LSTM-GPT-4 Integration for Interpretable Biomedical Signal Classification

Approximately 3.8 billion people lack access to essential health services, and diagnostic interpretation remains a major bottleneck in remote and resource-constrained settings. Limited access to specialists and the complexity of biomedical signal interpretation (eg, ECG and EEG) contribute to delays in recognizing cardiovascular and neurological conditions.

To develop and evaluate a technical framework integrating Long Short-Term Memory (LSTM) networks with GPT-4 to provide automated biomedical signal classification and human-readable interpretations, suitable as a foundation for future deployment in resource-constrained environments.

The two-layer LSTM architecture (128→64 units) was selected based on preliminary experiments comparing configurations ranging from single-layer networks (64, 128 units) to deeper architectures (128→64→32 units). The chosen configuration balanced model capacity against overfitting risk and computational efficiency. Preliminary validation showed that deeper networks (≥3 layers) provided negligible performance gains (<1% accuracy) while increasing training time by 40-60%. A baseline 1D-CNN (3 convolutional layers, 64 filters each) achieved 2-3% lower accuracy across all datasets, supporting the use of recurrent architecture for temporal biomedical signals. We implemented a modality-adaptive preprocessing pipeline and single-lead selection strategy. The framework was evaluated using public PhysioNet datasets: MIT-BIH Arrhythmia, PTB Diagnostic ECG, PTB-XL, Chapman-Shaoxing, MIMIC-III Waveforms, and Sleep-EDF. A patient-level split protocol (70/15/15) was used to reduce leakage risk. The LSTM architecture (128→64 units) performed temporal feature extraction with softmax-based classification for mutually exclusive classes. GPT-4 was integrated via API with structured prompts to generate clinical interpretations from model outputs.

For the expert evaluation, we randomly sampled 50 test cases per dataset (150 total: 30 from each class for MIT-BIH, 25 per class for PTB, 20 per class for CHB-MIT) ensuring balanced class representation. Three board-certified physicians (2 cardiologists for ECG datasets, 1 neurologist for EEG dataset) independently reviewed GPT-4 generated interpretations. Reviewers were blinded to whether signals were correctly or incorrectly classified by the LSTM model. Each interpretation was rated on a 5-point Likert scale (1=clinically inappropriate, 5=highly accurate and clinically useful). Inter-rater reliability was assessed using Fleiss' kappa (κ=0.78, substantial agreement). On held-out test sets, classification performance was: MIT-BIH 92.3% accuracy (F1=0.91, AUC=0.95), PTB Diagnostic 94.7% (F1=0.94, AUC=0.97), PTB-XL 88.9% (F1=0.88, AUC=0.93), Chapman-Shaoxing 91.2% (F1=0.90, AUC=0.94), MIMIC-III 89.5% (F1=0.89, AUC=0.92), and Sleep-EDF 87.3% (F1=0.86, AUC=0.91). Expert evaluation of generated interpretations (three board-certified cardiologists) rated clinical accuracy 4.3/5.0, clarity 4.6/5.0, and actionability 4.2/5.0, with strong inter-rater agreement (κ>0.85).

This proof-of-concept demonstrates a explicit methodological integration of deep learning-based biomedical signal classification with GPT-4 based interpretation. The framework provides a technical foundation for future prospective clinical validation, field studies, and regulatory review prior to clinical deployment in underserved settings.