AGEI: Adaptive Guided Evidence Interviewing for LLM-Based Depression Screening
This repository provides the implementation of AGEI, an LLM-based adaptive interviewing framework for depression-related evidence elicitation.
AGEI dynamically selects questions based on simulated patient responses and maintains a structured evidence state across 21 symptom domains.
Methodology
AGEI consists of four main components:
- Evidence Extraction: Identifies symptom-related information and assigns NONE, POSSIBLE, or SUPPORTED evidence states.
- Adaptive Planning: Selects subsequent symptom targets based on accumulated evidence and dialogue history.
- Question Generation: Produces context-sensitive questions with repetition controls.
- Interview Control: Supports 10 questions, early stopping and limits repeated targeting of symptom domains.
Implementation Details
- Base Model: Meta-Llama-3-8B-Instruct
- Quantization: 4-bit NF4
- Evidence Extraction: Greedy decoding, maximum 300 tokens
- Adaptive Planning: Greedy decoding, maximum 350 tokens
- Question Generation: Temperature 0.25, top-p 0.90, maximum 80 tokens
- Initial Question: Temperature 0.20, top-p 0.90
- Similarity Threshold: 0.93
- Random Seeds: 42, 123, 2026
Reproducibility
The AGEI_Implementation.ipynb notebook contains the adaptive interviewing pipeline, exact prompting templates, generation configurations, and decision rules.
The framework was evaluated using 20 LLM-simulated patient personas with three planned independent runs per persona.
The downstream BDI-II scoring module is separate and is not included in this repository.
Usage
Open AGEI_Implementation.ipynb in Google Colab, select a GPU runtime, install the required dependencies, and execute the notebook cells.
The implementation requires access to Meta-Llama-3-8B-Instruct and the corresponding simulated-patient adapters.
Citation
If you use this implementation, please cite:
AGEI: Adaptive Guided Evidence Interviewing for LLM-Based Depression Screening.
Disclaimer
This framework is intended for research using simulated patients. It has not been clinically validated and is not intended for real-world diagnosis or clinical decision-making.