AI-DxMH-Artificial-Intelligence-Diagnosis-for-Modern-Health

---

Overview

AI-DxMH ("Alfy") is a full-stack AI health assistant that lets patients in remote areas of India describe symptoms in plain language and receive a preliminary triage, likely diagnosis, and guidance on whether to seek in-person care. It targets regions with limited access to qualified physicians.

Core capabilities:

• Real-time streaming chat powered by an open-source LLM via the HuggingFace Inference API
• QLoRA-fine-tuned StableLM Zephyr 3B on a 16 000-sample medical Q&A corpus for domain adaptation
• Edge-deployed Next.js frontend with light/dark theming and mobile-responsive layout
• Published research with formal evaluation — see the IJCRT 2024 paper

Disclaimer: AI-DxMH is a research prototype. It is not a substitute for professional medical advice, diagnosis, or treatment.

---

Architecture

flowchart LR
    User(["Patient (browser)"])
    FE["Next.js 14 Frontend\n(NextUI + Tailwind)"]
    API["Edge API Route\n/api/chat"]
    HF["HuggingFace Inference API\nOpenAssistant Pythia 12B"]
    FT["Fine-tuned StableLM 3B\n(QLoRA / PEFT)"]

    User -- "symptom message" --> FE
    FE -- "POST /api/chat" --> API
    API -- "HuggingFace stream" --> HF
    HF -- "token stream" --> API
    API -- "StreamingTextResponse" --> FE
    FE -- "rendered reply" --> User

    subgraph Offline Training
        MedQA[("MedQuad dataset\n16 407 Q&A pairs")]
        MedQA --> FT
    end

---

Quick Start

Prerequisites

• Node.js ≥ 18
• A HuggingFace API token with inference access

Run the frontend

cd "AI-DxMH FrontEnd"
npm install

Create a .env.local file:

HUGGINGFACE_API_KEY=hf_your_token_here

npm run dev
# → http://localhost:3000

Fine-tune the model (optional)

Open Training_Code.ipynb in a GPU environment (tested on Google Colab with a T4/A100):

pip install -q bitsandbytes transformers peft accelerate datasets wandb
# Then run all cells in Training_Code.ipynb

The notebook fine-tunes stabilityai/stablelm-zephyr-3b with QLoRA (r=16, alpha=32) on the keivalya/MedQuad-MedicalQnADataset dataset and pushes the resulting adapter weights to HuggingFace Hub.

---

Project Structure

AI-DxMH-Artificial-Intelligence-Diagnosis-for-Modern-Health/
├── AI-DxMH FrontEnd/       # Next.js 14 web application
│   ├── app/
│   │   ├── page.tsx        # Chat interface (main entry point)
│   │   └── api/chat/       # Edge streaming API route
│   ├── components/         # UI components (navbar, splash screen)
│   ├── config/             # Site config and fonts
│   └── public/             # Static assets (bot avatars)
├── Training_Code.ipynb     # QLoRA fine-tuning notebook
├── G15FINALREPORT.pdf      # Full project report
└── G15 Final Review PPT.pptx

---

Training

Setting	Value
Base model	stabilityai/stablelm-zephyr-3b
Dataset	keivalya/MedQuad-MedicalQnADataset (16 407 samples)
Method	QLoRA (4-bit NF4 quantization + LoRA)
LoRA rank / alpha	r=16, alpha=32
Target modules	q, k, v, o projections + gate/up/down/lm_head
Trainable params	~25.9 M / 1 552 M (1.67 %)
Training loss (1 epoch)	0.5656
Optimizer	paged_adamw_8bit
Sequence length	512 tokens

---

Paper

AI-DxMH: Artificial Intelligence Diagnosis for Modern Health Published in IJCRT (International Journal of Creative Research Thoughts), January 2024.

Read the full paper (PDF)

---

Tech Stack

Layer	Technology
Frontend	Next.js 14, NextUI, Tailwind CSS, TypeScript
Streaming	Vercel AI SDK, HuggingFace Inference
Inference model	OpenAssistant / oasst-sft-4-pythia-12b
Fine-tuning	HuggingFace PEFT, bitsandbytes, Accelerate
Training data	MedQuad Medical Q&A Dataset