PSY410 Northeastern Illinois University Future of Psychiatry Assessment Paper Reading the article and then answer the question: Why the paper is the future

PSY410 Northeastern Illinois University Future of Psychiatry Assessment Paper Reading the article and then answer the question: Why the paper is the future of assessment, where is ten or twenty years from now will improve, and connect with a Tedtalk from Xiaosi Gu named what big data can tell us about what the brain believes Molecular Psychiatry
https://doi.org/10.1038/s41380-019-0441-1
REVIEW ARTICLE
Conceptualizing mental disorders as deviations from
normative functioning
Andre F. Marquand1,2,3 Seyed Mostafa Kia
Christian F. Beckmann1,2,5
?
1,2
?
Mariam Zabihi1,2 Thomas Wolfers1,2 Jan K. Buitelaar
?
?
1,2,4
?
1234567890();,:
1234567890();,:
Received: 29 December 2018 / Revised: 15 April 2019 / Accepted: 29 April 2019
© The Author(s) 2019. This article is published with open access
Abstract
Normative models are a class of emerging statistical techniques useful for understanding the heterogeneous biology
underlying psychiatric disorders at the level of the individual participant. Analogous to normative growth charts used in
paediatric medicine for plotting child development in terms of height or weight as a function of age, normative models chart
variation in clinical cohorts in terms of mappings between quantitative biological measures and clinically relevant variables.
An emerging body of literature has demonstrated that such techniques are excellent tools for parsing the heterogeneity in
clinical cohorts by providing statistical inferences at the level of the individual participant with respect to the normative
range. Here, we provide a unifying review of the theory and application of normative modelling for understanding the
biological and clinical heterogeneity underlying mental disorders. We ?rst provide a statistically grounded yet non-technical
overview of the conceptual underpinnings of normative modelling and propose a conceptual framework to link the many
different methodological approaches that have been proposed for this purpose. We survey the literature employing these
techniques, focusing principally on applications of normative modelling to quantitative neuroimaging-based biomarkers in
psychiatry and, ?nally, we provide methodological considerations and recommendations to guide future applications of these
techniques. We show that normative modelling provides a means by which the importance of modelling individual
differences can be brought from theory to concrete data analysis procedures for understanding heterogeneous mental
disorders and ultimately a promising route towards precision medicine in psychiatry.
In most areas of medicine, biomarkers that objectively
indicate disease state have revolutionized diagnosis and
treatment allocation. In contrast, psychiatric disorders are
still diagnosed exclusively on the basis of symptoms and
* Andre F. Marquand
a.marquand@donders.ru.nl
1
Donders Centre for Cognitive Neuroimaging, Donders Institute for
Brain, Cognition and Behaviour, Radboud University,
Nijmegen, the Netherlands
2
Department of Cognitive Neuroscience, Radboud University
Medical Centre, Nijmegen, the Netherlands
3
Department of Neuroimaging, Centre for Neuroimaging Sciences,
Institute of Psychiatry, King’s College London, London, UK
4
Karakter Child and Adolescent Psychiatric University Centre,
Nijmegen, the Netherlands
5
Oxford Centre for Functional Magnetic Resonance Imaging of the
Brain (FMRIB), University of Oxford, Oxford, UK
biological tests to assist diagnosis or treatment allocation
remain to be developed [1]. This yields clinical groups that
are highly heterogenous, both in terms of clinical presentation and underlying biology, which is a major barrier
to understanding underlying mechanisms and developing
better treatments [1–3]. This is widely recognized at a
theoretical level and over the years, many different theoretical models have been proposed to explain the heterogeneity of psychiatric disorders [4–9]. These emphasize the
myriad pathological mechanisms that may converge on
the same symptoms in different participants [4, 5] and that
the same underlying biological risk factors may result in a
different clinical phenotype in different individuals
depending on the context, genetic background and critical
time window [4, 9]. Heterogeneity remains the dominant
theme even in recent large-scale theoretical initiatives such
as the Research Domain Criteria (RDoC) [6, 7] and the
European Roadmap for Mental Health Research (ROAMER) [8]. All these theoretical models are founded on an
implicit recognition of the importance of modelling
A. F. Marquand et al.
individual differences within and across clinical cohorts
[10, 11]. It is very important to recognize, however, that this
broad theoretical recognition is not re?ected in the data
analysis strategies employed in practice. Instead, the overwhelming majority of analysis approaches remain focussed
on group averages (e.g. the ‘average patient’) and regard
individual differences principally as noise.1
Normative modelling is an emerging approach that can
address this challenge by providing statistical inferences at
the level of the individual with respect to an expected pattern [12]. This is analogous to the widespread use of normative growth charts in paediatric medicine to map child
height or weight as a function of age with respect to centiles
of variation in a reference population [13]. Normative
modelling generalizes this notion by substituting these
variables for clinically relevant variables then applying
automated statistical techniques to map centiles of variation
across the cohort. This is increasingly used to map variation
between cognitive, clinical or demographic variables and
quantitative biomarkers derived from neuroimaging
[12, 14–17]. The key feature of normative modelling that
makes it useful for stratifying cohorts is that it permits the
detection and mapping of distinct patterns of abnormality in
individuals without requiring a consistent neurobiological
signature across all individuals.
Whilst early applications focused on brain development
and ageing [16, 18], normative modelling has recently been
shown to be highly promising for psychiatry
[12, 14, 17, 19–21]. First, to map variation related to brain
development and ageing in psychiatric disorders, which is
appealing given the neurodevelopmental basis of mental
disorders [22]. For example, neurodevelopmental normative
models have been used in the context of schizophrenia
[17, 19], attention de?cit/hyperactivity disorder (ADHD)
[20, 23] and autism [21, 24] to help understand the emergence of mental disorders as deviations from an expected
developmental trajectory and identify individuals following
an atypical trajectory. Second, normative modelling can
also be abstracted beyond development to chart the spectrum of functioning across any cognitive domain. It has
been used, for example, to chart variation in reward systems
via mappings between trait measures of reward sensitivity
and reward-related brain activity [12]. Finally, normative
models can help to understand healthy variation and move
beyond simple dimensional theories of mental disorders
[25].
1
This includes classical analysis methods based on the general linear
model (e.g. t-tests and analysis of (co-) variance) plus classical and
supervised machine learning classi?cation models (e.g. linear discriminant analysis, logistic regression and support vector machines)
and unsupervised clustering methods (e.g. k-means), where—at best—
confound regressors are used to remove some of the cross-subject
variation.
Here, we provide a unifying review of normative modelling for charting individual variation across different
behavioural, demographic and biological dimensions,
thereby helping to understand heterogeneity within clinical
cohorts. We ?rst provide a statistically grounded, yet nontechnical overview of its conceptual underpinnings and
propose a framework to link the many different methodological approaches that have been proposed. Second, we
outline connections between normative modelling and
existing approaches for tackling heterogeneity including
clustering [26] and ‘brain age’ approaches that characterize
subjects in terms of a difference between a brain-derived
predicted age and true chronological age [27, 28]. Third, we
survey the literature employing normative modelling in
clinical conditions. Fourth, we discuss normative modelling
as a tool for ?nding structure in large cohorts, which is
important given the recent shifts towards ‘big data’ neuroscience and large population-based cohorts [29–31].
Finally, we provide recommendations for future studies and
critically evaluate the limitations of normative modelling.
Introduction to normative modelling for
understanding heterogeneity in clinical
cohorts
Normative modelling is a statistical framework for mapping
between behavioural, demographic or clinical characteristics and a quantitative biological measure, providing
estimates of centiles of variation across the population
(Fig. 1a). Normative modelling provides a concrete method
for studying individual differences and parsing heterogeneity across cohorts because it provides statistical inferences at the level of the individual participant as to the
degree to which each individual deviates from the normative pattern and allows these deviations to be mapped in
each individual. In other words, normative modelling provides a way to quantify and characterize the manner in
which different individuals deviate from the expected pattern, and from one another. Importantly, this does not
require that atypicalities overlap across participants (e.g. in
the same brain regions) or even that a consistent pattern of
deviation exists. Therefore, this accommodates the convergence of multiple pathological pathways on the same
symptoms in different individuals [4]. This is clearly different from case-control analyses, which all focus on ?rst
order statistics (group means), thereby seeking a consistent
pattern of atypicality (i.e. the ‘average patient’). In a casecontrol context, heterogeneity becomes apparent via in?ation of the model residuals and ultimately decreases sensitivity for detecting disorder-related effects. In contrast,
normative modelling explicitly models heterogeneity
because it focuses on modelling individual variation around
Conceptualizing mental disorders as deviations from normative functioning
Fig. 1 Conceptual overview of normative modelling. a Normative
modelling is similar to the use of growth charts in paediatric medicine,
except the conventional response variable (e.g. height or weight) is
substituted for a quantitative biological readout (e.g. regional brain
activity). The classical covariates (age and sex) can also be substituted
for clinically relevant variables. Normative modelling provides statistical inference at the level of each subject with respect to the normative model (red ?gure). b Procedural overview of normative
modelling. After the choice of reference cohort and variables, the
normative model is estimated, before being validated out of sample on
new response variables and covariates (y* and x*, respectively).
Finally, the estimated model can be applied to a target cohort (e.g.
clinical cohort). c A common con?guration for normative modelling of
neuroimaging data, where a separate normative model is estimated for
each sampled brain location. This can be described by a set of functions (y = f(x)) predicting neurobiological response variables (y) from
clinical covariates (x). d Normative models can also be estimated for
the opposite mapping, where brain measures are chosen as covariates
and age or other covariates are chosen as a response variable. See text
for further details
the mean using second order statistics (variances). Therefore, normative modelling explicitly characterizes and
quanti?es the heterogeneity underlying clinical conditions
at a ?ner grained level than is afforded by group averages.
Normative modelling also does not require that the clinical
group can be cleanly partitioned into subtypes [26] although
it can be used to generate features for clustering. Normative
modelling can be used to estimate many different kinds of
mappings based on the variables chosen, but here we focus
on mappings between behavioural or demographic measures and a quantitative biological readout, most commonly
derived from neuroimaging.
Procedurally, normative modelling involves four steps
(Fig. 1b): First, a reference cohort and a set of variables are
chosen to de?ne the mapping and population over which
variation is measured. Second, a statistical model is estimated to model variance in a response variable (a.k.a. target
or dependent variable) from a set of clinically relevant
covariates (predictor or independent variables) across the
reference cohort. For example, one may estimate a normative model for cortical thickness as a function of age and
gender using a population-based reference cohort. Third, it
is necessary to assess the accuracy of the normative model
for predicting the response variable (e.g. mean-squared
A. F. Marquand et al.
error, explained variance). To ensure accurate estimates of
generalizability, this must be performed on withheld data
(e.g. under cross-validation). Finally, this model can be
applied to quantifying the deviations of samples from a
target cohort (e.g. clinical cohort) with respect to this
reference model.
Many regression models have been proposed for normative modelling, including hierarchical linear models,
polynomial regression, quantile regression, support vector
regression and Gaussian process regression (Table 1). The
estimation of normative models is conceptually similar to
classical growth charts, for which many approaches have
been proposed [13]. In both cases, the data make multiple
demands from the regression model including ensuring
precise estimation of outer centiles (where data are sparsest), ensuring centiles vary smoothly as a function of the
covariates (and do not cross) and the ability to estimate
deviations for individual samples via analytical formulae
(e.g. Z-scores) [32]. In order to provide a conceptual framework linking these approaches, we categorize different
approaches according to three criteria: (i) the choice of
covariates and response variables; (ii) the degree to which
the model separates different sources of variation and (iii)
the degree to which the model permits statistical inference
the individual level. However, other features are also
important, for instance ability to model non-linear
relationships.
Choice of covariates and response variables
A simple way to categorize different approaches is in terms
of the variables that de?ne the mapping. One common
con?guration (Fig. 1c) uses age as a covariate, often in
combination with other clinical or demographic variables to
predict a quantitative biological readout. However, there are
other possibilities: for example, the mapping can be inverted such that age is the response variable which is predicted
from clincal or demographic variables (Fig. 1d). This is the
approach used by ‘brain age’ models [27, 28] which use
multivariate regression to predict age from a pattern of
brain-derived measures. For linear models, it is obvious that
an association can be detected in either direction simply by
inverting the linear model. However, the interpretation of
the centiles of variation and regression coef?cients differs
and we consider that charting variation over the biological
readout is more appealing because it directly mirrors the use
of growth-charting in paediatric medicine.
As noted, normative models are also not restricted to
charting variation across development. By substituting age
for other variables, normative models can chart variation in
any kind of mapping, for example to link cognitive scores
with brain activity patterns [12].
Separating different sources of variation
across the cohort
Normative modelling principally aims to model variation
across the cohort over and above estimation of mean
effects. To achieve this effectively, it is important to
separate different sources of variation, most importantly
to differentiate actual variation within the data (i.e. across
participants) from variability due to parameter and model
uncertainty (i.e. induced variability due to a lack of data).
In normative modelling, we quantify these variabilities
using two types of uncertainty commonly de?ned in
machine learning [33]: (i) irreducible (or ‘aleatoric’)
uncertainty that re?ects true underlying variability that
cannot be reduced with more data; (ii) reducible (or
‘epistemic’) variation that re?ects parameter uncertainty
or ignorance about the true model and can be reduced by
more data. Aleatoric uncertainty is of primary interest for
strati?cation because it re?ects variation across subjects
whereas epistemic uncertainty is nuisance variation that it
is desirable to minimize. The degree to which different
approaches account for these sources of variability can be
classi?ed hierarchically (Fig. 2): the simplest approach
involves estimating the mean effect only and assessing
deviations from the expected pattern using the model
residuals (Fig. 2a, b) [19, 27]. Whilst this is appealing in
its simplicity, it provides no estimate of variation across
the cohort and cannot provide statistical inferences at the
individual level (see below). This has been addressed in
different ways, for example via estimating con?dence
intervals via a post hoc regression between the residuals
of the model against the true response variable [14] or
using quantile regression to directly estimate centiles of
variation in the data [34]. These approaches provide statistical estimates of variation within the population and
can indicate if a particular participant deviates from the
expected pattern at a given con?dence level, but they do
not fully account for different sources of uncertainty, e.g.
uncertainty in the estimation of the centiles (Fig. 2c).
Bayesian methods such as Gaussian process regression
[12, 16, 35] provide one solution to this problem (Fig. 2d)
by estimating distinct variance components and providing
predictions for each participant that account for all sources of uncertainty. This is important for two reasons: ?rst,
it provides estimates of centiles of variation within the
reference cohort that are not in?uenced by data density;
second, it allows all sources of uncertainty to be taken into
account when making predictions. This provides the
desirable property that inferences become more conservative in regions of the input space where data are
sparse. With these complementary purposes in mind, it
may be desirable to report different variance components
separately.
Healthy participants, participants
with mild cognitive impairment and
dementia
Developmental study cohort
containing typically developing
adolescents and adolescents with
psychosis spectrum symptoms (2
levels) and other psychopathologies
Age, sex, total grey- and
Multi-study healthy reference cohort
white matter volume, total
cerebrospinal ?uid, MRI ?eld
strength
Developmental study cohort
containing typically developing
adolescents and adolescents with
psychosis spectrum symptoms (2
levels) and other psychopathologies
Age
Cortical Thickness
Cortical thickness
Alpha band brain activity
derived from electroencephalogaphy
Brain volume
Age
Task-related fMRI data
Functional connectivity
measures derived from
resting fMRI
Brain volume
Brain volume
Autism
Autism
Bipolar disorder
Cognition:
processing speed
Cognition: Inhibitory
Control
Sustained attention
Mild cognitive
impairment and
dementia
Mild cognitive
impairment and
dementia
Brain volume
Longitudinal cortical
thickness measures
derived from
structural MRI
Schizophrenia
Schizophrenia
Age
Age and gender
Cognitive scores measuring
executive function, memory,
complex cognition, social
cognition and sensorimotor
processing
Healthy reference cohort
Healthy reference cohort
Population-based cohort
Population-based cohort
Healthy volunteers
Healthy volunteers
Healthy reference cohort
Typically developing cohort
Typically devel…
Purchase answer to see full
attachment