sasasassaszzd/PHQ8-kNN_reg

Model Name: BertWithKNN for PHQ-8/9 Depression Score Prediction

How to use

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("username/PHQ8-prototype")
model = AutoModel.from_pretrained("username/PHQ8-prototype", trust_remote_code=True)

inputs = tokenizer("I feel tired and down.", return_tensors="pt")
outputs = model.inference(**inputs)
print(outputs)

---

Model Description

This model combines a BERT-based encoder with a k-Nearest Neighbors (kNN) regression head to predict PHQ-8/9 depression scores.
It was trained on:

• MHD dataset for pretraining (Masked Language Modeling task)
• DAIC-WOZ dataset for downstream fine-tuning

Additionally, a simple rule-based sentiment/depression keyword classifier is integrated to adjust the prediction.

---

Intended Use

• Task: Mental health score prediction (PHQ-8/9) from text
• Input: English conversation transcripts (e.g., patient–clinician dialogues, daily self-reports)
• Output: Predicted depression score in the range [0, 27]

This model is for research and educational purposes only, not for clinical or medical use.

---

Training Details

1. Pretraining

   • Dataset: MHD
   • Objective: Masked Language Modeling (MLM)

2. Fine-tuning

   • Dataset: DAIC-WOZ
   • Objective: Depression score regression/classification

3. Post-processing

   • Integrated with a simple rule-based classifier that captures keywords like depressed, hopeless, worthless and adjusts the score.

---

Evaluation Results

• Binary classification (depressed vs. non-depressed):
  • Accuracy: 0.823
  • F1-score: 0.724
• Regression (PHQ-8/9 score):
  • MAE: 3.414

---