README.md

metadata

title: Phq 9 Clinician Agent
emoji: 🐢
colorFrom: indigo
colorTo: gray
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
short_description: MedGemma clinician chatbot demo (research prototype)

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

Technical Design Document: MedGemma-Based PHQ-9 Conversational Assessment Agent

1. Overview

1.1 Project Goal

The goal of this project is to develop an AI-driven clinician simulation agent that conducts natural conversations with patients to assess depression severity based on the PHQ-9 (Patient Health Questionnaire-9) scale. Unlike simple questionnaire bots, this system aims to infer a patient’s score implicitly through conversation and speech cues, mirroring a clinician’s behavior in real-world interviews.

1.2 Core Concept

The system will:

Engage the user in a realistic, adaptive dialogue (clinician-style questioning).

Continuously analyze textual and vocal features to estimate PHQ-9 category scores.

Stop automatically when confidence in all PHQ-9 items is sufficiently high.

Produce a final PHQ-9 severity report.

The system will use a configurable LLM (e.g., Gemma-2-2B-IT or MedGemma-4B-IT) as the base model for both:

-A Recording Agent (conversational component)

-A Scoring Agent (PHQ-9 inference component)

2. System Architecture

2.1 High-Level Components Component Description -Frontend Client: Handles user interaction, voice input/output, and UI display. -Speech I/O Module: Converts speech to text (ASR) and text to speech (TTS). -Feature Extraction Module: Extracts acoustic and prosodic features via librosa (lightweight prosody proxies) for emotional/speech analysis. -Recording Agent (Chatbot): Conducts clinician-like conversation with adaptive questioning. -Scoring Agent: Evaluates PHQ-9 symptom probabilities after each exchange and determines confidence in final diagnosis. Controller / Orchestrator: Manages communication between agents and triggers scoring cycles. Model Backend: Hosts a configurable LLM (e.g., Gemma-2-2B-IT, MedGemma-4B-IT), prompted for clinician reasoning.

2.2 Architecture Diagram (Text Description) ┌───────────────────────┐ │ Frontend Client │ │ (Web / Desktop App) │ │ - Voice Input/Output │ │ - Text Display │ └─────────┬─────────────┘ │ (Audio stream) │ ┌───────────────────────┐ │ Speech I/O Module │ │ - ASR (Whisper) │ │ - TTS (e.g., Coqui) │ └─────────┬─────────────┘ │ ▼ ┌────────────────────────────┐ │ Feature Extraction Module │ │ - librosa (prosody pitch, energy/loudness, timing/phonation)│ └─────────┬──────────────────┘ │ ▼ ┌───────────────────────────────┐ │ Recording Agent (MedGemma) │ │ - Generates next question │ │ - Conversational context │ └─────────┬─────────────────────┘ │ ▼ ┌───────────────────────────────┐ │ Scoring Agent (MedGemma) │ │ - Maps text+voice features → │ │ PHQ-9 dimension confidences │ │ - Determines if assessment done│ └─────────┬─────────────────────┘ │ ▼ ┌───────────────────────────────┐ │ Controller / Orchestrator │ │ - Loop until confidence ≥ τ │ │ - Output PHQ-9 report │ └───────────────────────────────┘

3. Agent Design

3.1 Recording Agent

Role: Simulates a clinician conducting an empathetic, open-ended dialogue to elicit responses relevant to the PHQ-9 categories (mood, sleep, appetite, concentration, energy, self-worth, psychomotor changes, suicidal ideation).

Key Responsibilities:

Maintain conversational context.

Adapt follow-up questions based on inferred patient state.

Produce text responses using a configurable LLM (e.g. Gemma-2-2B-IT, MedGemma-4B-IT) with a clinician-style prompt template.

After each user response, trigger the Scoring Agent to reassess.

Prompt Skeleton Example:

System: You are a clinician conducting a conversational assessment to infer PHQ-9 symptoms without listing questions. Keep tone empathetic, natural, and human. User: [transcribed patient input] Assistant: [clinician-style response / next question]

3.2 Scoring Agent

Role: Evaluates the ongoing conversation to infer a PHQ-9 score distribution and confidence values for each symptom.

Input:

Conversation transcript (all turns)

OpenSmile features (prosody, energy, speech rate)

Optional: timestamped emotional embeddings (via pretrained affect model)

Output:

Vector of 9 PHQ-9 scores (0–3)

Confidence scores per question

Overall depression severity classification (Minimal, Mild, Moderate, Moderately Severe, Severe)

Operation Flow:

Parse the full transcript and extract statements relevant to each PHQ-9 item.

Combine textual cues + acoustic cues.

Fusion mechanism: Acoustic features are summarized into a compact JSON and included in the scoring prompt alongside the transcript (early, prompt-level fusion).

Use the LLM’s reasoning chain to map features to PHQ-9 scores.

When confidence for all ≥ threshold τ (e.g., 0.8), finalize results and signal termination.

4. Data Flow

User speaks → Audio captured.

ASR transcribes text.

librosa/OpenSmile extracts voice features (prosody proxies).

Recording Agent uses transcript (and optionally summarized features) → next conversational message.

Scoring Agent evaluates cumulative context → PHQ-9 score vector + confidence.

If confidence < τ → continue conversation; else → output final diagnosis.

TTS module vocalizes clinician output.

5. Implementation Details

5.1 Models and Libraries Function Tool / Library Base LLM Configurable (e.g. Gemma-2-2B-IT, MedGemma-4B-IT) Whisper gTTS (preferrably), TTS Coqui TTS, gTTS, or Bark Audio Features librosa (RMS, ZCR, spectral centroid, f0, energy, duration) Backend Python / Gradio (Spaces) Frontend Gradio Communication Gradio UI

5.2 Confidence Computation

Each PHQ-9 item i has a confidence score ci ∈ [0,1].

ci estimated via secondary LLM reasoning (e.g., “How confident are you about this inference?”).

Global confidence C=minici. ​ Stop condition: C≥τ, e.g., 0.8.

5.3 Example API Workflow

POST /api/message { "audio": , "transcript": "...", "features": {...} } → { "agent_response": "...", "phq9_scores": [1, 0, 2, ...], "confidences": [0.9, 0.85, ...], "finished": false }

6. Training and Fine-Tuning (Future work, will not be implemented now as we do not have the data at the moment.)

Supervised Fine-Tuning (SFT) using synthetic dialogues labeled with PHQ-9 scores.

Speech-text alignment: fuse OpenSmile embeddings with conversation text embeddings before feeding to scoring prompts.

Possible multi-modal fusion via:

Feature concatenation → token embedding

or cross-attention adapter (if fine-tuning allowed).

7. Output Specification

Final Output:

{ "PHQ9_Scores": { "interest": 2, "mood": 3, "sleep": 2, "energy": 2, "appetite": 1, "self_worth": 2, "concentration": 1, "motor": 1, "suicidal_thoughts": 0 }, "Total_Score": 14, "Severity": "Moderate Depression", "Confidence": 0.86 }

Displayed alongside a clinician-style summary:

“Based on our discussion, your responses suggest moderate depressive symptoms, with difficulties in mood and sleep being most prominent.”

8. Termination and Safety

The system will not offer therapy advice or emergency counseling.

If the patient mentions suicidal thoughts (item 9), the system:

Flags high risk,

Terminates the chat, and

Displays emergency contact information (e.g., “If you are in danger or need immediate help, call 988 in the U.S.”).

9. Future Extensions (Not implemented now)

Fine-tuned model jointly trained on PHQ-9 labeled conversations.

Multilingual support (via Whisper multilingual and TTS).

Confidence calibration using Bayesian reasoning or uncertainty quantification.

Integration with EHR systems for clinician verification.

10. Summary

This project creates an intelligent, conversational PHQ-9 assessment agent that blends:

The MedGemma-4B-IT medical LLM,

Audio emotion analysis with OpenSmile,

A dual-agent architecture for conversation and scoring,

and multimodal reasoning to deliver clinician-like mental health assessments.

The modular design enables local deployment on GPU servers, privacy-preserving operation, and future research extensions into multimodal diagnostic reasoning.