Psychologists have one of the most important jobs in the world, and that is to help people deal with their mental illnesses through tangible treatment plans; to help them find coping mechanisms within themselves, and improving their quality of life. However this cannot happen without considerable knowledge of a patient's past, to learn what makes them, them. And we cannot have that happen, without first realizing that people in today's age leave a subconscious trace of their activities and thoughts everywhere they go; a digital trace. That is a medium of data that is being wasted which can be used in a multitude of possibilities, when it comes to learning about a patient's mind, and empowering psychologists with the one thing they need; insight into their patient.
What it does
eDepress is a solution that provides proactive insight into depressive media expressed by a patient. It uses artificial intelligence to reflect on a patient's public social media feeds, and provides the doctor with an early risk-factor analysis, that can lead to depression & suicidal ideation.
How I built it
I first utilized the Instagram (FB) Scraper API & Twitter REST API, to create a script that can extract text, photos, and external links from a desired user's public profiles. I then created a web-crawling software that searches for depressive posts and related image media. Once I had several thousand records (text and images) of depression, I did the same for normal, non-depressive entities, and built a central repository of freshly labeled records "Depression" & "Non-Depression". I then build a pipeline that takes in the text, tokenizes it, preprocesses it (cleaning & scaling), and converts it into embedding sequences, which I feed (partitioned - training & validation) into my custom 1-D convolutional neural network, and run validations proving a 96% final test accuracy on differentiating depressive posts. I then use the labeled images, and augment 2.5x more (adding filters, manual noise & manipulations), and feed them into my custom 2-d convolutional neural network, which proves accuracies to be upwards of 76%. I then integrate both of these models, in my main platform that runs predictions of each media against them, and averages the sentiment of Depression to compute a final MDD-risk score.
Challenges I ran into
The biggest challenge I ran into was the time constraint, because of which I didn't have sleep for 2 consecutive nights! The sheer task of integrating the pipelines of the models with my main platform was in itself quite daunting to take under. Additionally, going about creating my own dataset and making sense of it, though was strangely satisfying, but was a very obstacle-driven problem.
Accomplishments that I'm proud of
I am very proud that I finished a prototype of a full-stack software, that can be used by psychologists to aid them in helping their patients, and bettering our world collectively. I am also proud that I advanced even the limits that I had set for myself when I starting out this project.
What I learned
I learned that I can accomplish anything regardless of the time or resources constraints that I thought were barriers, and I will be ever grateful for HackTheNorth for igniting that feeling! (Besides that I also learned how to use python's TkInter library for building small-scale apps :)
What's next for eDepress
My next steps for this application are also considering emotion, and the way people speak to others, to hopefully pinpoint a distress call there, in terms of depression risk. I would also like to be able to use a larger set of data to improve the performance of the models from even where they are today, in order to provide a more stable means of proactive depression diagnostics. I'll also be talking to various physicians and psychiatrists to get feedback on this application to improve it even beyond, and also to hopefully get access to national registries of this type of data for training & validation.