Mental Illness Classifier Platform: WeCare-mind.com
Motivations
Depression is a condition that reportedly affects one in ten Americans at one point or another, and the incidence of depression is higher in some states than others. According to the research conducted by Healthline, 44% of college students report having symptoms of depression, and 75% of them do not seek help for their mental health problems. Suicide is the third leading cause of death among children and young adults aged 10 to 24. As for people over 30, about 50% of all adults experiencing symptoms of depression will not talk to a doctor or seek help for depression.
These statistical findings merely cover depression not to mention including other forms of mental illnesses. Some depressed people avoid seeking help from external resources because they fear social judgment. Others may be more concerned by the high cost of treatment -- even with insurance.
As Ayush, Chetan, Suprith, Shi and I share the same values and goals, we teamed up at the HackUMass event creating the WeCare-mind.com to solve the problems stated above.
WeCare-mind.com
WeCare-Mind.com is a mental illness classifier platform for individuals and families affected by mental illness to give them a better understanding of their mental health condition. WeCare-mind.com uses machine learning to analyze text entered by users and predicts the kind of disorder that could be affecting them. There is still a lot of stigmas associated with mental illness and that prevents people from reaching out to others, for this reason, we created this website to enable people to be heard and to find possible solutions easily.
Data Collection
Ayush and I chose the mental health forums because people diagnosed by trained practitioners mostly post there about their day to day life. That gave us a good basis for assuming labeled data. For example, if a user posts an article in PTSD group, we put it in the PTSD category. We also scraped through Reddit original posts. We thought they would be a good fit because of the popularity of the post and tags of mental diseases. In the end, we gathered 520 reviews across mental diseases like depression, PTSD (Post Traumatic Stress Disorder), ADHD (Attention-Deficit/Hyperactivity Disorder), and PPD (Paranoid Personality Disorder) and labeled them accordingly.
Codes are available on GitHub.
Machine Learning
After gaining our labeled reviews, Ayush and I pre-processed the data by removing stop words, tokenizing, and stemming. Subsequently, we trained the multinomial Naive Bayes classification model with five-fold grid search cross-validation to tune the best hyperparameters. The accuracy score of the test set is about 70 percent.
Suprith, Shi, and Chetan collaboratively built a web app that is supported by Django.
Web App
The Front-end of our project was created using Node JS, and the back end was created using Django. The Django project was put into production on a Digital Ocean droplet with minimum processing and space requirements. Gunicorn was used to spawn threads for the Django server to serve real-time user requests, and Nginx framework was used to create an interface between Network and the Django threads. An object of the trained classifier was created and stored using pickle. For serving real-time diagnosis requests each time the classifier object is unpickled and used to classify the user request string. Based on the classification a custom page is created and returned which consists of disorder-specific coping mechanisms i.e. a diagnosis of ADHD would return a list of curated resources for ADHD and likewise for every disorder that our model predicts. In the future, we hope to add information about healthcare practitioners within a 20-mile radius who specialize in the predicted disorder.
Conclusions
- Unlike other online mental health assessment platforms, the WeCare website asks open-ended questions which allow users to fully describe their feelings in a more detailed manner.
- A free and real-time diagnosis allows people to learn more about their mental health condition and figure out possible solutions to their problems.
- Since the dataset is small and a little bit unbalanced, it would be better if we can scrape more data and try to get each disease proportionally. We will work on Convolutional Neural Network (CNN) approach to classify mental illnesses once a larger dataset are obtained.
Team
Ayush Sharma: CICS, UMass AMherst (M.S. in Computer Science)
Chetan Manjesh: CICS, UMass AMherst (M.S. in Computer Science)
Suprith Aireddy: Illinois State University (B.S. in Computer Science)
Shi Zhang: New York University (B.S. in Computer Science)
Yu-Han Chen: New York University (M.S. in Management and Systems)
About Author
