Episode 5 - AI & Facebook to screen for depression

Interactive Transcript

spk_0 Welcome to the intersection, the podcast about artificial intelligence and healthcare.

spk_0 Hi everybody, this is Maxime.

spk_0 Thank you for tuning in today in episode five of the intersection podcast.

spk_0 We are discussing a PINAS paper titled Facebook Language, Critics, Depression in Medical

spk_0 Records.

spk_0 So I assume everybody knows about Facebook.

spk_0 So let's ask Jean, here with me today, to tell us a little bit about depression before

spk_0 talking about the machine learning that has been done and the results.

spk_0 Yeah, so hi Maxime.

spk_0 Well, what people call depression is actually called major depressive disorder medicine.

spk_0 And it's actually a pretty common and serious medical illness that can negatively affect

spk_0 you, how you feel, the way you think, how you act.

spk_0 So how is depression typically diagnosed today?

spk_0 What are the symptoms?

spk_0 Well, you have to look at a few symptoms and the diagnosis relies on the combination

spk_0 of symptoms, right?

spk_0 It's not because you have one of those that you are depressed, you need to have several

spk_0 of those and during a long period of time.

spk_0 I see.

spk_0 And how is it as said that I present those symptoms?

spk_0 Is that through medical recording or is it through kind of talking with a therapist and

spk_0 with a medical doctor?

spk_0 Well, there are scales, but basically the first step would be to discuss with your medical

spk_0 like your PCP and then with a psychiatrist, for example.

spk_0 But let's focus on symptoms first.

spk_0 Okay.

spk_0 So the symptoms include feeling sad or having a depressed mood.

spk_0 Having a loss of interest or pleasure in activities that you use to enjoy.

spk_0 Changes in appetite like weight loss or weight gain and related to dieting, travel sleeping

spk_0 or even sleeping too much.

spk_0 A loss of energy or increased fatigue.

spk_0 Increase in a purposeless phase of activity, for example pacing or slow movements and

spk_0 speech.

spk_0 So this is typically reported by other people, not yourself.

spk_0 Feeling worse less or guilty.

spk_0 Having difficulty thinking and even in the most extreme cases, having thoughts of

spk_0 death of suicide.

spk_0 So it's a whole spectrum of things.

spk_0 It's not only one of those symptoms.

spk_0 I see.

spk_0 And what I can think already that it might be quite difficult to actually screen for

spk_0 depression because for most of the symptoms that you describe right now, even myself that

spk_0 I don't consider myself depressed right now, I experience them on a most weekly basis.

spk_0 I assume the screening is quite challenging.

spk_0 Exactly.

spk_0 That's why it's important to keep in mind the concept of having several of those and

spk_0 also having several of those for at least two weeks in a row.

spk_0 So it's everyone, each of us has some like done and up times.

spk_0 So it's important that if these downs keep coming up and keep getting longer and longer,

spk_0 then you might need and seek help.

spk_0 I see.

spk_0 And also, so it means you have to go to the therapies, which is quite an important step, right?

spk_0 And then you have to kind of provide a nuclear rate assessment of whether you have

spk_0 experience of symptoms in the past two weeks, right?

spk_0 Which is not something that is necessarily trivial either because those are quite qualitative

spk_0 symptoms.

spk_0 Yeah, absolutely.

spk_0 And you also should keep in mind that some medical conditions, typically cyroad problems,

spk_0 brain tumor or vitamin deficiency, could actually mimic the symptoms of depression.

spk_0 So before you actually diagnose depression, you really need to rule out any other medical

spk_0 condition that could mimic that, right?

spk_0 So it's like, it's not a default diagnosis, but it's not the first diagnosis that you

spk_0 need to take.

spk_0 Thank you.

spk_0 And thus depression, I guess what is the scope of the disease?

spk_0 Well, it's actually a disease, right?

spk_0 That's how it's called.

spk_0 So it's a disease with very serious consequences and it's actually pretty frequent.

spk_0 So for example, one in six people, so that's basically 16%, with experience, depression

spk_0 at some point in their life.

spk_0 Okay, so one in six people basically means that if I consider a family, let's say a family

spk_0 unit of six people in every family unit, there is an average one person who is affected by

spk_0 that disorder.

spk_0 So it's a pretty, pretty important topic I assume.

spk_0 Right, and you should also remember that depression can strike at any time of your life,

spk_0 but on average, the first episodes appear during the late teens to the mid-20s.

spk_0 And unfortunately, women are also more likely than men to experience depression.

spk_0 I see.

spk_0 And so what can you do when you have depression?

spk_0 How important is it to be screened?

spk_0 And then I guess to be diagnosed on the outcome, right?

spk_0 Right, because depression is actually a very treatable illness, it's actually very important

spk_0 to get screened.

spk_0 Because once you get screened and diagnosed, then you can get the appropriate treatment,

spk_0 right?

spk_0 So treatments usually include psychotherapy, medication, or in the most extreme cases,

spk_0 electroconversive therapy.

spk_0 But there are also a number of things that people can do to help reduce the symptoms of

spk_0 depression.

spk_0 For example, for many people, regular exercise helps create positive feelings and improve

spk_0 mood, getting enough quality sleep on regular basis, eating a healthy diet, avoiding alcohol.

spk_0 All these stuff can help a lot in that regard.

spk_0 And machine learning is actually also used with all of these activities that you mentioned,

spk_0 right?

spk_0 Machine learning used to help set tracks the amount of calories in your diet, machine learning

spk_0 being used to make better sleeping recommendations and so on.

spk_0 But that's not a topic of today's article, but it's interesting to see, right, that the

spk_0 extent of the impact that machine learning can have for that particular problem.

spk_0 So let's focus on the scope of that article, which is basically screening.

spk_0 And it's extremely important in the case of depression because most people with depression

spk_0 do not get the care the need, even if treatment exists.

spk_0 For example, only 20% of people treated for depression have received what the National

spk_0 Institute for Mental Health defines as a minimally adequate treatment.

spk_0 So that leaves 80% of people not getting the right treatment.

spk_0 So I mean four out of five patients do not get the treatment?

spk_0 Because they are undereignosed or because they do not get simply even if they are diagnosed

spk_0 the right treatment.

spk_0 So hence the importance of screening better diagnoses and better treatments?

spk_0 That makes sense.

spk_0 And this is actually exactly what this article is trying to tackle, right?

spk_0 Everybody, I guess most of the people use Facebook.

spk_0 People are very active on Facebook.

spk_0 People post and through the post, they reveal things about the way they feel.

spk_0 They reveal things about their friends and so on.

spk_0 And so in fact, what this article is trying to do is to build a somewhat non-invasive

spk_0 or effortless prediction model of depression using contents from the Facebook profile of a given

spk_0 in video.

spk_0 So let me say a bit more about the inputs that this prediction model takes, right?

spk_0 So there are five inputs.

spk_0 One is the textual content of the Facebook post.

spk_0 So what do you write in your post?

spk_0 The second one is how long your posts are.

spk_0 The third one is how frequently do you post?

spk_0 The fourth one is is there a particular patterns in the way in which you post, right?

spk_0 More at night, more in the day and so on.

spk_0 And then the last thing which it doesn't really have to do with Facebook but is an additional

spk_0 kind of input is a demographics, right?

spk_0 So some variables about your background and so on.

spk_0 Okay, and so now we know the inputs.

spk_0 What is the actual output of the model?

spk_0 So this model outputs the probability between 0 and 1 for an individual to be experiencing

spk_0 depression based on this input.

spk_0 And the way that this model is trained, right?

spk_0 So we have inputs we have outputs is using supervised learning.

spk_0 So in other words, the parameters of the model are adjusted to maximize the accuracy of

spk_0 the model prediction on a cohort of patients for whom the presence or absence of depression

spk_0 has been previously established and is available in their medical records.

spk_0 Okay, and let me just maybe take that opportunity to discuss what is the difference between

spk_0 supervised and unsupervised learning?

spk_0 Yes, so that's an important point.

spk_0 So in supervised learning, we are basically training a model on a data set for which we know

spk_0 the input and we know the output and we are basically trying to learn a function that maps,

spk_0 let's say X to Y input X output.

spk_0 Right.

spk_0 In unsupervised learning, the simple definition would be to say that we are trying to learn

spk_0 mapping from X to Y, but we don't know the ground truth why for the training samples X.

spk_0 So in unsupervised learning, we are basically trying to find pattern that define a population

spk_0 or any kind of outcome without actually knowing the outcome.

spk_0 Exactly.

spk_0 And one of the things that makes unsupervised learning quite important and in fact, even some

spk_0 pioneers in AI say that it might be the future is that you don't need to have labels.

spk_0 Right.

spk_0 So you can gather much bigger data sets more.

spk_0 Yeah, and as we already discussed, labels, especially in healthcare, are extremely costly,

spk_0 extremely rare.

spk_0 So being able to do prediction without any label would be a huge, huge step up.

spk_0 Exactly.

spk_0 And this is actually to come back to the article.

spk_0 That's a nice example of this because this article extracts the ground truth label from

spk_0 the medical records of the patients, which means, which means that for every patient in

spk_0 the training set, you have to access to have access to the medical records, you have

spk_0 to extract that label, you have to verify that the label is accurate and you have also

spk_0 a lot of other logistically issues with this.

spk_0 Right.

spk_0 And this is what actually is making the main quality of the article because we know that

spk_0 there were already other studies published about depression prediction from Facebook.

spk_0 But these previous studies, like, were more based on declarative labels from patients.

spk_0 So patients saying, I am depressed, I am non-depressed, which is very different from actually

spk_0 using the ground truth from a real medical record.

spk_0 That's right.

spk_0 So yeah, in fact, this article uses in house that I said from one academic center in the

spk_0 US that has about 700 patients.

spk_0 And in that group of 700 patients, there is about 100 patients who have had a diagnosis

spk_0 of depression in their medical records.

spk_0 So again, like we discussed during the last episode, a very significant class imbalance

spk_0 in that dataset.

spk_0 Definitely.

spk_0 Here you have an imbalance, which is of basically for every, out of seven patients,

spk_0 only one, as a positive label, in the sense.

spk_0 Okay.

spk_0 So maybe you can tell us the performances of that model, is that good?

spk_0 Yes.

spk_0 So the performance, the performance is all right.

spk_0 I think that the goal of this paper is not so much to establish an amazing performance

spk_0 in the sense that the performance actually is of the same order as the performance of

spk_0 the algorithms that have been reported previously.

spk_0 What I found most interesting in this article is actually to make an assessment of what are

spk_0 the inputs that actually are the most important when it comes to predicting depression.

spk_0 Right.

spk_0 And in that regard, again, we are reaching to interpretability for the models.

spk_0 Knowing which feature is important to do prediction.

spk_0 Exactly.

spk_0 And it turns out that the model that is using that study is in fact quite simple.

spk_0 It's a logistic regression model in a few words.

spk_0 And the fact that this model is simple actually simplifies its interpretation.

spk_0 And so the main finding of this study is that the feature that by far is the most relevant

spk_0 in predicting depression is a language, the Facebook language that is used in the post

spk_0 of the patient.

spk_0 And I can tell you a little bit about how the authors actually have assessed this.

spk_0 Right.

spk_0 So they have taken, for example, one Facebook post from one individual in the data sets.

spk_0 And from that Facebook post, they extract language topics using the latent Dirichlet allocation

spk_0 method.

spk_0 They extract language topics that this Facebook post basically exhibits.

spk_0 Right.

spk_0 And just for the sake of discussion, you mentioned that the algorithm is basically logistic

spk_0 regression.

spk_0 Yes.

spk_0 Is that actually machine learning?

spk_0 Can we discuss that?

spk_0 Yeah, that's a tricky question.

spk_0 I think it depends really on what is the definition of machine learning.

spk_0 But if you think of even some of the most complicated networks today, deep neural networks.

spk_0 For example, let's say deep convolutional neural networks in computer vision, the building

spk_0 block of those networks.

spk_0 Right.

spk_0 When you look at a single neuron, this is essentially a logistic regression.

spk_0 Exactly.

spk_0 So I would say that the answer to that question is not necessarily that important.

spk_0 It's a matter of language.

spk_0 But my answer will be that I think it's okay to cause this machine learning.

spk_0 Okay.

spk_0 And maybe we can also stress on the fact that logistic regression is much more easy to understand

spk_0 because it lets complex than a deep neural network.

spk_0 So that is in that particular setting, an advantage of that technique.

spk_0 Exactly.

spk_0 And actually to come back to my answer, I think that what is important is razors and trying

spk_0 to say that logistic regression is not machine learning, but machine learning is something

spk_0 else that have emerged in the past decade.

spk_0 It's actually a better answer to say that it is machine learning.

spk_0 And to recognize that machine learning has been around for actually a long time.

spk_0 The only thing that really is evolving and changing is the complexity of the model

spk_0 architecture that we use.

spk_0 Right.

spk_0 So just to give a little bit of context, there is some kind of war waging between statisticians

spk_0 and machine learning or scientists.

spk_0 But what you're saying basically that we need to go a little bit beyond that and focus

spk_0 on the interest and the importance of the question that is asked in a given study.

spk_0 I would agree with that.

spk_0 And so to come back to this study, so a simple, as we said, a simple logistic regression model

spk_0 that takes a few different inputs to predict the probability of experiencing depression.

spk_0 The main finding of this article is that the most important feature is a language in your

spk_0 Facebook post.

spk_0 And it turned out that the authors are able to extract what are the topics that they

spk_0 are found to be most positively associated with the future depression diagnosis.

spk_0 And those in fact are pretty common sense.

spk_0 So for example, this is deep-press mood and feelings.

spk_0 So words like tears, cry, topic of loneliness.

spk_0 If you write, for example, the word miss in your post, topics of hostility with a word

spk_0 like hate, somatic complaints.

spk_0 If you write about hurt and sick and any word that also refer to medicine like hospital

spk_0 and pain.

spk_0 Okay, great.

spk_0 Anything else to add?

spk_0 Yeah, I think maybe we can close by obviously reminding that this is an important medical

spk_0 issue.

spk_0 Depression is a very debitating illness that affects too many people in the US and worldwide.

spk_0 And given the number of Facebook users around the world, the potential of screening

spk_0 individuals, the other Facebook activity is a very appealing prospect that we hope to

spk_0 see reach its promise in the future.

spk_0 Yeah, I really, really agree with you.

spk_0 But just to give the other side of the coin here, we also know and there have been studies

spk_0 published about that.

spk_0 That's Facebook users are more prone to depression than other people.

spk_0 So we need also to keep that in mind and it could also be a potential bias of that study.

spk_0 But still, I agree with you.

spk_0 Yeah, you mean, yeah, if you step back even more and look at the net effect, you mean

spk_0 of having Facebook?

spk_0 This goes beyond the scope of this study and definitely would be worse in investigating.

spk_0 What I take away from this article personally is the potential not only of screening potentially

spk_0 for depression, you know, through Facebook activity, but also the potential of reacting to

spk_0 the prediction that will be made by such a model.

spk_0 Yeah, actionable outcome.

spk_0 Exactly.

spk_0 For example, you know, if this model were to find out that a Facebook user is depressed

spk_0 and maybe in the worst case, you know, it's depressed with thoughts of death and suicide

spk_0 attempt, that there could be a reaction from Facebook, for example, by suggesting positive

spk_0 messages or ads for mental health resources and so on.

spk_0 The second thing that I take away is that data sharing is often seen as harmful.

spk_0 There's a lot of articles in the media about the danger of data sharing, but really what

spk_0 this article shows is that if it's not used maliciously, it also has a real positive potential

spk_0 when used to tackle issues related to human health.

spk_0 Yes, exactly.

spk_0 What we need to keep in mind is that everything is not black or white.

spk_0 It's a balance.

spk_0 And it's also the responsibility of the people who own that data and we are talking about

spk_0 Facebook, but we could also discuss about medical data from the hospital.

spk_0 The responsibility of the people owning that data to actually do something with it for

spk_0 the better, for the better good of everyone.

spk_0 So it's not as easy as do not use that data or use that data, but we need to try to think

spk_0 about how can we use that data in the best way possible.

spk_0 I see we are entering a new age and somehow we are not yet ready for it in the sense of

spk_0 we don't have yet the right laws, the right maybe even attitude, the way of thinking about

spk_0 what to do with this data, what is right, what is not and how to basically handle this.

spk_0 Right.

spk_0 Okay.

spk_0 So that's it for today.

spk_0 We hope you liked it.

spk_0 If you enjoyed it, subscribe and we'll see you next time.

spk_0 Bye bye.

spk_0 Merci.

spk_0 A bientôt.

Episode 5 - AI & Facebook to screen for depression

Interactive Transcript

Topics Covered

Related Episodes

Exploring Integrative Medicine and Holistic Health with Dr. Sandeep Gupta - Part One

The New Science of Women's Health with Tamsen Fadal & Dr. Amy Killen

216. Dr. Labib Ghulmiyyah: On How To Prepare Your Body for Pregnancy, Fertility Decline, and Preconception Health

AI Therapy: An Open Conversation with Therapists

Share Episode