QLoRA Medical AI Practice: Clinical Decision Support System on Phi-3 Mini

This article introduces a clinical decision support model built using Microsoft Phi-3 Mini and QLoRA technology. It can accept patient information in natural language and output death risk assessments with reasoning processes. The project demonstrates best practices for fine-tuning large language models on consumer-grade GPUs, addressing key issues in medical AI implementation such as resource constraints and interpretability.

Artificial intelligence has great potential in the medical field, but its implementation faces unique challenges: sensitive data privacy, the need for model interpretability, complex clinical workflows, and medical scenarios often require models to run in resource-constrained environments rather than relying on expensive cloud computing resources.

Project Overview

An open-source project by ArjunJagdale that uses QLoRA technology to fine-tune the Phi-3 Mini model on consumer-grade GPUs to build a clinical decision support system.

Data Preparation

• Uses the UCI Heart Disease Dataset (299 patients, 13 features), converting structured data into natural language with domain context (e.g., explaining the medical significance of ejection fraction values).
• Training samples include 100 samples with intentionally missing features to teach the model to identify incomplete inputs.

Model Configuration and Training

Attribute	Configuration
Base Model	microsoft/Phi-3-mini-4k-instruct
Parameter Count	3.8B
Quantization Method	4-bit NF4 (QLoRA)
Trainable Parameters	3,145,728 (0.08%)
LoRA Rank	16
LoRA Alpha	32
Training Epochs	3
Hardware	Kaggle T4 x2 GPU
Training Time	~20 minutes

QLoRA Advantages

The trained LoRA adapter is only 12.6MB, can be trained on free GPUs, easily deployed on edge devices, and reduces storage and transmission costs.

Input Example

The patient is a 72-year-old male with hypertension and diabetes. Ejection fraction is 22%, serum creatinine is 2.4 mg/dL, serum sodium is 128 mEq/L, and CPK is 1400 mcg/L. Follow-up period: 30 days.

Output Example

This patient has a high risk of death. Key influencing factors:

• Severe reduction in ejection fraction (22%)
• Elevated serum creatinine
• Severe low serum sodium
• Elevated CPK indicates muscle damage Result: The patient did not survive during the follow-up period.

Incomplete Input Handling

⚠️ Incomplete clinical information detected. Serum creatinine and CPK levels are not provided. These are key indicators for a complete assessment.

Technical Highlights

1. Natural language reasoning: Input is free text, automatically extracts information and marks missing content, lowering the threshold for clinical integration.
2. Interpretability design: Each assessment is accompanied by a detailed reasoning process to help doctors understand the basis for judgment.
3. Resource efficiency: Training completed in 20 minutes on a free Kaggle GPU, suitable for resource-constrained scenarios.

Application Scenarios

• Emergency triage: Quickly assess death risk and prioritize resource allocation.
• Telemedicine: Provide specialist-level risk assessment for primary care doctors.
• Clinical research: Screen high-risk patients.
• Medical education: Help students learn comprehensive risk assessment.

1. Data scale: Based on a dataset of 299 patients, generalization ability for rare cases may be insufficient.
2. Regulatory compliance: Actual deployment needs to meet medical device regulatory requirements.
3. Doctor supervision: AI assessment is only an auxiliary tool and cannot replace professional judgment.
4. Data privacy: Strict security measures are required when processing real patient data.

Open-Source Resources

Provides complete open-source content: training data (heart_lora_ready.jsonl), training script (heart_lora_train.py), Kaggle notebook, LoRA weights (12.6MB), etc.

Technical Insights

1. Key to data engineering: Converting structured data into context-rich natural language unleashes LLM capabilities.
2. Uncertainty quantification: Identifying and reporting uncertainty is more important than accuracy.
3. Efficiency first: Technologies like QLoRA make it possible to develop AI tools with limited resources.
4. Interpretability: Medical AI must provide reasoning processes, not black-box predictions.

Conclusion

This project provides an example for the democratization of medical AI and is a reference implementation for medical applications of large language models.