CoolFace
Apppublic

absiddiq735/Mental_Health_Risk_Analysis_Detector

sourceHugging Facemitupdated 3mo agoView on Hugging Face
0likes
App README

🧠 Mental Health Sentiment & Depression Severity Analyser

Thesis Project — Multi-Task DistilBERT for Joint Sentiment & Depression Severity Detection

Model Architecture

ComponentDetail
Base encoderdistilbert-base-uncased
MAX_LEN256 tokens
Sentiment head768→512→256→3 (LayerNorm + GELU, 3-layer MLP)
Cross-task gate2-layer MLP 3→128→256 (Sigmoid)
Severity branch768→512→256 (LayerNorm + Residual)
Severity refine256→256→256 (2-layer + Residual)
CORAL layer4-class ordinal (3 boundaries)
ThresholdsLearnable per-boundary CORAL thresholds
LossUncertainty weighting (Kendall et al., 2018) + ordinal smoothing

Labels

Sentiment (3-class): negative · neutral · positive

Severity (4-class, PHQ-9): minimum · mild · moderate · severe

Training Datasets (6 sources)

  • —Combined_Data.csv (sentiment for mental health)
  • —Redditdepressiondataset.csv
  • —combined.json (DepressionEmo)
  • —reddit-depression-cleaned (HuggingFace)
  • —mentalhealthcounseling_conversations (HuggingFace)
  • —Suicide_Detection.csv

Disclaimer

Research prototype for academic/thesis purposes only. Not a clinical diagnostic tool.