Dartmouth Researchers Propose A Deep Learning Model For Emotion-Based Modeling of Mental Disorders Using Reddit Conversations

According to the World Health Organization (WHO), mental diseases impact one out of every four people at some point in their lives, according to the World Health Organization (WHO). However, due to the social stigma associated with seeking professional care, patients in many parts of the world do not actively seek it. As a result, there is a need to passively use interactions to detect mental problems.

It is feasible to persuade patients to seek diagnosis and treatment for mental problems through passive (i.e., unprompted) detection. That is precisely what Dartmouth University academics have proposed. The objective is to concentrate on a subgroup of mental illnesses marked by different emotional patterns.

The research team’s proposed model is entirely based on emotional states and their transitions using conversations on Reddit. Content-based representations, such as language model embeddings, have long been the focus of research in this area.

Domain and topic bias affect content-based representation; hence it does not generalize. As a result, the necessity to suppress topic-specific expression and generalize well across all domains and timelines is becoming increasingly important. The experiments are based on the model’s capacity to detect a variety of emotional problems and its generalizability.

The researchers used their model to categorize the emotions conveyed in user posts and map emotional transitions between them. This allowed each post to be labeled as ‘joy,’ ‘anger, “sadness,’ ‘fear,’ ‘no emotion,’ or a combination of these emotions. The map is a grid that depicts the likelihood of moving from one state to another. Different emotional diseases have their unique patterns of emotional transitions, which have been used as the basis of this model. The model can detect emotional problems by creating an emotional “fingerprint” for a user and comparing it to existing signatures of emotional diseases.

Results:

After being fine-tuned with train and validation data, each classifier was tested on test data. The four metrics used to analyze were accuracy, F-1 score, precision, and recall. As tested by four criteria, all three classifiers (SVM, LogReg, and RF) exhibit adequate performance with ER features. The RF classifier has the best performance across all standards; hence it has been chosen as our model for further applications. 

• Comparison with baseline models:

Model parameters were fine-tuned for each model and tested individually to compare ER and content-based features’ efficacy.

The proposed models’ performance was evaluated using a 5-fold cross-validation technique. In all of the tasks, the random forest model trained using ER characteristics outperforms the baseline model. According to the results, the model performs slightly better in the anxiety and major depressive tests but marginally worse in the bipolar challenge. This phenomenon could indicate that these two approaches to user modeling are complementary.

Although the researchers do not look at intervention options, they expect that this research will show the way to prevention. They make a compelling case in their study for a more thorough examination of models based on social media data. The model’s performance will always come first, but it does not excuse a lack of understanding of the model, its biases, and limitations.

Paper: https://arxiv.org/pdf/2201.09451.pdf

Reference: https://home.dartmouth.edu/news/2022/03/ai-model-detects-mental-disorders-based-web-posts

Suggested