Deep Learning Model with Individualized Fine-tuning for Dynamic and Beat-to-Beat Blood Pressure Estimation

UTransBPNet: A General Deep Learning Model for Cuffless Blood Pressure Estimation under Activities

“…According to our previous work [14] and some preliminary results of this study, CNN, CNN-LSTM and transformer based model structures cannot correctly track intra-individual BP changes, although they have shown good performance under static conditions in the literature. One possible reason is that the BP variation pattern under activities is significantly different from that of static conditions due to the involvement of the various BP regulation mechanisms, thus exhibiting both shortand long-term variations, while existing deep models are incapable of learning effective short-and long-term features from the high- 2) Three different finetuning schemes are devised to regress SBP and DBP from continuous BP waveform, and the scheme with all model parameters finetuned performed the best, outperforming the method that directly detecting maximal and minimal feature points from the continuous BP waveform;…”

“…Deep Learning Model Structures. The recent advances in deep learning have brought the surge of fully data-driven models that take raw physiological signals as inputs, such as electrocardiogram (ECG), photoplethysmogram (PPG), ballistocardiogram (BCG), or the combination of them, which can automatically learn representations from these signals without handcrafted feature design [14][15][16][17][18]. Various deep learning models have been proposed for cuffless BP estimation in recent years such as deep neural network (DNN) [19] and one-dimensional (1D) convolutional neural network (CNN) [20], etc.…”

“…In addition, recurrent neural network such as long short-term memory neural network (LSTM) [12,22] and gated recurrent unit (GRU) [14,23] have been widely employed for BP estimation, which can make predictions of current time stamp with relatively long periods of previous information, which however suffer from the high variations of input features. To combine the strengths of the two structures, CNN-LSTM or CNN-GRU based models have been proposed and become the most popular BP estimators in recent studies [14-16, 24, 25].…”

“…A practical way is to build general models based on data from a group of subjects. For general models, individual calibration is not a necessitate, depending on the availability of individual calibration data, but they usually showed lower accuracy compared to calibrated models [14,15]. However, most studies built individualized models based on the MIMIC database which recorded continuous BP as the references thus providing sufficient individual samples for training the model, i.e., normally the first 70-80% individual data were used for training the model and the remaining data for testing.…”

“…Another issue that hindered the application of these models in real-life settings is that, very few studies validated their models under sufficient intra-individual BP changes induced by interventions as required in the IEEE Standard 1708. As far as we know, two recent studies tested their deep models under scenarios with induced BP changes, and the mean absolute error of systolic BP significantly degraded compared to those in static conditions, i.e., 14.3 mmHg and 9.5 mmHg for interventions of coffee drinking [19] and water drinking [14] respectively.…”