Paper reading and analysis: Deep PPG: Large-Scale Heart Rate Estimation with Convolutional Neural Networks

1. View the time spectrum of the data

The middle one on the left is the heart rate signal, and the others are motion artifacts and other noise signals;

2. DaLiA dataset

Collect data in different scenarios. The device using the EEG signal simultaneously collects the heart rate as the real heart rate value;

3. Proposed neural network architecture DeepPPG

The main steps:

1. Extract spectrum

(1) The ACC data of the three channels is the three-axis acceleration, and FFT respectively to obtain the frequency spectrum;

(2) Single-channel PPG data, FFT transform to obtain spectrum;

(3) After FFT transform, take 240 points in the frequency range of 0-4HZ, add the endpoints, and get 257 points;

Network Architecture:

layers	parameterparam	feature map shape
Conv	kernel_num(8)、kernel_size(1,1) 、stride(1,1)
maxpool	size(1,2)、stripe(1,2)	128*8
Conv	kernel_num(16)、kernel_size(1,3) 、stride(1,1)
maxpool	size(1,2)、stripe(1,2)	64*16
Conv	kernel_num(32)、kernel_size(1,3) 、stride(1,1)
maxpool	size(1,2)、stripe(1,2)	32*32
Conv	kernel_num(64)、kernel_size(1,3) 、stride(1,1)
maxpool	size(1,2)、stripe(1,2)	16*64
Conv	kernel_num(16)、kernel_size(1,1) 、stride(1,1)	16*16
flatten		1*256
FC1	64	1*64
FC2	1	1*1

4、评价指标
$$MAE=\frac{1}{W}\sum _{w=1}^{W}BP{M}_{are}(w)-BP{M}_{ref}\left(w\right)$$

5. Experimental results

Comparing the MAE of the two methods

Comparing MAE of different network parameters

Compare heart rate graphs for different activities

Comparing the MAE of different methods for a single experimenter