Adaptive RAG for Evidence-Grounded Depression Care

Project Overview

Project Title: Adaptive RAG for Evidence-Grounded Depression Care

Core Idea

Build a Retrieval-Augmented Generation (RAG) system that:

• Grounds responses in verified clinical sources

• Tracks patient progress over time

• Monitors treatment adherence

• Detects risk signals

• Adapts engagement strategy

• Escalates safely when required

This is not a therapist. This is a clinical follow-up assistant.

1. Problem Statement

Depression patients often:

• Stop medication early

• Miss therapy sessions

• Drop out of treatment

• Fail to report worsening symptoms

Existing chatbots:

• Are mostly scripted

• Are not evidence-grounded

• Do not track longitudinal risk

• Do not enforce safety citations

Your system fixes this.

2. System Objectives

• Ensure responses are evidence-based (RAG)

• Maintain longitudinal patient state

• Predict dropout risk

• Detect self-harm risk

• Personalize engagement

• Provide escalation pathways

• Maintain audit traceability

3. System Architecture

Layer 1: User Interface

• Mobile/web chat interface

• Mood check-ins

• Medication reminders

• Progress dashboard

Layer 2: Risk & State Engine

Tracks:

• PHQ-9 scores

• Mood trend

• Medication adherence

• Engagement frequency

• Sentiment change

• Risk score

This becomes your longitudinal state model.

Layer 3: RAG Engine

Pipeline:

User message
↓
Risk classifier
↓
Retriever (vector database)
↓
Evidence selection
↓
LLM generation
↓
Safety filter
↓
Response with citation

All medical claims must be backed by retrieved evidence.

Layer 4: Adaptive Engagement Model

Based on user state:

• Low engagement → gentle reminder

• Stable mood → maintenance check-in

• Worsening trend → increased follow-up

• High risk → crisis protocol

This is where your novelty sits.

4. Technical Components

A. Data Collection

Sources to index:

• WHO depression guidelines

• National clinical guidelines

• CBT manuals

• Patient education materials

• Crisis management protocols

Documents are chunked and embedded.

B. Embeddings & Vector DB

Tools:

• SentenceTransformers

• FAISS / Chroma

Each chunk stores:

• Source

• Section

• Trust score

• Date

C. Risk Scoring Model

Risk Score = weighted function of:

• PHQ-9

• Negative sentiment

• Sudden language shift

• Missed medication logs

• Drop in engagement

Example:

Risk = (0.4 × PHQ) + (0.3 × sentiment) + (0.2 × engagement drop) + (0.1 × adherence gap)

High risk triggers escalation.

D. Safety Filter

Before sending response:

• Block harmful advice

• Check for hallucination

• Verify citation exists

• Add emergency contacts if needed

If suicide ideation detected: Immediate escalation protocol.

5. Evaluation Plan

Clinical:

• PHQ-9 change over 8 weeks

Engagement:

• Retention rate

• Response frequency

Safety:

• False negative risk detection

• Escalation accuracy

RAG Quality:

• Evidence citation accuracy

• Hallucination rate

Use simulated users first. Human trials require ethics approval.

6. Tech Stack

Backend: Python + FastAPI

AI: Llama 3 / Mistral, LangChain / LlamaIndex

Vector DB: FAISS

Database: PostgreSQL

Frontend: React / Streamlit

Deployment: AWS / GCP / Local server

7. Challenges

• Ethical clearance

• Data privacy compliance

• Avoiding hallucinations

• Avoiding medical liability

• Ensuring human-in-loop safety