An Adaptive Weight Learning-Based Multitask Deep Network for Continuous Blood Pressure Estimation Using Electrocardiogram Signals

. 2021 Feb 25;21(5):1595. doi: 10.3390/s21051595

Algorithm 1 Training the proposed adaptive weighted multitask network.

Require: Training ECG signals

X_{t r a i n}^{E C G}

, validation ECG signals

X_{v a l}^{E C G}

Ensure: Multitask neural network model

M O D E L

to BP estimation

1:
Initialize task loss’ weight $θ^{i}$ ← 1/3
2:
Initialize multitask network weights including sharing weights $W^{s h a r e}$ and task-specific weights $W^{i}$
3:
Initialize maximum iterated epochs N
4:
for $k = 0 \to N$ do
5:
# Training phase
6:
Load $X_{t r a i n}^{E C G}$
7:
Update parameters of $M O D E L$ with loss function of Equation (4)
8:
# Validation phase
9:
Load $X_{v a l}^{E C G}$ , dividing into M batches
10:
Load trained $M O D E L$ and validate
11:
Compute mean value $L_{m e a n}^{i} (k)$ and standard deviation $L_{s t d}^{i} (k)$ of losses of M batches
12:
if $k > 0$ then
13:
Compute trend of mean value of losses $T_{m e a n}^{i}$ based on Equation (8)
14:
Compute trend of standard deviation of losses $T_{s t d}^{i}$ based on Equation (9)
15:
Compute task-specific weight $θ^{i}$ based on Equation (10)
16:
end if
17:
end for