Table 7 Comparative analysis of SleepNet with similar previous works.
From: Integrating physiological signals for enhanced sleep apnea diagnosis with SleepNet
Feature | SleepNet model | Study 149 | Study 250 |
|---|---|---|---|
Data modalities | Multi-modality: ECG, SpO2, airflow for holistic analysis | Single-modality: ECG signals only | Single-modality: ECG signals only |
Model architecture | Hybrid architecture combining CNN (spatial features) and Bidirectional GRU (temporal dependencies) | Deep RNN with stacked LSTM and GRU layers for sequential ECG data processing | Comprehensive comparison of six models: 1D CNN, 2D CNN, RNN, LSTM, GRU, and DNN. Best: 1D CNN and GRU |
Focus | Spatial + temporal feature extraction for accurate apnea detection using multi-modal data | Sequential temporal analysis of ECG signals to detect SDB events | Focus on both temporal analysis and feature extraction but limited to single-modality ECG |
Generalizability | Validated across diverse datasets and demographics, ensuring robustness in varied populations | Limited validation on a dataset of 92 patients, with sequential ECG data segmented into 10s windows | Validation on a dataset of 86 patients; comparison across models, but limited generalization |
Performance metrics | Achieves high accuracy, specificity, sensitivity, precision(PR), and F1 score(F1). Accuracy: 95.19% | High F1-score: 98.0% (LSTM) and 99.0% (GRU) | Accuracy and recall: 99.0% for 1D CNN and GRU |
Computational efficiency | Optimized for clinical deployment with computational efficiency for real-time applications | Moderate efficiency; focused on deep RNNs, which are computationally demanding | Higher computational cost due to comparison across six models, particularly CNNs |
Key strengths | Comprehensive multi-modal integration, hybrid architecture, and robust generalization | Excellent performance with sequential models (LSTM and GRU) for time-series ECG data | Robust comparative analysis of multiple deep learning models; outstanding accuracy with 1D CNN and GRU |
Key weaknesses | Requires diverse patient data for validation to maintain generalizability | Limited by single-modality ECG signals and small dataset size | Limited to ECG signals, lacking multi-modal data, and computationally heavier for CNN models |
