Long Short-Term Memory-GPT-4 Integration for Interpretable Biomedical Signal Classification: Proof-of-Concept Study.

Area of Science:

• Biomedical Signal Processing
• Artificial Intelligence in Healthcare
• Machine Learning for Diagnostics

Background:

• Millions lack access to essential health services, with diagnostic interpretation being a key challenge in remote areas.
• Limited specialist access and complex biomedical signal analysis (ECG, EEG) delay cardiovascular and neurological condition diagnosis.

Purpose of the Study:

• To develop and assess a framework combining Long Short-Term Memory (LSTM) networks and GPT-4 for automated biomedical signal classification and interpretation.
• To create a foundation for deploying AI-driven diagnostics in resource-constrained environments.

Main Methods:

• A 2-layer LSTM (128→64 units) architecture was chosen for temporal feature extraction and classification.
• The framework was evaluated on diverse PhysioNet datasets (ECG, EEG) using a patient-level split.
• GPT-4 was integrated via API for generating human-readable clinical interpretations from model outputs.

Main Results:

• High classification accuracy achieved across datasets (e.g., 92.3% for MIT-BIH Arrhythmia, 94.7% for PTB Diagnostic ECG).
• Expert physicians rated GPT-4 interpretations highly for clinical accuracy (4.3/5), clarity (4.6/5), and actionability (4.2/5).
• Strong interrater reliability (κ=0.78 for classification, κ>0.85 for interpretations) was observed.

Conclusions:

• This proof-of-concept demonstrates a novel integration of deep learning and large language models for biomedical signal interpretation.
• The developed framework provides a technical basis for future clinical validation and deployment in underserved healthcare settings.