In a piercing letter to his brother, Vincent van Gogh captured the mental anguish of depression in a devastatingly perfect visceral metaphor: “One feels as if one were lying bound hand and foot at the bottom of a deep dark well, utterly helpless.”Anyone who has suffered from this debilitating disease knows that the water in that well is qualitatively, biochemically different from the water in the puddle of mere sadness. And yet, even as scientists are exploring the evolutionary origins of depression understanding and articulating the experience of depression remains a point of continual frustration for those afflicted and a point of continual perplexity for those fortunate never to have plummeted to the bottom of the well.
This explains well why more than one-fourth of people with major depressive disorder
are undiagnosed and fewer than half receive treatment. However, not all people request help for depressive symptoms,
reason being social stigma and lack of awareness of help platforms. Not just this, sometimes individuals don't associate their feelings
to depression until years later. It is typical of people to go for long periods of time feeling different, uncomfortable, scared and in pain
without attaching the notion of depression to their situations. A sample of comments from an interview with 20 individuals diagnosed with
clinical depression includes:
"I don't think I was aware of what was going on. I just knew a lot of times, I just felt uncomfortable relating to people"
"Did I know what it was? It was pain, but I don't think I would have called it depression. I think I would have called it pain"
"I had no label to put on my feelings, like depression"
In 2017, an estimated 264 million people in the world experienced depression. A breakdown of the number of people with depression by region is shown below. Latest estimates of mental health disorder prevalence, disease burden rates, and mortality impacts across a number of disorders can be found here.
The major hindrance in the curing process is that diagnosing depression has been a difficult task since it can manifest in so many different ways. For example, some clinically depressed individuals seem to withdraw into a state of apathy while others may become extremely emotional. Eating and sleeping patterns can differ from being either in excess or being completely eliminated form one's routine. Diagnosing and tracking methods that still rely mainly on assessing self-reported depressive symptoms usually involve filling out surveys or engaging in face-to-face interviews, provide limited accuracy and reliability and are costly to track and scale. Other reasons such as stigma from the society, lack of awareness of the help and support groups etc. adds up to the challenges.
It has been found that those in misery and depression, tend to find comfort in having companions in woe. Not just this, but studies also suggest that those in a sad mood tend to prefer sadness in the music they listen to. While the surveys and self-descriptive tracking methods fail to diagnose symptoms of depression, peculiarities as those of rumination which come exclusively with depression can be analyzed to predict the state of mind of an individual.
On analyzing the comments as a function of videos present on YouTube, we found out that viewers tend to ruminate a lot on videos with Depressive triggers (e.g. Self-Injury, NSSI, Suicidal short films and multi fandoms). The nature of comments suggest that users tend to find a validation amongst others who are going through or have gone through the same situation. Not only do they feel comfortable, but at times it also gives them the courage to confess about their own situations. This acted as the base of our solution wherein we aim to analyze the watching patterns of an individual on YouTube by analyzing the affective content of the videos watched by her over a window size of 24 days with an overlap of 10 days.
We propose a solution to aid in early diagnosis of depression by giving a zoomed-in perspective of patient’s state of mind and experiences (usually undisclosed to physician) prior to process of seeking treatment in order to help target the at-risk population and improve diagnostic accuracy and efficiency. A pipeline of the solution is shown below:
For training our model, we'll require it to train on the CES-D scores of individuals corresponding to the patterns recorded in their YouTube watching history. To be precise, we would require the YouTube watching history of individuals over a period of time (preferably 3 months) and their score on the CES-D scale. We are in the process of building a platform that would anonymously record this data for our model to train upon.
Once we have this data, we'll calculate the affect of each video by predicting their arousal and valence values and calculating the affective score by cumulating them. The patterns observed in the affective scores of the videos watched over the duration will then be mapped to the CES-D score of the corresponding individual to help the model learn the watching patterns which correspond to a certain state of mind. The kind of patterns we expect to observe for individuals in or heading towards depression are:
What's important to note here is this whole line of research is not focused on predicting the affective score of the video itself, but the kind of mental experience it induces or attracts.
Once the model is ready, it can be used by an individual to track his/her change of watching patterns over a period of time. This behavioral pattern can be useful in determining the mental state of the viewer. This can be used to monitor recommendations and restrict the recommendations of triggering videos based on the watching activity. In case the patterns reveal the signs on unhealthy mental state, this tool can provide early diagnosis by warning the users about the change points in their mood and also direct them to the right portals for help. This would have a substantial impact for individuals both in the initial stage of depressive disorders and those who are already suffering from a depressed mental state. Our proposal aims at providing earlier intervention to improve recognition and management of depression in primary care by understanding patients’ inner experiences prior to and during the process of seeking treatment.
In an ideal scenario, our model should:
This post explains the model trained for classification of videos as depressive. It also discusses the data collection methodology and various other models that we experimented with.
In this post, the CES-D score calculation along with the analysis of comments is discussed. We talk about our evaluation methodology and results produced by the classifier in a real-life setting.
Calculation of Arousal Valence values from a video is discussed here. We talk about our proposed approach for prediction of arousal valence, the dataset used and the results form various experiments performed.