than

HC (59% vs 72%. respectively); moreover, the distributed representations of visual objects were compromised in

SCZ. In both groups, the accuracy of the classiﬁcation

was signiﬁcantly correlated with behavioral measures.

This impairment was directly correlated with a decrease in performance in the 1-back matching task. Patterns of cortical activity (prefrontal, sensory and visual cortex) are reﬂected in altered behavior. The authors pointed out that

the impairment emerging from fMRI data could reﬂect the

neural abnormalities at the basis of poor performance in the matching task.24

Reavis et al showed four different objects and an outdoor scene to 51 BD patients, 50 SCZ patients and 31 HC during the execution of a fMRI and then performed a MVPA classi-ﬁcation analysis to assess the distinctiveness of activity cor-responding to the perception of each stimulus in the lateral

occipital complex. As the authors did not ﬁnd differences

between groups (accuracy 41%), they hypothesized that if a deﬁcit exists, it is linked to speciﬁc categories of objects (not between-group but within-group differences).25

In a recent paper, Wang et al explored long- and short-range functional connectivity (lFC) (sFC) through rs-fMRI

in 48 ﬁrst-episode, drug-naïve AOS patients and 31 HC

using SVM. They found abnormalities of lpFC and spFC in some brain networks (both lpFC and spFC increased in the anterior DMN and decreased in the posterior DMN and lpFC only decreased in the SN). They were able to classify patients and controls with high accuracy (up to 92.4%), sensitivity (89.6%) and specificity (96.8%). This study was thefirst to examine lFC and sFC in the whole brain infi rst-episode drug-naïve AOS patients through rs-fMRI. The authors concluded that the functional alterations could point to a role of DMN and SN in the neurobiology of

SCZ that is already present in patients in theﬁrst episode, hence their ability to discriminate AOS patients from HC with great diagnostic accuracy.26

Another paper published by Guo et al27also examined

the role of s-lFCs in fMRI in differentiating patients from

HC or their own relatives. Speciﬁcally, SCZ patients

showed an increase in spFC and lpFC in the DMN, while, unlike in the previous manuscript, they both decreased in sensorimotor circuits. In addition, the combi-nation of spFC values in the right superior parietal lobule and lpFC in the left fusiform cerebellum gyrus was able to differentiate patients from their relatives with an adequate level of sensitivity and speciﬁcity. Wang et al enrolled 49 drug-naïve AOS patients and 32 HC to investigate and identify brain abnormalities using ReHo input in SVM analysis through rs-fMRI. The study found that AOS

patients showed signiﬁcantly increased ReHo values in

the bilateral superior medial prefrontal cortex and decreased ReHo values in the left superior temporal gyrus, right precentral lobule, right inferior parietal lobule,

and left paracentral lobule compared with HC.28

Liu et al used a similar sample with 48 AOS and 31 HC to analyze the coherence regional homogeneity (CoHe-ReHo) through ML methods with rs-fMRI. The Cohe-ReHo value was reduced in the left postcentral gyrus, left superior temporal gyrus, left paracentral lobule, right precentral gyrus, right IPL, right middle frontal gyrus and bilateral praecuneus. It did not appear to increase in

any cerebral area in AOS patients compared with HC.29

Chyzhyk et al used the amplitude of low-frequency

ﬂuctuations, fractional ALFF, voxel-mirrored homotopic

connectivity (VMHC) and ReHo as input data for voting-ELM (v-voting-ELM) (SVM+RF) analysis. The accuracy of this method proved to be close to 90%. The study compared 72 SCZ patients and 75 HC. The anatomic brain regions with major alterations were the inferior temporal gyrus, anterior division of the parahippocampal gyrus, planum polare,

temporal fusiform cortex and left and right thalamus.30

Zhu et al developed a new method named the

“Adaptive Learning Algorithm” to distinguish 27 SCZ

patients from 28 HC using rs-fMRI data. They identiﬁed

23 ROI of the resting-state functional language network that are considered important in language comprehension and production (eg, Wernicke’s area, Broca’s area, inferior parietal, inferior temporal). They reached an accuracy of 83.6%, a sensitivity of 81.5% and a speciﬁcity of 85.7%.31 Cao et al integrated fMRI data with single-nucleotide polymorphisms (SNPs) in their research with 92 SCZ