A new method for heart rate prediction based on LSTM-BiLSTM-Att

This paper proposes a new method for heart rate prediction based on LSTM-BiLSTM-Att model (Long Short Term Memory, Bidirectional LSTM, Attention Mechanism). In this LSTM-BiLSTM-Att model, LSTM is used to construct the long-term relationship of the heart rate data, and then abstract the high-dimensional features of the heart rate. BiLSTM has the ability to capture the forward and backward correlation information of these heart rate data, and then effectively learn the features of the heart rate data extracted by LSTM. The attention mechanism is added to this proposed model, which can further improve the performance of this heart rate prediction method. The number of neurons, the length of sliding window, and the depth of the LSTM-BiLSTM-Att model are optimized. Two volunteers were randomly selected from 30 volunteers to test the resting heart rate prediction by using random forest (RF) method, ARIMA, the feed forward neural network (FNN), LSTM model, BiLSTM model, LSTM-BiLSTM model, and LSTM-BiLSTM-Att model. The experimental results show that the root mean square error (RMSE) of the resting heart rate of the male volunteer by using this proposed LSTM-BiLSTM-Att model is 2.520, which is 43.9% of RF, 45.9% of ARIMA, 92.1% of FNN and 98.2% of LSTM, respectively; the RMSE of the male volunteer by using the LSTM-BiLSTM-Att model is 1.729, which is 45.5% of RF, 49.5% of ARIMA, 93.6% of FNN and 96.2% of LSTM, respectively. This proposed LSTM-BiLSTM-Att method effectively improves the accuracy of heart rate prediction, and the experimental results prove the effectiveness of this method.