Fig. 1

InsightSleepNet architecture

This was done to discern which segment of the PPG signal influenced the prediction for each epoch in continuous PPG. Additionally, to identify light sleep, we incorporated the ‘3-minute rule’ [35], an established sleep staging technique that analyzes data over 3-minute intervals, into our local attention module. To achieve this, we set the kernel size of the causal convolutional layer, the initial layer of the local attention module, to 7168 (equivalent to 7 epochs) and a stride of 1. This configuration allowed the layer to cover the preceding 3 minutes of the epoch to be predicted. The causal convolutional layer of the TCN model has the characteristic of applying zero padding equal to ‘kernel size - 1’ on both sides of the input data sequence. This feature enabled us to perform computations as intended from the very first epoch. After the causal convolution, we added a 1D convolutional layer with a kernel size of 1 and a stride of 1 to generate an output with a size of 1. Following that, we applied a sigmoid operation every 1024 epochs, corresponding to the PPG signal length within a single epoch. This enabled the calculation of an attention score for each epoch across the entire PPG sequence, considering causality and the influence of the previous 7 epochs. Consequently, an attention score ranging from 0 to 1 was calculated for each epoch