Multi-Agent QLoRA-Optimized LLMs for Personalized Healthcare

Team: M. Nithyashree, Pavithra N, Krupa P. Nadgir

Summary: A lightweight, memory-efficient framework for deploying multi-task Large Language Models (LLMs) on edge devices in healthcare settings. Integrates real-time IoT sensing with quantized LoRA adapters to deliver diagnostic reasoning, health education, and one-day diet planning—all on resource-constrained hardware.

Features

• Modular adapter-based fine-tuning
• 4-bit quantization with QLoRA for minimal memory usage
• Three specialized agents:
• • Diagnostic reasoning
• • Personalized health tips generation
• • One-day diet planning
• Real-time IoT integration (heart rate, SpO₂, ECG, temperature)
• Edge deployment of memory-efficient LLM models

Architecture

• Base Model: DeepSeek-R1-Distill-LLaMA-8B

Hosted on Hugging Face, distilled and quantized to 4-bit

• Adapters:

LoRA for low-rank parameter updates. QLoRA adds adaptive scaling for quantized weights. Stored as .safetensors, dynamically loaded per task

• IoT Pipeline: Hardware sensors stream vitals to the edge device

• Deployment: Single quantized model in memory, Task switching by loading specific adapter

Components

Component	Description
Diagnostic Agent	Interprets symptom descriptions and suggests possible medical conditions
Health Education Generator	Produces tailored preventive care and wellness guidance
One-Day Diet Planner	Generates culturally relevant, individualized meal plans based on real-time health metrics
Web Dashboard & API	User interface for text/voice input, visualization of vitals, and real-time recommendations

Prerequisites

Python 3.10+

PyTorch 2.x

bitsandbytes

transformers

peft (Parameter-Efficient Fine-Tuning library)

NVIDIA Jetson Orin Nano (8–16 GB RAM)

Results

• Fine-tuned models achieve coherent, context-aware outputs vs. generic base model responses

• Training convergence within ~1 k samples per task

• GPU power usage under 80 W; inference on 4-bit model with < 16 GB RAM

• Real-time switching between agents with minimal overhead

Key Contributions

• Demonstrated feasibility of running advanced LLMs on low-power edge hardware

• Introduced dynamic adapter loading to support multiple healthcare tasks from a single base model

• Integrated live IoT data for truly personalized, context-aware AI-driven health insights

Research paper published

Here is the link to our research paper on this for detailed information:

[Multi-Agent QLoRA-Optimized LLMs for Personalized Healthcare : M Nithyashree, Krupa P. Nadgir, Pavithra N and Sindhu D. V.] (https://drive.google.com/file/d/1iOIQPCFURs6PIhn7Y7YVTBb5yoASxk0u/view)