Access will only be granted to users who fill this Data Request Form

ANGST: Anxiety-Depression Comorbidity Diagnosis Dataset

Overview

ANGST (ANxiety-Depression Comorbidity DiaGnosis in Reddit PoST) is a novel dataset developed to facilitate the diagnosis of comorbid mental health disorders, specifically depression and anxiety, from social media posts. Unlike traditional datasets, which typically treat mental health conditions as isolated, ANGST is designed for multi-label classification, allowing posts to be labeled as indicative of depression, anxiety, both, or neither. The dataset reflects the complex and nuanced nature of real-world mental health discourse, making it an invaluable resource for researchers and practitioners in the field of digital psychiatry and mental health diagnosis.

Access to the Dataset

Given the sensitivity of the dataset and the importance of protecting user privacy, we require interested users to complete a consent form before accessing ANGST. This ensures that the dataset is used responsibly and ethically for research purposes.

Please fill out this Google Form to request access to the dataset. The authors will carefully review each request before granting access, considering the potential impact and relevance of the work to mental health research. We appreciate your understanding in maintaining the integrity and ethical use of this resource.

1. Data Collection

The dataset was compiled from publicly available posts on Reddit, focusing on mental health-related subreddits. Using the PRAW (Python Reddit API Wrapper), we scraped over 400,000 posts spanning from January 2018 to December 2022. Posts were then filtered based on several criteria, including post length (at least 75 words) and the number of posts per author (at least 10). This process ensured that the dataset focused on substantial discussions surrounding mental health, ultimately resulting in 25,000 posts most relevant to depression and anxiety. From this pool, a random sample of 3,000 posts was selected for expert annotation.

2. Annotation Process

The final dataset, ANGST, consists of 2,876 posts meticulously annotated by a team of three expert psychologists. The annotators independently reviewed each post, labeling it as indicative of depression, anxiety, both, or neither. The annotation process also included a review mechanism to resolve any conflicting annotations, ensuring high-quality and clinically relevant data. Inter-annotator agreement metrics, such as Krippendorff’s alpha and Fleiss kappa, were used to evaluate the consistency of the labels, resulting in acceptable agreement scores that reflect the subjective nature of mental health diagnosis.

3. ANGST-Silver

In addition to the gold-labeled posts, we have created ANGST-SILVER, a silver-labeled corpus comprising 7,667 posts. These posts were sourced from the same distribution as ANGST but were labeled using a prompting technique with GPT-3.5-turbo. This supplementary dataset serves as an ancillary resource, enabling researchers to explore semi-supervised and few-shot learning paradigms in mental health diagnosis.

Citation

If you use this dataset, please cite the following:

@misc{hengle2024quitethereevaluatinglarge,
      title={Still Not Quite There! Evaluating Large Language Models for Comorbid Mental Health Diagnosis}, 
      author={Amey Hengle and Atharva Kulkarni and Shantanu Patankar and Madhumitha Chandrasekaran and Sneha D'Silva and Jemima Jacob and Rashmi Gupta},
      year={2024},
      eprint={2410.03908},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.03908}, 
}

---