Algorithm 3 Trend Attention Mechanism.

Input:      physiological time series S Output:      trend level attention weights 1:  $n{s}^{ch}=getNor\left(s^{ch}\right)$; // normalize each channel of the time series 2:  $m{s}_k^{ch}=getMean\left(n{s}_k^{ch}\right)$; // calculate the mean of the k-$th$ segment $n{s}_k$ 3:  $d{s}_k^{ch}=getDiff(m{s}_k^{ch},m{s}_{k^{{}^{\prime }}}^{ch})$; // calculate the difference between all the segments 4:  $tr{c}^{ch}=getMax\left(d{s}_k^{ch}\right)$; // obtain the maximum value of $d{s}_k^{ch}$ 5:  $AttO$ = getSim(X, $tr{c}^{ch}$); // calculate the similarity between X and $tr{c}^{ch}$ 6:  $AttT$ = getFull($AttO$); // convert $AttO$ into $AttT$ through the fully connected layer 7:  $\beta$ = softmax($AttT$); // calculate the trend level attention weights