Heavyweight review: strategies for reliable fMRI measurements

Abstract ：fMRI Has considerable potential , It can be used as a conversion tool , Used to understand risks 、 Prioritize interventions , And improve the treatment of brain disorders . However , Recent research has found , Many of the most widely used fMRI measurement methods have low reliability , Undermines this potential . ad locum , We believe that many fMRI measurements are unreliable , Because they are designed to identify group effects , Rather than accurately quantifying individual differences . then , We highlighted four emerging strategies [ Extended aggregation 、 Reliability modeling 、 Multiple echo functional magnetic resonance imaging (ME-fMRI) And exciting design ], They are based on established psychometric properties , To produce more accurate and reliable fMRI measurements . By adopting these strategies to improve reliability , We are right. fMRI Optimistic about its potential as a clinical tool .

1.   Can we reliably measure individual differences in brain function ？

      Cognitive neuroscience has revolutionized our understanding of how the brain supports behavioral functions from basic senses to complex cognitive processes . Based on these basic understanding of the brain and behavior , An emerging translational neuroscience project seeks to identify individual differences in these patterns , And provide information for the development of clinical biomarkers , These biomarkers can be used to predict disease risk , Prioritize interventions , And improve treatment . The core of these efforts is functional magnetic resonance imaging (fMRI), Because it provides a non-invasive measure of brain activity during a lifetime of human behavior .

      In recent years ,fMRI There has been a surge in research on individual differences in the field of clinical significance , The expectation of clinical application also increases . Use fMRI The expansion of individual difference studies conducted raises questions about whether they are ready to meet the measurement characteristics required for clinical transformation , The most important one is reliability .

      Psychometrics has long established that reliability is the first step in validity . for example , To study how brain function makes a super old man resilient during neurodegeneration , Or adjust the brain stimulation according to the individual's unique functional terrain , We must first be able to reliably measure the characteristics of brain function . To establish reliability , There is no significant change in individuals ( Such as disease progression 、 Receive treatment ) Under the circumstances , Repeated measurements of brain function must produce convergent estimates . lately , We report on a number of tasks that are widely used in clinically relevant behaviors - Functional magnetic resonance imaging (task-fMRI) Methods of measuring brain activity ( Such as episodic memory 、 Executive control ) The test retest reliability of is low , therefore , In the present state , They cannot be used as clinical biomarkers . The results of similar studies also show that , Other widely used fMRI measurement methods have low reliability , Including functional connection measurement generated by short scan . Fortunately, , Methods to improve reliability have been used in related fields of psychology ( for example , Personality or cognitive assessment ) It has been developed and applied for a long time . However , These methods are in fMRI It has not been fully applied in the research .

      In this review , We begin to describe historical trends , This has led to the widespread use of unreliable in the study of individual differences fMRI measurement . then , We highlighted four emerging strategies ( Extended aggregation 、 Reliability modeling 、ME-fMRI And exciting design ), Each strategy has its roots in psychometrics , It allows researchers to reliably measure individual differences in brain function ( chart 1, Key diagram ). Our conclusion is , Despite some false starts and dead ends , But use fMRI The accumulation of 、 The neuroscience of transforming individual differences has a bright future , But to realize this future , We must subvert the status quo , Support the sound principles of psychometrics , To promote the development of measurement .

chart 1 A typical fMRI The average time of study passage ( Usually 5 - 10 Minute data ) To measure the activation or function connection

2. fMRI A brief history of the study of individual differences

      1992 year 3 month ,Kwong Et al Ogawa Et al. Reported the first batch of utilization MRI The study of mapping human brain function . In every study , Contrast the activity between alternating darkness and visual stimuli by implementing the subject's internal design , Use blood oxygen level dependent (BOLD) Signals to measure activity patterns in the visual cortex . These simple and powerful experiments prove that fMRI (fMRI) Potential for noninvasive measurement of human brain activity , And triggered a series of upsurge to further study human brain function .

      fMRI The first decade of began with a technical trigger , Then there is iterative development 、 Widely used and used in fMRI Translate into our understanding and treatment of depression 、 Expectations of progress from schizophrenia to Alzheimer's disease ( chart 2).

      In the second decade , Functional magnetic resonance imaging (fMRI) has expanded in both breadth and depth . More powerful scanners provide measurements of brain activity with greater spatial and temporal resolution . meanwhile , Functional magnetic resonance imaging (fMRI) is increasingly used in the study of special populations ( Like children 、 Brain disorders ).

It is in this expansion phase , Some researchers began to use fMRI tasks , The task was originally developed to stimulate powerful effects within the subject , To explore individual differences between subjects .

      stay 2010 years , Two emerging trends have attracted more attention and examination of functional magnetic resonance imaging at the same time . First , stay fMRI Driven by maturity and Prospects , Large scale joint research ( Such as the human connectome project , The establishment of the British biological bank , The clear goal is to use experimental cognitive neuroscience fMRI Task to measure individual differences in brain function . secondly , In the repeated crisis of replication in Psychology , A wave of research has found the key limitations of mainstream fMRI practice , The discovery of many individual differences has been questioned .

      These observations have inspired methodological innovation , Directly address many of these limitations , Including more accurate statistical inference methods 、 A larger sample 、 Multivariate modeling to improve reliability 、 Motion review and advanced data processing technology . However , Rely on short scans ( namely 5-10 minute ) In small samples (n≤100 ) And the ability of rigid stimulus control and group average continuous restriction fMRI to reliably measure individual differences in brain function , Represents convertible and easy to operate mechanism risks 、 Pathophysiology and therapeutic response . Even if there are advanced methods to reduce artifacts and statistical inference , frequently-used fMRI Methods often produce unreliable measurements . therefore , This unreliability continues to represent a fundamental threat to our ability to translate neuroscience rigorously to individual differences . Since reliability is the lowest level of effective individual difference research 、 Necessary prerequisites , More and more fMRI Researchers are trying to build a new framework for translational neuroscience by asking a basic question : Under what conditions , Functional magnetic resonance imaging can produce reliable 、 The measurement of individual brain function , And inform clinical practice accurately enough ? Next , We will focus on four complementary strategies , These four strategies are designed to answer this question , And the psychometric principles behind these strategies .

chart 2 fMRI Research timeline

3.  Neuroscience to establish reliable individual differences

3.1 Improve reliability by extending aggregation

      BOLD The signal is from the heat 、 Physiology ( Like sports 、 breathing ) And other non physiological ( Such as scanner drift ) Surrounded by the noise of the source , Only for fMRI A small part of the data , about 5-20%.

      In order to solve the problem of BOLD The separation of variance subsets , More and more precision fMRI (pfMRI) The research adopts the reliable principle of classical testing theory : Everyone collects more data . As the evaluation length increases , Reliability tends to increase , Because random 、 The variability of unstructured error scores has more opportunities to offset itself .

      When that happens , Changes in real scores form a larger part of the measurement , This leads to higher reliability . for example , Single item measurements are usually dominated by noise and item specific variances , With the addition of additional items and the use of comprehensive scores for multiple items , These variances offset themselves . Usually , In psychometrics , Evaluation length refers to the number of items in the questionnaire or survey ; However , The same principle applies to fMRI Scan length . In various queues 、 Scanner and research design , It has been shown that fMRI The reliability of measurement increases with the increase of scanning length . To illustrate this point , Multiple scans of different days session When the data is combined , The improvement of reliability is particularly obvious , This shows that , Due to transient factors ( Like a day 、 Head position 、 Wakefulness and scanner effects ) And unnecessary differences , Often covered up in a short time , Stable individual differences hidden in single session scan measurements .pfMRI The study collected data from a small number of subjects over a long period of time ( chart 3), It is further proved that 、 Individual specific signals exist in the spatial organization, time structure and time scale of brain network . Besides ,pfMRI It is helpful to discover the new cortical network that was previously covered by the group average , And began to move towards clinical application , Used to guide transcranial magnetic stimulation , To detect the recovery of traumatic brain injury , And the measurement of individual specific cortical reorganization .

      pfMRI Our new findings show that , If there's enough data , A particular functional organization may be a rule , Not the exception . This shows that the traditional short fMRI The use of scanning and population averaging has hampered many translational neuroscience efforts , Because individual differences are submerged in an ocean of unknown changes and noise . However , current pfMRI The method requires several hours of data from each individual to achieve a high level of accuracy . Many anatomical targets in translational neuroscience ( Like the amygdala 、 Nucleus accumbens 、 Orbitofrontal cortex ) in , The signal dropout Mixing is unreliable , More data is needed . therefore , For most developmental and clinical samples ,pfMRI The burden on participants is high , For them , Lying motionless in the scanner for hours is particularly challenging . therefore , In the present form ,pfMRI It has not been widely studied in the field of population neuroscience , This is for the realization of fMRI Broad transformational value is essential . However , up to now ,pfMRI By using a relatively crude method , Allow unstructured variances to offset over time , Thus, higher reliability is realized to a great extent . So that we can understand the generation source of the real score change , We want to measure ( for example ,BOLD The signal ) And wrong score changes , We want to delete ( Noise ), Alternative strategies with shorter scans can more effectively achieve accurate and reliable measurements and reduce the burden on participants .

chart 3 pMRI Can reveal reliable 、 Individual specific brain functional characteristics

3.2 By modeling the stability variability , Can improve reliability

      The focus of translational neuroscience research is often on measuring stable disease risk 、 Biomarkers of status and prognosis . However , Many of the most widely used fMRI The modeling method uses a large number of fMRI The measurement is reduced to the average estimate of a single individual to mix stable and transient variability . in other words , Functional magnetic resonance imaging studies often reduce regional brain function to a single activation or functional connection estimation . for example , In task fMRI (fMRI) in , This is usually done by fitting a single regressor or by comparing interests , This comparison represents the control and experimental conditions ( for example ,BOXCAR Model ) The alternate structure of tasks . Similarly , Functional connectivity is usually estimated by integrating the whole fMRI Associated with scanned activities .

      These modeling methods were originally designed for experimental cognitive neuroscience , among , The variance between subjects is the source of error , Need to minimize , To maximize the estimation of the experimental effect within the subjects and the group average . However , Since each individual has only one estimate ( for example , Task comparison beta Or edge function connection ), The stability of the 、 The individual specific variance cannot be correlated with the instantaneous source of variance within the subject ( for example , thought 、 Fluctuations in emotional state or attention ) Separate from noise .

      Recent research shows that , By using tools designed for repeated measurements ( Such as latent variables and hierarchical Bayesian models ) Explicitly isolate the stable variance , It can greatly improve the reliability of task activation and function connection measurement . The key is , These modeling methods can be applied to multiple scans of each individual , It can also be used when there is only one scan . This is because fMRI scans essentially contain many estimates of brain activity or connections . for example , Multiple activation estimation can be achieved by fitting regression variables to fMRI Scan the first half and the second half to generate ( namely , Segmentation analysis ), perhaps , At a more granular level , By fitting independent regressors to each test in the scan . Similarly , Multiple functional connectivity estimates can be made by dividing a single scan into two halves , Thus, two functional connectivity estimates are generated , perhaps , In extreme cases , Through to each fMRI Volume or data points produce covariance estimates . Once multiple estimates are generated for each individual , You can use the tool of repeated measurement modeling to fMRI The stable components of variance are separated from instantaneous variance and noise . in general , This modeling has been found to improve the reliability of activation and functional connectivity measurements , Especially from the short fMRI Scanning , Can be improved by up to 60% The reliability of the . Besides , These stable components showed higher heritability and greater behavioral Association , Further improve the transformation value .

      These methods illustrate a measurement principle that may seem counterintuitive : take fMRI The data is decomposed into multiple noise estimates , More reliable measurement results can be generated at the level of potential variables , Instead of simply aggregating across components .

      Besides , This solution and the recent structure MRI Found consistent , Many times 、 Fast 、 Low resolution scanning can produce better results than single 、 Longer 、 Higher resolution scans for more accurate brain structure estimation ( Such as the thickness of cortex ). However , It should be noted that , This kind of reliability modeling is not omnipotent , Nor can it replace careful measurement . In its simplest form , This modeling will absorb all forms of stable variance through design , Including stable artifacts , Such as head movement 、 Respiratory and vascular dynamics . therefore , In the modeling process , The extent to which these physiological illusions are not completely removed , They are also absorbed by stable ingredients , And continue to undermine the effectiveness of brain behavioral connections .

3.3 Stability can be improved by removing physiological artifacts

      Not bold Sources of variability , Such as head movement , It is usually a stable characteristic of an individual . therefore , These sources are not necessarily removed by aggregation or latent variable modeling , Because their variances are nonrandom , And unconsciously simulates the individual differences of interest . Besides , The mainstream data processing technology can not completely remove these physiological artifacts .ME-fMRI It's a new kind of 、 Methods based on biophysical principles , It can be downloaded from fMRI Separating and removing noise and non noise from data bold Variance of source . To do this ,ME-fMRI Multiple whole brain images were acquired during each excitation pulse ( That is, multiple echoes ), Instead of a single image that is normally captured ( That is, a single echo ). This allows for the removal of many physiological artifacts , because BOLD The signal will decay with the echo , Instead of BOLD Artifacts and noise do not .

      However ,ME-fMRI Prudent data cleansing practices are still needed , Because in the non nerve BOLD effect ( for example , Breathing patterns throughout the scan ) Moderate and stable individual differences will still interfere with the study of individual differences .

3.4 Reliability can be improved by designing stimuli to induce individual differences

      As mentioned earlier , The vast majority of fMRI tasks are designed for experimental manipulation within subjects , Group average effect , Rather than best evoking individual differences in brain function between subjects . therefore , Improve fMRI Another strategy for measuring individual differences is , Design new tasks from scratch , The clear objective is to optimize reliability and accuracy . especially , Using psychological measurement tools from item response theory and general theory , Choose stimuli based on their ability to evoke reliable individual differences , This is largely an untapped opportunity . without doubt , Building new tasks and stimuli from scratch takes a lot of time and is expensive fMRI Pilot study , To assess a large number of stimuli and task items , Test their psychological characteristics , And repeatedly choose stimulation , Effectively produce the most accurate 、 Reliable measurement . However , Relevant research shows that , Proven with known measurement characteristics fMRI Stimulation can produce great benefits , Including the ability to create more complete models , To study how the individual brain processes feelings 、 Memory and language information . Besides , Preliminary evidence shows that , A small part fMRI The point in time drives reliable individual differences in functional connectivity disproportionately . Because these highly reliable time points are often caused by the same movie clips in the individual , By selecting the stimulus that most effectively evokes individual differences in brain function , It is possible to gain greater efficiency gains .

      The stimulation of naturalism , Like a movie 、 Speeches and complex social scenes , There may be special benefits in this regard , Because they keep participants engaged 、 Awake and relatively static , So as to reduce the head movement 、 The artifact of attention and lucidity . Naturalistic stimuli also tend to have higher ecological validity than traditional tasks , Can easily target a wide variety of content , Including vision 、 Emotional and social characteristics , The goal is the psychological structure of interest . However , These benefits come with tradeoffs . for example , The complexity of naturalistic stimuli can not be easily controlled to the level of traditional cognitive neuroscience stimuli ( for example , Color composition , Space frequency ).

4.   summary

      The strategy we emphasize here is not exhaustive or prescriptive . contrary , By emphasizing these strategies and their basis on psychometric principles , We hope to promote the design of fMRI research , It can better produce reliable measurement methods of individual differences in brain function . Use fMRI Translational neuroscience cannot be the secondary goal of experimental cognitive neuroscience , It's something that needs to be iterated 、 Explicit development to optimize reliable measurement of individual specific variation . Despite some recent setbacks , But now that we know the present better than ever fMRI Limitations of measurement , And there are new strategies to build more precision 、 Reliable measurement methods , We see a bright future for the cumulative transformation of individual differences into neuroscience .