Apnea_LLM

This project explores Obstructive Sleep Apnea (OSA) classification using physiological signal features and Large Language Models (LLMs) via few-shot prompting.

🔍 Project Phases

1. Data Preprocessing

   • Segment ECG signals
   • Align apnea labels from _respevt.txt

2. Feature Engineering

   • Time-domain + frequency-domain feature extraction
   • Hjorth parameters, spectral entropy, dominant frequency, etc.

3. LLM Reprogramming & Prompting

   • Convert features into natural language prompts
   • Few-shot classification using Claude-3

📈 Results

Claude-3 Few-Shot Performance:

• Accuracy: 93.5%
• Precision (Apnea): Low — class imbalance present

🧠 Next Steps

• Balance prompts for LLM
• Experiment with GPT-4 or Mistral
• Evaluate SHAP-based feature importance

---

🔗 Repository

https://github.com/zobi-logs/Apnea_LLM