Depression is the no. 1 cause of bad health and disability worldwide According to the World Health Organization, more than 300 million people across the globe are suffering from depression. The WHO also reports that “depression is the leading cause of disability worldwide.” The numbers are rising even faster among millennials (up to 47%) and adolescents (an increase of 47% for boys and 65% for girls) compared to adults.
Low exposure or low rate-of-depression data in some regions Taboos against mental health disorders may drive down diagnosis rates, for example in East Asia. Some regions also have poor public health services. The paper further cautions that reliable depression surveys don’t even exist for some low-income countries -- a common issue with global studies -- forcing the researchers to come up with their own estimates based on statistical regression models.
Untreated depression or incorrect diagnosis Only fewer than half of those affected in the world (in many countries, fewer than 10%) receive effective care. The barriers to effective care include lack of resources, lack of trained health-care providers, and social stigma associated with mental disorders. Another barrier to effective care is inaccurate assessment. In countries of every income level, both people who are depressed or not are often misdiagnosed; the ones who do not have the disorder are often prescribed antidepressants.
Lost of productivity and high economic cost Mental health disorders cost the global economy $1 trillion in lost productivity a year, with depression being the leading cause of ill health and disability, according to the WHO. The total economic burden of depression is estimated to be $210.5 billion per year with 48 to 50 percent of it are attributed to absences from work as well as decreased productivity caused by depression.
What it does
As we know depression is rising faster among millennials and adolescents. The median age of depression onset is 32.5 years old.
Nowadays, people from nearly all ages in the world use social media as a platform to express their feeling about everything.
Deprazer will leverage social media as our source for detecting depression distribution in a region. What does Deprazer do? (a) Deprazer will do depression rate analysis in Twitter to detect depression level in many countries and cities in the world. However, in our prototype, we only support regions that use English as their language. (b) Deprazer will inform top 5 keywords that most commonly appear in the depression tweets in a region. (c) For the personal user, Deprazer will inform the severity of the person’s depression by giving percentage of her/his tweets that are detected as depression tweet. (d) Our stakeholders are parents, hospitals or mental health services, researchers, and governments. (e) We won’t provide our depression rate analysis from Twitter to other parties except to the relevant requester (individuals, relatives, or hospital staffs). Besides, we will analyze a person’s Twitter account only after he/she has already given his/her permission to the requester.
How does Deprazer solve the 3 main depression problems?
Deprazer will help researchers to solve low exposure or low rate-of-depression data in some regions by providing holistic insights from Twitter post analysis of people in the region. Researchers don’t need to come up with their estimation of depression distribution data because Deprazer will perform the analysis.
It will help the government to take care of the issue seriously. In the future, we also want to provide the depression data segmented by users’ age. The government can then utilized this to evaluate their policy. For example, (a) If the issue is high among adolescents, there might be parenting issues or too much pressures in their school. (b) If the issue is high among adults, there might be too much pressures in their working environment. (c) On the other hand, we can see that depression is a serious problem and we can increase the availability and adequacy of mental health services in regions where the severity is high.
After the government is more aware to the depression issue, they can persuade the citizens to also become more aware of the issue. We can facilitate people who want to know if their relatives are depressed or not to analyze them. Deprazer will provide a page to let user analyze his/her Twitter account personally, through hospital, or treatment facility. In the future, Deprazer will also provide suggestion containing a list of mental health services nearby. Therefore, it will eliminate untreated depression or incorrect treatment.
Last but not least, for hospital staffs or mental health services, Deprazer will help them to provide the best way to treat a patient by providing information about his/her tweets. Deprazer will crawl his/her Twitter account; this is supposedly based on agreement with relevant parties.
In this way, we will make world a better place by reducing depression rate, the result of which is productivity increase that will impact the economy positively all over the world.
How we built it
After the idea is finalized, we start developing this solution by dividing the task into two parts: the modeling and the web part. In the modeling part, we develop a neural-network based model using TensorFlow. The model do not employ word-based embeddings but use character-based embeddings instead. That way, training another model for another language will be simple. The embedding for each character of a word is fed to a bidirectional LSTM (long short-term memory) layer. The last states of both layers are concatenated to form the “embedding” for a word. These word “embeddings” are then fed to another bidirectional LSTM. The last states of both layers are concatenated then fed to a fully-connected layer. The output of this layer is then used by a sigmoid unit to determine if a tweet is a depression tweet or not. For the web part, we build this prototype using Python flask for integration with the AI model and Twitter API for the backend. For the frontend, we utilize d3js for the map visualization. The front-end itself is divided into several pages, one of which is the homepage that shows the United States map. The map is designed to be clickable for each state. Each state has its own page, we called it detail page. It shows us the map of the state with depression level for each county. Apart from map, we create a personal page that will be used by hospital or treatment facilities to help psychiatrists or other similar professions to provide better treatments for their patients. This page is designed to preserve user’s privacy, so it will utilize more authentication features.
Challenges we ran into
We are working on the Twitter API integration with our AI model. The challenge is to provide the best accuracy while preserving the user’s privacy. We even think to introduce some minor aberration, limit the access, and make a stricter authorization to protect our user’s privacy.
Although there is Sentiment140 data that contains ~1.6 million of tweets, it is still not suitable enough for training our model and we need to process it further. The Sentiment140 data only contains general sentiment and not depression level of the tweet. By some automatic processing, we only acquire ~300 thousand of tweets worth training.
Accomplishments that we're proud of
In terms of creativity, we have already done some researches and yes, there are some applications that help users to detect whether they are depressed or not based on their interaction with their phones. There are also some applications that help depressed people to feel better by providing community support similar to group therapy. However, we try to solve depression issue differently. We want to solve this issue by reducing low exposure or low rate-of-depression data. We believe, with sufficient data, people’s awareness will increase and the depression issue will be reduced because depressed people will get correct treatment. What makes us unique in terms of analyzing depression rate issue? Without using word embedding, we can predict the depression probability in tweets with up to 95% accuracy. To replace the word embedding and to determine the depression rate, we only use the last state of LSTMs. That makes our model simpler. The strength of our model lies in its simplicity.
The first point of concern in terms of complexity lies in building the model. We do not use word embedding but use character embedding instead. Word embedding makes it more difficult to train the model for another language because we need big text corpus to train the embedding more accurately. On the other hand, character embedding can be trained more accurately because the number of characters is far more limited than the number of words in a language. As said before, the strength of our model lies in its simplicity. The more simple the model the easier it is to train the model. Therefore, our model can be expanded to another language more simply. To support the model, another layer of complexity is added in form of the web application serving this product. We use python flask in the back-end with jinja2 and several other libraries to serve the front-end. Since this product is geared toward visualization more, we choose d3js to visualize the map in order to improve the user experience. And to create a reliable and scalable service, we use AWS services to serve this product. We use Elastic Beanstalk as the scalable host. We plan to crawl tweets in daily to update our map. We also plan to update our product in the next phase to support time-based map so relevant users can see changes of depression rate in each region over times. We plan to use AWS RDS or Redshift for this update.
As we know, low exposure or low depression-data in some regions will affect the treatment for depressed people negatively and it will make their health worse than before. Mentally unhealthy people will result in loss of productivity and eventually economic loss. Accurate model and data will give useful insights for both the researchers and the governments. a. For the social impact, Deprazer will increase awareness of the danger of untreated depression. Automatic prediction using our model can also preclude an individual from prejudice and social stigma attached to mental disorder the individual suffers from. Therefore, depressed people will hopefully get correct treatment on the proper time. b. It will reduce the number of depressed people resulting in productivity increase that will impact the country’s economy positively.
Moreover, for the monetization, once we perfect our prototype, we will try to partner with a. Hospitals or mental health services to help them explore the background of a patient’s depression based on Twitter analysis. b. Governments by providing depression-rate analysis in their own region.
What we learned
TensorFlow can provide smooth transition from the high-level stand-alone Keras API to the prepackaged Keras API inside it. This gives scientist more time to think about how to develop the model than how to code it, without sacrificing the flexibility and the efficiency of the low-level API the TensorFlow provides.
After doing a deep research for this product, we know some sad facts about depression such as (a). Most people in the world do not aware that they are already depressed, therefore they just ignore it and live their life as usual. However, this untreated depression will create huge impact to their life and surroundings. The most harmful impact is to their health. Depression will affect the way they eat, sleep, and think until it reaches a crisis point where they are more likely to commit suicide. (b). Researchers have already done some researches to overcome this issue. However, since mental health issue is still stigmatized or taboo in some regions, there is a tendency to under-report it. Difficulty to diagnose the issue and insufficient number of mental health services in the region can also contribute to the under-reporting. Under- reporting it can lower the depression rate in the region while concealing the real issue there. (c). That is why we really want to solve this issue and make the world a better place by building Deprazer.
What's next for Deprazer
- Support more languages.
- For personal use, after we analyze a user’s Twitter account and predict his/her depression level, we will suggest him/her to visit a nearby mental health service if necessary. It will help him/her to get proper treatment as soon as possible.
- Once it is ready to be tested in public, we will partner with governments, hospitals, or mental health services, starting with countries that use English as their main language.
- Recap daily tweets by region to generate time-based map so relevant users can see the changes over times and determine what should be done to decrease the depression level or keep it low.
https://wtop.com/medstar-washington/2017/12/depression-hospital-patients-seeking-help-treatment/ https://www.webmd.com/depression/guide/untreated-depression-effects#1 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3137804/ https://www.weforum.org/agenda/2018/05/depression-prevents-many-of-us-from-leading-healthy-and-productive-lives-being-the-no-1-cause-of-ill-health-and-disability-worldwide/ https://www.washingtonpost.com/news/worldviews/wp/2013/11/07/a-stunning-map-of-depression-rates-around-the-world/?fbclid=IwAR39YXN1_5m93KPxpw5u5rftsS3Ublz7suX-bg1lagflLJSTQmEMTUrPXTw&noredirect=on&utm_term=.0b9c9f51df04 https://www.who.int/news-room/fact-sheets/detail/depression https://www.verywellmind.com/depression-statistics-everyone-should-know-4159056