Yale Psychiatry Grand Rounds: October 30, 2020

October 30, 2020

Information

Sarah Yip, PhD, MSc, Assistant Professor of Psychiatry, Yale School of Medicine: "Connectome-based Prediction of Abstinence Across Drugs and Brain States"

ID5829

To CiteDCA Citation Guide

00:00Before we get started,
00:01I'd like to say a few introductory remarks.
00:061st, just a reminder to keep your
00:08microphones on mute during the talk
00:09and during the discussion less you
00:11are speaking during that discussion.
00:13Also, if you'd like to ask questions,
00:15please hold them to the end of the
00:17talk or place them in the chat.
00:19In doing so, I'd like to encourage
00:21everyone to maintain a respectful stance,
00:23nuances, sometimes hard to discern
00:25in the format of a zoom call,
00:27and so it pays to be particularly
00:30thoughtful about this. Next,
00:32for those of you who are joining
00:34us from other institutions,
00:36a special welcome like attendees from Yale.
00:38If you'd like to receive CME credit,
00:41you may Trisha Doll will post information
00:44in the chat about how to sign up.
00:47She'll also post information
00:48about how to receive credit for
00:50participation in today's lecture.
00:53Finally, I'd like to announce that
00:55the Grand Round speaker next week
00:57will be Dean moves from Caltech,
00:59who will be speaking on space, time and fear.
01:02Survival decisions along defensive circuits,
01:04so I encourage you to attend
01:06next week as well.
01:07So with these introductory remarks behind us,
01:09I'm pleased to introduce our
01:11Grand round speaker Doctor Sarah.
01:12Yep.
01:13Doctor Yip received her PhD in
01:16psychiatry from the University of Oxford
01:19and in the United Kingdom in 2013.
01:23Subsequently,
01:23she came to Yale to complete a
01:25two year fellowship in Addiction
01:27Psychiatry within the Division of
01:29Substance Abuse in our Department.
01:31She was then promoted to Associate
01:33research Scientist and subsequently
01:35to assistant professor in 2016,
01:37and she holds a joint appointment
01:39in the Child Study Center.
01:41Ann in psychiatry.
01:43While at Yale,
01:44she is used Nuro psychiatric
01:46research methods to identify the
01:48biological mechanisms of addictions
01:50and their treatments.
01:51To support her efforts,
01:53she's been remarkably successful
01:54in obtaining external funding,
01:56including from the UK's Medical Research
01:59Council, Beijing Normal University,
02:00the National Center for Addiction,
02:02Substance Abuse,
02:03the Brain and Behavior Research Foundation,
02:05Naida, and, most recently,
02:07in IEEE.
02:09In this work,
02:10she's applied machine learning
02:12approaches to identify predictive
02:14neural mark rows of cocaine and
02:16opioid use as well as collection of
02:18neuroimaging data from individuals
02:20receiving different forms of medication
02:22assisted treatments for opioid
02:24use disorder or pharmacological
02:25challenges with medications,
02:27and most recently she's received
02:28funding from an I AAA to include
02:31predictive modeling of neuroimaging
02:33data from three large developmental
02:35cohorts to identify neural
02:36markers of alcohol initiation
02:38and misuse. Promises to be a
02:41very important work.
02:42Of note, in 2019 her work on
02:45connectome based modeling to predict
02:48treatment response in cocaine use
02:50disorder was cited as a top as the
02:54top basic science achievement of 2019
02:56by night as director Norvo Cough.
02:59Today we'll have the opportunity
03:01to learn more about this important
03:03work in her talk entitled Connectome
03:06based prediction of absence across
03:08drugs and brain states welcome Sarah.
03:12Thank you Stephanie,
03:13for that great introduction.
03:14It was one of the top basic
03:16science achievements, not the talk.
03:20I'm happy to provide support.
03:22That's our perspective.
03:23OK, let me just get my screen sharing.
03:33Can anyone see my slides?
03:37Yep yeah great OK great.
03:38So yes, Stephanie said I'm going to
03:41be presenting today on connectome
03:43based prediction of abstinence
03:44across drugs in Britain states.
03:46So this is work that my lab has
03:48been doing for a number of years
03:49to try to identify brain based
03:51predictors of treatment outcomes.
03:54I'm sure everyone here is aware of the
03:57current substances epidemic in this country,
03:58but I think it's nonetheless important to
04:00mention so this color map shows population
04:02level deaths per 100,000 individuals in the
04:05United States back in 2014, and as we know,
04:07these rates have continued to rise.
04:09And so, for example,
04:11annual opiate associated fatalities
04:12have exceeded those calls by firearms
04:14and motor vehicles combined,
04:15as well as those caused by HIV at
04:17the height of the AIDS epidemic.
04:19And although it's been much less publicized,
04:22there has been this concurrent rise and
04:23cocaine and stimulant associated fatalities.
04:25Therefore, we really do need improve
04:27strategies to combat the current
04:29substance use epidemic in this country,
04:31which is really the motivation
04:33for my labs or imaging work.
04:35So I often get asked why we would
04:37even using our imaging to predict
04:39treatment outcomes in the 1st place.
04:41After all,
04:42we're pretty fortunate in the addiction
04:43field with a large number of really
04:45excellent evidence based treatments that
04:47are very effective for some individuals.
04:49However,
04:49we also have a lot of between
04:51patient heterogeneity,
04:52which unfortunately means that the overall
04:53efficacy of any given treatment tends to
04:55be highly variable across individuals.
04:57Given this heterogeneity,
04:58it's perhaps not surprising that
04:59even gold standard treatments and
05:01have a high rates of relapse and that
05:03most individuals with a substance
05:04use disorder go through multiple
05:06failed treatment attempts.
05:07In other words,
05:08when it comes to addiction treatment,
05:10one treatment really doesn't
05:12fit all relapse rates.
05:13Following treatment also remain high,
05:15and for some substances this
05:16is a critical vulnerability.
05:17Vulnerability period for
05:19overdose associated death.
05:20And this problem is really further
05:22exacerbated by the fact that what I like
05:24to call traditional clinical variables.
05:26So for example,
05:27things like baseline severity don't tend
05:28to be all that helpful for predicting
05:30treatment response for relapse rates,
05:32meaning that we have unexplained
05:33sources of her originality
05:35influencing treatment outcomes.
05:36So within this context,
05:37a number of research groups have
05:38turned in our imaging measures
05:40to try to identify brain based
05:41predictors of treatment response.
05:45And so over the past 10 to 15 years,
05:48there's been a number of proof of
05:51concept studies looking brain,
05:52function and structure to
05:53treatment outcomes and addictions.
05:55So, for example,
05:56work by our group and others
05:57indicate that individual differences
05:59in pretreatment reward related
06:01activations are prospectively linked
06:03to within treatment abstinence.
06:05That changes in reward related activations,
06:07Ann May Co occur with reductions
06:09in substances and that individual
06:11differences in functional connectivity
06:12relate to both current cocaine use,
06:15an feature relapse to cocaine.
06:18So collectively,
06:18these and other studies are really
06:20supporting this overarching hypothesis,
06:21that individual differences in some
06:23aspects of the brain are indeed linked
06:26to differences in treatment outcomes.
06:28However,
06:28this work also has some limitations,
06:30so in particular most prior studies
06:32have relied on prospective associations
06:34and have used methods such as
06:36correlation and regression that
06:37have a tendency to overfit the data,
06:39leading to inflated effect size estimates.
06:41And so the problem with using the term
06:43prediction in this context in the
06:45context of symbol correlation or regression,
06:47is that true prediction requires
06:49application of a model to novel data.
06:52Machine learning approaches seek to
06:54produce overfitting via creation of
06:56a predictive model in a data driven
06:58manner manner using the training data set.
07:00And then application of that model
07:02to an independent test data set which
07:04is used for model validation and
07:05that distinction is really important
07:07because whereas the aim of traditional
07:09statistical approaches is to explain
07:11the relationship between two variables,
07:13the aim of machine learning approaches is
07:15to generate predictions in novel data,
07:17and I would argue that this type of
07:19approach is critical for the eventual
07:21translation of research findings
07:22into clinical settings,
07:24which is really a primary
07:25challenge of modern psychiatry.
07:27In addition, if used correctly,
07:28these approaches can be used
07:30for normal biological discovery.
07:31Which to me is also really important.
07:34So in my opinion,
07:35in addition to generating predictions
07:37in novel data,
07:38a primary goal of brain based
07:40clinical modeling should really be
07:42elucidation of mechanism.
07:43However, as we know,
07:44even highly predictive models
07:46often can be very little to enhance
07:49our mechanistic understanding.
07:50So throughout this talk,
07:51I'm really going to focus on maximizing
07:54nor biological discovery within the
07:56context of predictive modeling.
07:57In order to maximize this type of discovery,
08:00we use data driven whole brain methods.
08:02The reason for this is that given
08:03that recovery from addiction involves
08:05complex interactions across across
08:06clinical and biological domains,
08:08we think it's likely that it's even
08:09of abstinence also involves distributed
08:11processes for multiple brain regions.
08:13Therefore,
08:13all of the data and we showed you
08:15today is using a whole brain
08:17connectivity based approach,
08:18in which we're focusing on patterns
08:20of functional activity across the
08:22entire brain.
08:23The specific method I'm doing focusing
08:24on throughout this talk is referred
08:26to as connectome based predicted modeling,
08:28and unlike some other machine
08:30learning approaches,
08:30this is an entirely data driven technique.
08:32It doesn't require any a priori
08:34specification of networks or
08:35regions of interest.
08:36Therefore it's not only a predictive tool,
08:38but it's also a method
08:41of identifying networks.
08:42And so, prior to the data will be presenting.
08:45Today, connectome based modeling had
08:46been used to generate robust predictive
08:48models of measures such as IQ and attention,
08:51but it had never been applied to predict
08:54future behavior or clinical outcome.
08:57When we use a connectivity based approach,
08:58what we're doing is we're extracting time
09:00courses activity from multiple regions.
09:02It together encompass the entire
09:03brain to represented here by these
09:05cards in green and red voxels,
09:06and we simply correlate the patterns
09:08of activity across the time course of
09:10the whole scan to get a single summary
09:11statistic that summarizes the temporal
09:13coherence between those two brain regions.
09:15And we can do this for every single
09:17voxel in the brain to create a
09:19functional connectivity matrix or
09:21what we refer to as a connectome.
09:22So this single matrix Now summarizes
09:24whole brain connectivity patterns
09:26for all of the voxels in the brain.
09:27Or a single person,
09:28and there's data to indicate that
09:30these connectomes are both relatively
09:32unique to the individual,
09:33but also the patterns of connectivity
09:35within the connectome.
09:36Forgiven individual may vary as a
09:39function of task performance or what
09:41we like to refer to his brain state.
09:44The consistent with this notion of
09:45brain stay work by colleagues here
09:47at Yale has demonstrated that the
09:49accuracy of predictive models generated
09:50from connectivity data is improved
09:52when the input is connectivity data,
09:54computed joint task performance
09:56as opposed to join resting state.
09:57So here on your left you have the percent.
10:00Variance in IQ is explained by a
10:02predictive model built using different
10:03tasks from the Human Connectome Project.
10:05An on your right you have the
10:06same thing for the Philadelphia
10:08in or developmental Gore,
10:09and what I want to point out here
10:11is it in both cohorts the predictive
10:13accuracy of models built using
10:14resting state data is relatively low,
10:16so in both cases the model is only able to
10:19account for about 4% of the variance in IQ.
10:21The second thing I want to highlight
10:23is that the predictive accuracy of
10:25models generated from different types
10:26of task data is also highly variable,
10:28with some tasks accounting for
10:29up to 12% of the variance in IQ.
10:32Another accounting for less,
10:33so together this is suggesting that
10:34specific brain statement elations mean
10:36preferable for predicting a given behavior,
10:38which is an issue that I'm going to
10:40return to you throughout this talk,
10:42which brings me to the first bit of
10:44data will be presenting today in which
10:46we're using a connected based approach
10:48to try to put it to the clinical outcome.
10:50In this case,
10:51abstinence from cocaine during
10:52a 12 week treatment.
10:54The study design for this
10:55is relatively simple.
10:56We recruited patients and during
10:57a 12 treatment trial and scan them
10:59at the start of treatment and
11:00then again following treatment.
11:02All of our participants receiving
11:03methadone maintenance therapy for opiate
11:05use disorder, but we're now entering
11:07treatment for cocaine use disorder,
11:08so these are polysubstance using individuals.
11:11I'm not going to talk about the specific
11:13aspects of the 12 week trial today,
11:16but if you're interested,
11:17this was an RCT of behavioral therapy,
11:19with or without treatment with a
11:21cholinesterase inhibitor lanta mean,
11:22so the results of that are
11:24published in the Journal.
11:25Kinda looks like hydrate and
11:27consistent with the overall arc TR,
11:29nor imaging subsample was dominantly male,
11:30there mostly unemployed and the primary route
11:33of cocaine administration was via smoking.
11:35They also did a number of
11:37prior treatment attempts,
11:38including an average of three or more
11:39prior failed inpatient treatment attempts,
11:41and three or more for three or more prior
11:43failed outpatient treatment attempts,
11:45and they also had significant legal problems,
11:47so this is a very treatment
11:49refractory population, but it's also,
11:50you know, kind of a.
11:53It's not an unusual population.
11:56I'm not really going to drill down too
11:58much into the methods of CPM today,
12:00but I'm happy to answer questions
12:02if people are interested,
12:03but I do just want to give you sort
12:05of a general overview of the approach.
12:07So practically when we're doing
12:08a connecting based model,
12:10what we're doing is we're taking
12:11individual connectomes from a training
12:13data set and relating them to a
12:14behavioral variable of interest.
12:15So in this case that within treatment,
12:17abstinence and we do this using simple
12:19correlation in order to identify
12:20positive productive connections,
12:21or what we refer to as edges,
12:23as indicated by these red squares
12:24here and then we also identifying
12:26negative prediction.
12:27Productive connections as indicated
12:29by the green squares.
12:30So positive predictive connections
12:32are connections for which increased
12:34connectivity predicts absence,
12:35whereas negative predictive connections
12:37or connections for which decreased
12:39connectivity predicts abstinence.
12:40So this is our feature selection phase.
12:43We then create individual participant
12:44summary scores via just summing
12:46the edge weights identified in our
12:48feature selection phase so that each
12:49participant now just has two values.
12:51One is that positive summaries for
12:52another is a negative summers,
12:54or we then create our brain behavior
12:55model by entering these summary
12:57scores into predictive models.
12:58In this case, assuming linear relationships,
13:00so a simple Y equals MX plus B and
13:02then finally we apply this model to our
13:04testing data so we take connectivity
13:06matrices from the independent data set.
13:08We extract connectivity values from the
13:10edges we identified in our training data,
13:11set some them to create.
13:13Summary scores for new participants
13:15and enter those values into our brain
13:17behavior model to generate individual
13:19predictions of within treatment abstinence.
13:21And finally,
13:21although it's not shown here,
13:23we evaluate the performance of our
13:25model by comparing those predicted
13:26abstinence values with the actual
13:28within treatment abstinence values.
13:30And so we can compare actual
13:31predicted values.
13:32A number of different ways,
13:33which is an issue.
13:34I'm not really going to go into in this talk,
13:36but really the simplest way
13:38is just brain correlation.
13:39So that's what I'm going to be
13:41using to assess model performance
13:42throughout the talk today.
13:44For this study, with a couple of different
13:46types of functional data to choose from.
13:48So during pretreatment scanning,
13:49everyone perform two tasks.
13:50One was a basic cognitive control task
13:52and the other one was a classic reward
13:55task and monetary incentive void task.
13:57And so, given that we were
13:59interested in cocaine use.
14:00We chose to either word
14:02task because at that time,
14:03at least to my mind,
14:05the data linking cocaine use to
14:06reward was stronger than the data
14:08linking it to cognitive control,
14:10and so using the data acquired
14:11Dreamer Award task,
14:12we compute individual participant
14:14connectivity matrices and we feed this
14:15into our connectome based model along
14:17with our measure of instrument abstinence,
14:19which was defined as the percentage
14:21of BI weekly urine specimens
14:22that were negative for cocaine
14:24during the 12 week treatment.
14:26So is the biologically verified
14:29dimensional measure of absence?
14:31So here we have our initial model
14:33results here on the Y axis we have
14:35individual participant abstinence values
14:37as predicted by our brain behavior
14:38model and on the X axis we have actual
14:41absolute values for each participant.
14:43And so typically when we use correlation,
14:45we're trying to explain the
14:46variance between two variables.
14:47But here we're just using it to
14:50characterize predictive accuracy
14:51or the correspondence between
14:52actual and predicted values.
14:54And so you can see that our model
14:56has relatively good predictive
14:57accuracy with this German 0.42,
14:59which means that about 20% of the
15:01variance in within treatment abstinence
15:02is accounted for by connectivity
15:04within our abstinence networks,
15:05and interesting Lee,
15:06this correspondence is greater
15:08than that absorbed.
15:09If we just relate absence
15:11to another level variable,
15:12for example related to baseline cocaine use,
15:14indicating that are connected based
15:16model has greater predictive accuracy
15:18than traditional clinical variables.
15:19So in addition to generating predictions,
15:21this approach also,
15:22of course,
15:23identifies networks just as a heads up.
15:26Networks identified using whole
15:27connectome based approaches are
15:29typically complex and can be
15:30composed of multiple adjacent
15:32and non Jason's brain regions.
15:33As was certainly the case here.
15:35So here on your left you can see the
15:38positive network shown in red as a reminder.
15:41The positive network includes connections
15:43which increased connectivity predicts
15:45absence and on the right you can see
15:47the negative network which corresponds
15:49to connections which decreased
15:51connectivity predicts abstinence.
15:52And so while these networks are
15:54certainly complex and arguably
15:55pretty hard to interpret when
15:56presented in this fashion,
15:58it's important to point out that actually
16:00the combined number of connections
16:01within the positive and the negative
16:03network together is only around 500,
16:05which is actually less than than
16:062% of all possible connections.
16:09In other words,
16:10despite this visual complexity,
16:12these are actually quite
16:13specific connections.
16:17So there are a number of standard
16:19ways in which we can now start
16:21to summarize these connections
16:22to facilitate interpretation.
16:23So one simple way of characterizing
16:25our networks is to summarize
16:27them by a connection distance.
16:28So for each nodal connection we can
16:30compute Euclidean distance using
16:32central coordinates for each node.
16:33And if we apply this to
16:35our abstinence networks,
16:36what we find is that both networks include
16:38both short and long range connections.
16:40However, positive absence network,
16:41the network for which increased connectivity
16:43predicts more within treatments,
16:44is characterized by predominantly
16:46longer interactions.
16:47Where is our negative network,
16:48the network for which decreased connectivity
16:50predicts more within treatment.
16:52Abstinence is characterized by
16:53predominantly shorter range connections,
16:54which makes intuitive sense
16:56because longer range connections
16:57tend to be involved in the higher
16:59order more complex processes.
17:03Another way in which we can summarize
17:04these networks is of course,
17:06anatomically so these circle plots
17:07summarize network connectivity based on
17:09the number of connections between mappers.
17:10They'll brain regions which
17:11are listed on your left,
17:13so from the top of each circle brain brain
17:15regions are represented in grafana top,
17:17and as amical order,
17:18with lines coming from the
17:19red portions of the top,
17:21or responding to prefrontal cortical
17:22connections in lines coming from
17:24the purple bits at the bottom
17:26corresponding to brainstem connections.
17:27And from left to right,
17:28network connectivity is threshold
17:29with a different levels based on the
17:31total number of connections for brain,
17:33region and so you can see that if we
17:35use a liberal threshold, for example,
17:37we look at all brain regions with five
17:39or more connections showing here on the left.
17:41Our networks remains somewhat
17:42difficult to interpret.
17:43However,
17:43if we take a more conservative threshold,
17:45for example only focusing on regions with
17:4712 or more connections on the right.
17:49Our networks begin to be a bit
17:51more interpretable,
17:52so for example,
17:53we can see that the positive abstinence
17:55network shown in red at the top is
17:57characterized by right prefrontal node
17:58with connections to temporal limbic
18:00and left frontal cortical regions,
18:01and we can also see that the negative
18:04network here at the bottom is
18:06characterized by a temporal node with
18:08connections to limit and cerebellar nodes.
18:10Another way we can begin to understand
18:12the anatomy of our absence networks
18:14is via their spatial overlap
18:15with Canonical neural networks.
18:16So for example,
18:17via their overlap with the frontal
18:19parietal salients and default mode networks,
18:21all of which had previously
18:23been implicated in addictions.
18:25When we do this,
18:26we can characterize connectivity
18:27based on the number of connections
18:29between these established networks.
18:31So here we have matrices summarizing
18:32overlap between these networks for the
18:34positive and negative network separately,
18:36and for each matrix the cells represent
18:38the total number of connections
18:39within and between each network,
18:41with darker colors indicating a
18:42greater number of connections,
18:44and so we can see the positive
18:46network is characterized by a large
18:47number of frontal parietal to me
18:49and medial frontal connections,
18:50as well as by a lot of salience
18:53and motor sensory connections as
18:55indicated by these dark red boxes.
18:57We can also see that the negative network
18:59includes a large number of salience,
19:01default mode, and medial frontal connections,
19:03as indicated by the dark blue boxes,
19:05and these differences become even
19:06more apparent if we directly compare
19:08the number of connections within
19:10each network so we can see that the
19:12positive network includes more frontal
19:13parietal to medial frontal connections,
19:15as well as more salient subcortical
19:16motor sensory connections.
19:17In contrast,
19:18the negative network includes more
19:20connections between the medial
19:22frontal network and the salience
19:24in default mode connections.
19:25So based on these differences,
19:27we generated a theoretical network
19:28based model of cocaine abstinence.
19:30We propose the cocaine abstinence
19:32is positively predicted by increased
19:33connectivity within and between frontal
19:35parietal and medial frontal networks.
19:37As a reminder, these are networks
19:39involved in coordination of attention
19:40and executive control processes,
19:42and so we think they might contribute
19:44within two and abstinence via
19:45coordination of the top down process
19:47is necessary for treatment engagement.
19:50So for example,
19:50things like acquisition of new
19:52skills or enhanced control over
19:54impulsive behaviors or data further
19:55suggests that cocaine abstinence.
19:57Is positively predicted by
19:59increased connectivity.
20:00Within salient subcortical motor sensory
20:02regions and so these are networks
20:04involved in coordination of salience,
20:06encoding and reward.
20:07So we think that these networks
20:09could support motivational
20:10processes and relevant in treatment.
20:12For example,
20:12willingness to change in shoring up
20:14of non drug reward processing or
20:17attending double turn it rewards.
20:18Finally,
20:19our data indicate that absence
20:20is further predicted by decreased
20:22connectivity between these two systems
20:24and so based on prior connectivity
20:26working cocaine use disorder,
20:28we actually think that appropriate
20:29separation between these two systems.
20:31Could relate to greater behavioral
20:33flexibility or decreased compulsivity
20:34as might be required for behavior
20:36change during treatment,
20:37so this model is really building
20:38on prior models of addiction
20:40emphasizing separation of frontal,
20:41parietal and salience networks
20:42such as model put forth by Elliot
20:44Stein colleagues where we're also
20:46incorporating medial frontal motor,
20:47sensory and subcortical networks.
20:49Provide a theoretical framework
20:50for future research,
20:51and so our hope is that by
20:53proposing this framework,
20:54we can encourage others to test
20:56these hypothesis in their own data.
20:59So the last thing I want to point out
21:01on this slide is that so when we're
21:03when we're computing our connectivity,
21:05matrices were basically we're
21:07taking the time course across the
21:08entire rewards get broad tasks,
21:10so we're not modeling events
21:11of interest at all.
21:13We're just basing it on the
21:14overall pattern of connectivity
21:16across the entire task.
21:17But Despite that,
21:18I think that our findings are
21:19actually somewhat intuitive when we
21:21consider them within the specific
21:23context of reward task performance,
21:24which of course would require
21:26coordination of both attentional
21:27and cognitive control processes
21:28as well as assailants encoding
21:30and reward response behaviors.
21:31So, so again,
21:32this is just starting to hint at
21:34this possibility that Brain State
21:36might matter for predictive modeling.
21:38Also tested the ability of the
21:39identified networks for abstinence
21:40following treatment.
21:41So for this analysis we simply create
21:43summary scores by summing edge
21:45weights for the nodes identified
21:46in our original analysis.
21:48So we're taking our absence networks
21:49and we're using them as a mask to
21:52extract values from post treatment data,
21:54and we enter those into correlational
21:55analysis with percent days of absence.
21:57During six month follow up and we find
21:59that post treatment network strengths
22:00were in fact significantly correlated
22:02with posttreatment abstinence,
22:04suggesting relative consistency
22:05of this relationship overtime
22:06further consistent with that.
22:07When we just compare pre
22:09versus post treatment.
22:10Connectivity strength we see no changes,
22:12overtime raising the possibility
22:13that these effects are somewhat
22:16stable within individuals.
22:17But the big question is,
22:18is not really whether or not these effects
22:20are stable within individuals overtime,
22:22but whether or not this fact
22:23replicates an independent sample,
22:25because that's really the whole
22:26point of using something like
22:28a machine learning approach.
22:30So the analysis of and show
22:31you up to now were run using.
22:33They gotta leave one out cross
22:35validation scheme in which for each
22:37for a single participants predicted
22:38value or what we refer to as the
22:40left out participant is generated by
22:42taking the data from all of their
22:44participants in the training data
22:45set an iterative manner and so all
22:47participants have a predicted value.
22:49But that approach can also be
22:50prone to overfitting,
22:51so we wanted to also have an external
22:54validation sample which was no small
22:56feat because pretreatment Norman
22:57data is pretty hard to come by.
22:59However, things to Kathy, Carolyn,
23:01Mark, but Enza,
23:01we managed to cobble together
23:03a replication sample.
23:04I'm not going to go too much
23:05into the details of that today,
23:07but it was 45 individuals about
23:09a third of whom were taking
23:10methadone for opiate use or now
23:12seeking treatment for cocaine.
23:13As in our original sample,
23:14but 2/3 of the replication sample were
23:16not on methadone and work independently,
23:18and they're also scan prior to
23:20us totally separate treatment.
23:21Trial involving a different medication.
23:22So this is a very heterogeneous
23:24replication sample, was at least.
23:25For replication sample were again just
23:27creating summary scores by summing
23:29edge weights with nodes identified in
23:31our original analysis and entering
23:33them into correlation analysis.
23:34Their biological measure absence
23:35percent cocaine for yarns and in fact
23:38we find the same relationship between
23:40network strengthen within treatment.
23:41Abstinence in this heterogeneous
23:43independent sample.
23:44Just to summarize that first data set.
23:46These data demonstrate the ability of
23:48a recently developed machine learning
23:49approach for Dick Truman Outcomes.
23:50In this case,
23:51abstinence from cocaine germ
23:52phobia treatment.
23:53I didn't show you the data today,
23:55but we recently replicated this
23:56in a second external sample.
23:58Networks that we are identified with her
24:00bus and they were relatively unchanged,
24:02involved analysis,
24:02controlling for other
24:03baseline clinical variables.
24:04So for example,
24:05methadone dose or days of
24:07past month calcaneus.
24:08Post treatment connectivity within
24:10these networks also predicted
24:11absent certain follow up and our
24:12networks are not changed from
24:13pre to posttreatment raising the
24:14possibility that this relationship
24:15might be somewhat consistent overtime,
24:17which is a point that will come
24:19back to at the end of my talk.
24:21But first I want to return
24:23this issue of Brain State.
24:24And which brings me to the
24:26second part of my talk,
24:27in which we've begun looking into
24:29whether brain states are drug specific.
24:30So this is work that was recently
24:32published and look at their psychiatry and
24:34was led by Doctor Sarah Legiance Teen,
24:36a very talented former postdoc in
24:38my lab who was recently who recently
24:40received her appointment to assistant
24:42professor here in Psychiatry.
24:43So the clinical rationale for
24:45this work really came from
24:46some MA work done by Epstein
24:48and colleagues a few years ago,
24:50so this was an electronic
24:51diary study of 114 methadone
24:52treat individuals with cocaine,
24:54an opiate dependence so very
24:55similar to our Poly substance.
24:57Using our image example,
24:58and in this study what they did
25:00was ask participants to track
25:01changes in mood and craving,
25:03and to document their substance use
25:05within an animated framework and
25:06what they found was a differential
25:08Association between positive versus
25:09negative mood States and future
25:10substances such that cocaine use
25:12was most robustly associated
25:14with having been exposed to the
25:16drug or being in a positive mood.
25:18Suggesting possible links
25:20with impulsivity and reward,
25:21but heroin craving was associated
25:23with increases in negative affect,
25:25suggesting possible links with emotion
25:28regulation or inhibitory control.
25:30And so, based on this finding,
25:32the different mood states predict opiate
25:33versus cocaine use in the same individuals.
25:36We wanted to test the hypothesis
25:37that maybe different brain states
25:39might be more closely linked
25:40to opiate versus cocaine use.
25:42In our image example.
25:44So for this analysis,
25:46instead of focusing on cocaine use,
25:47we're focusing on opiate use.
25:49However,
25:49the primary simple is the same,
25:51so as I mentioned in the beginning
25:53of this talk,
25:54when I introduced our sample despite
25:55seeking treatment for cocaine use,
25:57all of our participants at Coco
25:59get use disorder.
26:00And we're currently methadone
26:01maintained and in fact it did turn
26:03out that there was a fair amount of
26:05residual to be used in this sample.
26:06If you're interested in the details on that,
26:08I'll refer you to the clinical
26:10paper led by Kathy Carroll,
26:11but just briefly so you can sort
26:13of understand that data.
26:14Here we have survival curves indicating
26:15the time to 1st submission of a
26:17non methadone will be a positive
26:18urine screen during this whole week
26:20treatment and we can see that on
26:22average about 50% of participants
26:23in this study submitted at least
26:24one non methadone opiate positive
26:26urine during treatment.
26:27This is also true of the newer
26:29imaging subsample,
26:30for which the mean percentage
26:32of opiate negative yarns meted
26:33during treatment was about 65%,
26:35indicating of course at about
26:3735% of specimens tested positive
26:38for non methadone opiates.
26:43For this analysis we again just calculate
26:45functional connectivity matrices,
26:45but this time we're doing using the
26:47data acquired during performance
26:48or cognitive control task and we
26:50entered those matrices into our
26:52print behavior model along with their
26:54biological measure of abstinence.
26:56And so this is the result of that model
26:58here on the Y access we again have
27:01individual participant absence values
27:02as predicted by the brain behavior model
27:04and on the X axis we have the actual
27:06abstinence values for each participant,
27:08again as a reminder.
27:09Typically when we use correlation,
27:10we're using it to try to explain
27:12the variance between two variables,
27:14But here we're just using the up to
27:15characterize predictive accuracy
27:16or the correspondence between
27:18actual encrypted values.
27:18And as you can see,
27:20our model has relatively good
27:21predictive accuracy with aspirin 0.34,
27:23which means at about 12% of the
27:25variance and within treatment
27:26opiate use is accounted for by
27:27connectivity within this network.
27:29So for Contacts that's comparable
27:30to the amount of variance explained
27:32by similar approaches that seek to
27:34predict rate like behaviors such as IQ.
27:36As their cocaine network that
27:38will be at network,
27:39we identified his complex that includes
27:41connections between multiple adjacent
27:42and non adjacent brain regions and in
27:44fact or opiate network is actually
27:45somewhat larger than the cocaine network.
27:47So with the positive and
27:48negative networks together,
27:49including just under 1000 edges.
27:52So that's almost twice the size
27:54of the cocaine network.
27:55However, it's still less than
27:573% of possible connections,
27:58so again, and despite the visual.
28:01Complexity to actually quite
28:02specific connections,
28:03you can understand the anatomy
28:04of her opiate network.
28:05We can again summarize it by
28:07overlap with Canonical networks.
28:08So for example by overlap of medial,
28:10frontal and default mode networks
28:11when we do this,
28:12we can see that at least relative
28:14to the cocaine network,
28:15the opiate network is somewhat more sparse,
28:17and it does not include any within network,
28:20medial, frontal,
28:20or default mode connections.
28:22So despite including more edges overall,
28:24that will be it.
28:25Network is distributed across fewer
28:28Canonical networks.
28:29Directly compare positive versus
28:30negative networks.
28:31We find that positive network we
28:32follow the positive network included
28:34relatively more within network
28:35connections in the motor sensory network,
28:37whereas the negative network was
28:39characterized by more connections
28:41between the Motors and sorry
28:42network and rental prior to default
28:44mode and medial frontal networks.
28:46I have to say we were somewhat
28:47surprised that a large number of
28:49Motors motor sensor connections here.
28:51So what we've done is it also
28:52taken a virtual lesion approach
28:54with these data in which we just
28:56knock out all the nodes overlap
28:58with the given conical network,
28:59and we find that when we do this,
29:01despite the ceilings in child's
29:03motor sensory component,
29:03if we completely remove it,
29:05remaining connections still are
29:06sufficient in product abstinence.
29:07And actually we find the same
29:08thing if we if we knockout other
29:10individual clinical networks.
29:12For example,
29:12if we just knock out the default mode
29:14connections, really indicating that those.
29:16Single Canonical network alone is
29:18required to support abstinence.
29:20Salty understanding we again proposed
29:22a theoretical model that summarizes
29:24key aspects of this network,
29:25so this figure just emphasizes
29:27the absence was associated with
29:28increased within network motor sensor
29:30connectivity and increased between
29:32network connectivity of motor sensory
29:34and salience networks and of default
29:36mode and frontal parietal networks
29:38as indicated by the red lines here.
29:40And models also highlighting the absence,
29:42was further associated with decreased
29:44connectivity between the motor sensory
29:45network and medial frontal default
29:47mode and frontal parietal networks,
29:48as indicated by the blue lines,
29:50and again,
29:51our hope is that by summarizing the data
29:53in this somewhat simplistic manner,
29:55we can encourage others to test these
29:57theories in their own datasets to
29:59further guide nor biological risk.
30:01We've also made all of the masks of
30:03our actual absence networks public,
30:05and along with the associated code.
30:08Because we're interested in mechanism,
30:09of course,
30:10very interested in anatomical overlap
30:12between the cocaine and obeah networks.
30:14However,
30:14Interestingly,
30:15only compare edges across networks,
30:16we actually saw very little overlap,
30:18so less than 1% of edges shared.
30:20So here on your left we have edges their
30:22common to both cocaine and opiate networks,
30:25with the red lines indicating shared
30:26positive edges and the blue lines
30:28indicating shared negative edges.
30:29And as you can see,
30:31there's only a total of 8 shared edges.
30:33Overall,
30:34I realized I left off the figure legend here.
30:36So just for reference,
30:37the positive shared edges include
30:39a prefrontal, prefrontal,
30:40and limbic connections,
30:40as well as a subcortical to parietal
30:42connection and negative straight
30:44edges include parietal, still in bed,
30:46and supportable connections.
30:47In addition,
30:48we also identified several edges that
30:49have opposite opposite associations
30:51with opiate versus cocaine use,
30:53so these are edges for which,
30:55for example,
30:55increased connectivity is a positive
30:57predictor of cocaine abstinence,
30:58but for which decreased connectivity
31:00is a positive character.
31:01Opiate abstinence,
31:02or vice versa.
31:04Cities opposing edges are including
31:06connection between prefrontal
31:07and cerebellar regions as well as
31:09the temporal and vital cortices.
31:10And as you can see,
31:11there are more opposing edges
31:13than consistent edges.
31:14So these data together really
31:15indicating that the neural
31:16substrates of cocaine opiate use
31:18disorder may be largely disposable.
31:25We've also been looking into the
31:26specificity of these networks.
31:27Are predicting specific drugs
31:29across different brain states,
31:30so the data I just showed you
31:32indicated the cocaine obit network
31:33have pretty limited anatomical overlap.
31:35However one network was
31:36driver from reward task data,
31:38another was drive from cognitive data
31:39and I haven't yet shown you whether
31:41the opiate network also related to
31:43cocaine abstinence or vice versa,
31:45and I also haven't yet presented data
31:47to determine whether the relationship
31:48between network connectivity and
31:50substances is in fact task dependent.
31:52Or in other words,
31:53whether or not connect me findings
31:55hold across brain states.
31:56So to answer this question,
31:58first we looked at the impact of
31:59brain state on network identification.
32:01So we repeat our connectome based
32:03model of opiate abstinence using
32:05reward instead of cognitive task data,
32:06and we find that as I just showed you,
32:09we are able to predict opiate abstinence
32:11using cognitive control house data.
32:13However,
32:13we're not able to predict opiate
32:15abstinence using reward task data.
32:16Similarly,
32:17when we repeat our model of cocaine
32:19abstinence using cognitive data
32:20instead of reward task data.
32:22We find that we're only able to
32:24identify a cocaine network using
32:25the reward task data,
32:27demonstrating the identification
32:28of both obit cocaine abstinence
32:30networks was brain state specific.
32:31So having established that next we
32:33wanted to test whether once identified
32:36relationships between networks and
32:38specific substances might hold or
32:40generalize across brain states.
32:42To her on the left,
32:43with the Association between
32:44the in treatment,
32:45opiate abstinence and connectivity
32:47within the cocaine network,
32:48and as you can see,
32:49there's no relationship and on the right
32:51we have the relationship between cocaine,
32:53network connectivity and abstinence
32:54across different brain states.
32:55So for these analysis we're taking
32:57the cocaine cocaine network that were
32:59identified using reward task performance,
33:00using it as a mask to extract
33:02connectivity during cognitive task
33:04performance and also during resting
33:05state and what we find is that we
33:07again see a modest linear relationship
33:09between network connectivity and
33:10within treatment cocaine abstinence.
33:12So together these data indicate that all
33:14cocaine network does not generalize.
33:16Predict opiate abstinence,
33:17the relationship between cocaine,
33:19network connectivity and cocaine abstinence
33:21does in fact generalize across brain states.
33:24We see a very similar
33:25pattern with opiate network,
33:27so on the left we have the
33:28Association between within treatment,
33:30cocaine,
33:30abstinence and connectivity,
33:31even though be it network.
33:32This time during cognitive task
33:34performance and you can see that
33:35there is no significant Association
33:37where we take the network that we
33:39identified using cognitive task data.
33:40Use it as a mask to extract
33:42connectivity during reward task
33:43performance or during resting state.
33:45We again see a positive Association between
33:47network connectivity and within treatment.
33:49Abstinence would be absence.
33:51Again,
33:51indicating that while the opiate network
33:53does not generalize predict cocaine,
33:55use the relationship between
33:57opioid network connectivity and
33:59opiate abstinence does generalize
34:01across brain states.
34:02So I don't want to overly
34:04believer this point,
34:04but this is just to show
34:06you that we've run
34:07all possible combinations of this.
34:08So, for example, we've looked at whether
34:10opiate network connectivity during resting
34:11state relates to cocaine abstinence,
34:13and in all cases the specificity
34:14of these effects remains.
34:15So. In other words,
34:16we're seeing a double dissociation
34:18such that the cocaine network is
34:19consistently unrelated opiate use,
34:20and vice versa.
34:23As of the cocaine data,
34:24we've looked into the relationship
34:26between connectivity within the opiate
34:28network at post treatment and subsequent
34:30opiate use at 6 month follow up and
34:32we again find a similar relationship
34:33between post treatment activity
34:35and abstinence following treatment,
34:36potentially indicating some consistently
34:38this relationship overtime.
34:39And when we compare connectivity
34:40within the opiate network from
34:42free to post treatment,
34:43we also don't see any significant changes,
34:44further suggesting stability
34:46of these effects.
34:47Which is interesting to me,
34:49touches on an important point,
34:51which is that networks contributing
34:52to treatment response may in fact
34:54be distinct from those that change
34:56with treatment or that are directly
34:58implicated in disease pathology.
35:00So for example,
35:01brain regions for Tim's treatment
35:03responses and other disorders such as
35:05depression often have limited overlap
35:07with regions consistently found to
35:09differentiate patients from controls.
35:11Another possibility is the brain
35:12regions that predict treatment
35:13outcomes may just be different from
35:15those that change with treatment,
35:16and so at first I know that
35:18sounds counter intuitive,
35:18but when you think about it
35:20in a clinical context,
35:21we know that factors that
35:22predict treatment response.
35:23So for example,
35:24motivation to change can be distinct
35:25from those that change with treatment
35:27such as the acquisition of new skills.
35:29Thus the same may be true
35:31for neural networks.
35:32Further,
35:32it's possible that changes with
35:34an absence networks may take
35:35time to emerge and may only be
35:37detectable months after treatment,
35:38and that theory is consistent with
35:40data demonstrating the abstinence
35:41rates continue to improve following
35:43some behavioral treatments
35:44for cocaine use disorder.
35:45Thus,
35:45it stands to reason that the
35:47same may be true for the brain.
35:49Neural change may just take time to emerge.
35:53And so we began trying to
35:55explore that first possibility.
35:56The possibility that our networks are not
35:58only predictive of future substance use,
35:59but that they also just may be altered
36:01relative to healthy control individuals,
36:03and therefore perhaps linked to
36:05addiction pathophysiology more generally.
36:07Theoretical basis for looking into
36:09comparisons with healthy controls
36:10comes from really nicely by Hugh
36:12Garavan and colleagues from a few
36:14years ago in which they propose
36:15a couple of different scenarios
36:16for what prolonged absence might
36:18look like at the brain level.
36:20So here at the top,
36:21they propose that one scenario would
36:23be that recovery from addiction could
36:24involve some sort of restoration of
36:26premorbid brain function in the middle,
36:28they propose that an alternative
36:30hypothesis could be that prolonged
36:31recovery may in fact require some
36:33sort of hyper functionality.
36:34Brain regions involved in absence
36:36maintenance to above the level
36:38observed in healthy controls.
36:39Or finally,
36:40a third option at the bottom is the
36:42individuals in recovery from addiction may
36:44continue to exhibit decreased function,
36:46thereby confirming vulnerability for relapse.
36:47So just unpack that a bit
36:49more based on this model.
36:50If we were to compare absence
36:52networks to healthy controls and
36:53we found that absent individuals
36:55had similar levels of network
36:56strength relative healthy controls,
36:58we might conclude that what we're seeing
37:00is a return to premorbid functioning.
37:03Alternatively,
37:03if we were to find the individuals
37:05who achieve abstinence have increased
37:06network strength relative to controls,
37:08we might conclude that we're
37:10seeing is an elevation of brain
37:12function or a hyper recovery.
37:13And finally,
37:14if we were to find the abstinent
37:16individuals at decreased network
37:17strength relative to controls,
37:18we might interpret that as indicating
37:21continued vulnerability for relapse.
37:22So to test these theories,
37:24we've computed connectomes from
37:25identical task data for 38 age
37:27and sex matched non substance
37:28using control participants.
37:29And we've compared this to our
37:31Poly substance using sample.
37:32So here I'm just showing your binarization
37:34of the data I present earlier.
37:36So we have cocaine network sent
37:37for individuals to achieve
37:39some abstinence from cocaine.
37:40Drug treatment on the left and network
37:42show and for individuals who did
37:44not achieve absence on the right.
37:46And when we add in our controls what
37:47we find is it non substance using
37:49individuals or actually intermediary
37:51between responders and nonresponders.
37:52Such that our treatment responders
37:54actually have increased network
37:56strength relative to controls and
37:57our non responders have decreased
37:59network strength relative to controls.
38:00We again seem very similar pattern
38:02when we look at our opiate network,
38:04we find that our control group
38:05is again intermediary between
38:06responders and nonresponders with
38:08responders have increased network
38:09strength relative to controls.
38:10Although the difference between
38:12the controls and the non
38:13responders here isn't significant.
38:15So for both the cocaine and opiate networks,
38:17we're seeing this powder in the treatment
38:19responders have greater network,
38:20strengthen control participants.
38:22Again,
38:22consistent with this notion of an
38:24elevation of function relative to
38:26controls or hyper functionality.
38:28Most recently we've been applying
38:29this same approach to try to identify
38:31predictors of cannabis use outcomes,
38:33which I think is a really important issue,
38:35particularly within the context of the
38:38changing legislation in this country.
38:40Started my talk.
38:41I really focused on cocaine and
38:43opiates and I referenced increased
38:44overdose kristallen treatment as a
38:46significant motivator for identifying
38:48brain based predictors outcomes.
38:50And so while cannabis of course
38:52poses much less of an overdose risk,
38:54assuming it hasn't been mixed with
38:56another illicit substance cannabis use,
38:57none of nonetheless remains
38:59extremely prevalent,
38:59so approximately 15% of you as adults
39:02reported past year use back in 2017,
39:04and that number is likely higher
39:05now given all the ongoing changes
39:07to legislation in this country.
39:09So not only is cannabis use
39:11potentially becoming more prevalent,
39:12it's also becoming stronger,
39:14so the blue line here corresponds to
39:16the proportion of THC in cannabis
39:18samples over the past 20 years or so.
39:20And the green line corresponds
39:22to the portion of CVD,
39:23and as you can see,
39:24cannabis is becoming significantly
39:26stronger or composed of a significantly
39:28higher ratio of THC relative to CD.
39:30For these analysis we're
39:31combining aggregate pretreatment
39:32data from 2 separate arctis,
39:34one led by Kathy Carroll,
39:35another led by Brian Killick,
39:37and by combining data from these two samples,
39:39we end up with Nora merging data
39:41that needs our quality control
39:42criteria for 58 individuals of
39:44primary cannabis use disorder.
39:48For these analysis we again have
39:50connectivity matrices that we generate
39:52from both cognitive control and reward
39:53task data which we enter interpretive
39:55model along with their dimensional
39:57biological measure of abstinence. Surprise.
40:00What we find is that we're able to
40:02generate accurate predictive models of
40:04cannabis abstinence using both data types,
40:06suggesting that both cognitive and
40:07reward related brain states are relevant
40:09for understanding cannabis abstinence.
40:11So here, on the left we have the
40:13correspondence between predicted
40:14and actual abstinence values for
40:15the cognitive tasks shown in Gray,
40:17and for the reward tasks shown in white.
40:20And what we can see is that for
40:22both types of data,
40:23the spearings was hovering just under .4,
40:25indicating that about 16% of the
40:27variance and within treatment cannabis
40:29abstinence can be accounted for by
40:31connectivity within these networks.
40:32As with the cocaine and opiate networks,
40:34we've also tested the generalizability
40:35of this effect to an independent
40:37sample of individuals.
40:38The primary cocaine use disorder,
40:39the results of which are shown on your right,
40:42and as you can see,
40:43we found that the cannabis network
40:45does not generalize further,
40:46indicating substance specificity of
40:48our different abstinence networks.
40:50We're only just starting to dig down
40:52into the anatomy of these network,
40:54but so far the anatomy of the networks
40:56identified during reward and cognitive
40:58tasks do have some similarities,
40:59so these cord plots are summarizing
41:01positive network connections for the
41:03network identified during cognitive
41:04task performance on your left and
41:06reward task performance on your right.
41:08So these plants are really similar
41:10to the other circle plots are
41:12presented previously,
41:13but here,
41:14instead of summarizing macroscale
41:15regional connectivity there,
41:16summarizing connectivity between
41:17Canonical networks and we can see that
41:20for both tasks the cannabis network
41:21is characterized by high degrees of
41:23connections between frontal parietal and
41:25motor sensory networks as indicated by
41:27these sort of blue to pink arcs here.
41:30But we also see some differences.
41:32For example,
41:33we see differences in patterns of
41:35connectivity related to salience
41:36and visual networks.
41:37These differences are more striking.
41:39When we compare the negative
41:40connections so connections to which
41:42decreased connectivity is positive,
41:43productive cannabis abstinence.
41:44So here we're seeing clear differences
41:46in patterns of connectivity between
41:48salience and motor sensory networks,
41:49characterized by these sort of
41:51phase to pink arcs.
41:53And we can also see that joint
41:55cognitive task performance.
41:56The model is identifying a number
41:58of connections between the default
41:59mode and frontal network as shown
42:01by the screens beige dark,
42:02which we're not seeing in the
42:04reward task model at all.
42:06So as I said, these are brand new data.
42:08We're still working on drilling down
42:09into the anatomy of these networks.
42:11A lot of the work with this stuff comes
42:13after you've already done your model.
42:15But really,
42:16these findings are just again
42:17highlighting this idea that brain
42:19State may be a significant factor
42:20for generation of optimal models.
42:22So even if we're trying to predict the
42:24same behavior in the same individuals,
42:25the connections that we can identify should
42:28be partially dependent on the brain state
42:30that participant was in during acquisition.
42:32That's all the day I wanted to show you
42:35today, which I think demonstrates with
42:37ability of our approach to generate
42:39specific externally valid predictions,
42:41but also to provide more biological insight.
42:44So hopefully we've demonstrated today
42:45is that networks identified using this
42:48approach are clinically relevant. That is,
42:50there able to predict treatment response.
42:52Is there also externally valid or able
42:54to generalize product specific behaviors
42:56and novel symbols and individuals?
42:58In addition, despite their complexity,
43:00these networks are in fact
43:01biologically meaningful,
43:02and that they're composed the specific
43:05connections observing specific behaviors.
43:07And finally, these networks are a bust.
43:09A common sources of variance,
43:10so they predict even after
43:12controlling for severity treatments.
43:13I mentor medication status.
43:16If you're interested in applying predictive
43:18modeling approaches your own data,
43:19I recommend checking out the August
43:21issue of biological psychiatry,
43:22CNN I,
43:22which is dedicated to data driven approaches.
43:24This includes our our review paper that
43:26covers recommendations for best practices
43:27in cross validated biomarker research
43:29within the specific context of addictions.
43:31However,
43:31I think that many of the issues that
43:33we cover really applied to clinical
43:34predictive modeling more generally,
43:36which is probably why I figure summarizing
43:38the work for this approach made the cover.
43:40So again,
43:41if you're interested in these approaches,
43:42I suggest that you check that out.
43:46Just briefly,
43:46I'm not going to go into all
43:48the recommendations here,
43:49but I do want to just highlight
43:51a few key points,
43:52one of which is careful consideration
43:54of the window of assessment.
43:56So while scanning prior to a clinical
43:57intervention is often considered desirable,
43:59it's important term it's getting during this
44:01time can introduce unnecessary confounds,
44:03such as effects of acute
44:04intoxication or withdrawal.
44:05In the case of addiction or mood,
44:07state or psychosis.
44:08In the case of other disorders.
44:10In addition,
44:11scanning prior to clinical intervention
44:13may not be feasible in this treatment.
44:15Initiation is delayed,
44:16which may place unnecessary burden
44:18on the patient.
44:19I also want to emphasize the importance
44:21of employing multiple performance
44:22metrics for quantifying model accuracy,
44:24which I didn't go into too much today.
44:27But again,
44:27it's interview paper and also
44:29important of setting realistic
44:30expectations for effect size
44:32estimates when you're talking about
44:34results from cross validated models.
44:36Finally,
44:36I want to emphasize the importance
44:38of conducting post hoc testing to
44:40provide elucidation of mechanism,
44:41as without this findings or models,
44:43even with high accuracy,
44:44can do very little to advance our
44:47understanding of the underlying neurobiology,
44:49which is essential for informing
44:51novel treatment development.
44:52So in this context,
44:53it's important to consider model findings
44:55across multiple levels of interpretation,
44:57with the most basic level being the
44:59connections themselves and the most
45:00abstract level being the overarching
45:02or biological model which can be
45:03used to guide treatment and act
45:05as a basis for further testing.
45:08Within the context of a region
45:09of interest of region
45:10of interest based approach,
45:12one simple method of determining
45:13significance of different features.
45:15So significance of individual regions
45:16of interest or networks is to just rerun
45:19the model excluding specific features.
45:20This is that sort of virtual
45:22lesion approach that I talked
45:23about earlier with Soviet network.
45:25And we do this in order to determine
45:28which features were necessary
45:29for optimal model performance.
45:31So for example, to ask the question,
45:34does a middle volume contribute
45:36to overall model performance?
45:38Similarly, we can rerun the model
45:39only including selected features,
45:41and that will enable determination of the
45:42relative weight of specific features.
45:44So that would answer the question
45:46what is the predictability of
45:48amygdala volume by itself,
45:49and so these are really simple steps,
45:51but people often skip over them,
45:53so this is just a reminder that elucidation
45:56should also be a goal of prediction.
45:58We have several ongoing
46:00projects to extend this work,
46:01including our one focusing on
46:02neural markers of opiate use
46:04during the postpartum period,
46:05among women receiving methadone,
46:06we also have a pilot project which
46:08we're using real time fMRI to try
46:10to directly target connectivity
46:11within that will be at network,
46:13so we're interested in whether these
46:15connections are in fact modifiable,
46:16and we also have some medication.
46:18Studies were looking into that too.
46:21A sort of separate side of my
46:23lab work is also developmental,
46:25so we've also been using a
46:26connectome based approach to identify
46:28developmental mechanisms initiation,
46:29which is an ongoing collaboration
46:31with Dustin and Godfrey.
46:32Pearls in here at Yale,
46:33but also with Mary Heitzig in Hugh Garavan.
46:37There's also other studies that
46:39we wanted were not funded for yet,
46:41so we just received a very promising
46:42initial score from another or one
46:44focusing on developmental mechanisms
46:46of initiation, but this time,
46:47using the ABCD data set and including
46:49considerations ivkova 19 related factors,
46:51and that's a collaboration with
46:52Danilo Stockem ago.
46:53Finally,
46:53we're very interested in extending
46:55our work across different scanners in
46:57clinical settings and also interested
46:58in looking into potential sex
47:00differences in it will be at Nora markers,
47:02and so we've recently applied for
47:04funding to pursue that works,
47:05so fingers crossed on that one.
47:08I'll end there, but I want to just
47:10thank all of my collaborators,
47:11and because of the coauthors on the
47:13data presented today, and also my Cal,
47:15my mentor is Kathy Carolyn,
47:16Mark, but Enzo, also,
47:17of course,
47:17when I think everyone in my lab
47:18and all of the funders who make
47:20this research possible and thanks,
47:22all of you for listening.
47:26Thank you so much Sarah.
47:28We can we have a little time
47:30if there are any questions.
47:33It was a really great talk and
47:35very sophisticated systematic
47:36work that you've been doing.
47:40Thank you.
47:50Chris, go ahead.
47:52Yes Sarah great great talk.
47:53Thank you in your quick summary
47:55of the predictive modeling.
47:57You emphasize that you're
47:58looking at linear relationships.
48:01And the computational
48:02accessory for that is clear,
48:04but it also may be limiting
48:06because you may have threshold
48:07or other non linear relationships
48:09that are equally important.
48:10Or more important is there is it possible
48:13to incorporate nonlinear relationships?
48:14Or does the parameter space just blow
48:16up so quickly that that can't be
48:18done? That's a great question.
48:20I believe that Dustin Shine
48:22Host is working on some of that.
48:24I mean, you could use a kernel or
48:26something like that. Absolutely.
48:28I mean there's no reason why you couldn't.
48:30I mean, again, I find the interpretation
48:32even just with the linear so hard
48:35if we're talking about mechanistic
48:37understanding, but absolutely,
48:38especially if you have data that aren't.
48:40Yeah, yes, that would be great.
48:42We should do that next.
48:45Well, it enter related question
48:47that motivated that question.
48:48Is your predictions the amount of
48:50variance that you're explaining? You're
48:51getting up to 16% or something that's
48:54clearly impressive and validates that
48:55you're looking at something real.
48:57But it's a long way from
48:59clinical utility, absolutely.
49:00And so I was wondering whether incorporating
49:02nonlinear relationships would be a
49:04way to bump that up more quickly,
49:06but I wonder what other thoughts
49:08you have about what the future steps
49:11in this line of research might be.
49:13That will get us to. 85%.
49:17Whatever the threshold may
49:19may need to be for this to become clinical,
49:22clinically actionable?
49:22Yeah, absolutely, that's a great question.
49:24I mean, I think. You know,
49:26so we've done some stuff where
49:28we've been for the cocaine data in
49:30the external replication example,
49:32where we combine in other clinical features.
49:34So if we just add in, you know.
49:36So we take the network strength,
49:38but then we add in baseline
49:40severity or past month cocaine use,
49:41which are things that by themselves
49:43aren't sufficient to predict outcome.
49:44We do find that the model accuracy improves,
49:47so when we do that, if we're doing,
49:49just say yes, no any absence from treatment,
49:51we can get up to 71% accuracy.
49:53But again, this all needs to be validated
49:55in different clinical settings.
49:57And and all that stuff like this is,
49:59you know, and if.
50:00Magnets, so the multisite application
50:02that I mentioned at the end,
50:04I'm really excited about and I think
50:06it's sort of the next step to try to
50:09understand if these things are real,
50:11so we'll see what their viewers think.
50:15Thanks.
50:16Thinking.
50:18He said that was beautiful
50:19and really convincing data,
50:21especially the way you generalized
50:22across those discrete patient samples.
50:24And So what I was curious about is if
50:27you could have any task you wanted in
50:29the scanner because you're using these
50:31cognitive control and reward tasks
50:33that are sort of proxies for constructs,
50:35do you think are relevant and you've
50:37shown variants with different task states?
50:39What do you think the ideal
50:41task would be for prediction?
50:44So it's possible the
50:45answer might be drug use,
50:46and I'm not a huge huge
50:48cue reactivity person,
50:49but I think that that could be the answer.
50:52At least an opiate use disorder,
50:53but the clinics that I work with
50:55tend not to want us to show drug use,
50:58and so I don't know that that's
51:00necessarily the best.
51:01I mean,
51:02one thing that I think would be really cool,
51:04and it Brian Killick and Kathy Carolyn.
51:06I've been talking about is so the
51:08type of behavioral therapy that we
51:10were receiving in the Poly substance
51:12using sample is this computerized
51:13CBT that Kathy Carroll developed.
51:15And we think it would be really cool
51:17to scan patients while they're viewing
51:19the computerized CBT and so then see
51:21if that is a better predictor of treatment.
51:23Response to CDT itself, right?
51:25So if you get them in the
51:28brain state of therapy?
51:30So yeah,
51:30lots of different directions to go,
51:33thanks.
51:44Any other questions? Comments.
51:50Seen a few things in the chat saying what
51:53a wonderful talk it's been, Sarah. Oh yeah.
52:02OK, well if there are no other
52:04comments will close for today and
52:06thank you so much for a really
52:08terrific presentation and stimulating.
52:10And I'm sure this will lead to some
52:13additional collaborations with your
52:15colleagues here at Yale and elsewhere.
52:17So thank you.