Multi-View Deep Learning for ECG Sleep Apnea Detection

While single-lead Electrocardiogram (ECG) monitoring offers a low-cost, portable alternative, conventional artificial intelligence models face two core limitations which are (i) analyzing ECG data solely as a 1D time series or a 2D image misses complementary spatial, temporal, and frequency details, and (ii) computer vision backbones (e.g., standard ResNet models) are not optimized to preserve the subtle, multi-resolution temporal features unique to biological signals.

The Multi-View Interactive Convolutional Network (MVIC-Net) addresses these gaps by transforming a single 60-second ECG recording into four distinct data representations, processing them simultaneously to capture full physiological dynamics are (a) raw numerical signal capturing continuous amplitude changes over time, (b) reshaped signal sequence revealing spatial-structural relationships, (c) polar-coordinate transformation mapping signal periodicity and rhythm perturbations and (d) wavelet-transformed image displaying how energy distributes across frequencies over time.

Combining features at the deep-learning layer yields significantly higher accuracy than using any single representation or combining model predictions via simple voting. Ablation analyses confirm that time-frequency scalograms and 2D spatial structures contribute the most to diagnostic accuracy, while 1D and polar plots provide stability against noise.

MVIC-Net demonstrates that combining multi-modal signal representations with specialized feature-interaction backbones creates highly accurate, reliable, and interpretable non-invasive diagnostic tools. This multi-view design methodology provides a blueprint for expanding AI diagnostics across other signal-based healthcare applications, including cardiac arrhythmia tracking and multi-sensor wearable health platforms.