Multitask deep label distribution learning for blood pressure prediction

Abstract

Cuffless continuous blood pressure (BP) monitoring is of vital importance for personal health management. Currently, there are extensive studies devoted to cuffless BP prediction based on advanced machine learning techniques and by fusing a variety of physiological signals such as Photoplethysmogram (PPG) and Electrocardiogram (ECG) signals. However, the prediction accuracy still cannot meet the requirements, and it is inconvenient to collect multiple signals at the cost of additional sensors, which limits its potential application scenarios. Different from the conventional routine of modeling BP prediction as a classification or regression question, we model BP prediction as a label distribution learning question (sample level information fusion) for the first time and an end-to-end model is trained based on the proposed adaptive multitask weighted loss to predict systolic BP (SBP), diastolic BP (DBP) and mean BP (MBP) in parallel (task level information fusion), with only PPG signal as input. Resultly, not only precise BP but also predictive confidence interval are reported, and the normalize target technique usually used in regression modeling is no longer needed. To fully delve useful information for BP prediction from the only PPG signal, an end-to-end network is proposed for learning and fusing information from different modalities (original signal and its derivatives, time domain and time–frequency domain) of the signal (feature fusion). Besides, taking into account the varying informativeness of each learned feature accounting for different prediction tasks, task-specific attention module is introduced to learn the varied importance of each feature learned to different prediction tasks, under the hard parameter sharing mode of multitask learning (MTL) network. Extensive experiments on a publicly available database indicate that: (1) The proposed MTL model achieves superior performance over the corresponding single-task learning (STL) model at the cost of only about 1/3 times the amount of parameters. (2) The distribution learning mode enables superior generalization ability of the model over the regression modeling mode in both MTL and STL settings. (3) Compared with regression modeling, the distribution learning mode can alleviates the predictive bias of the trained model due to skewed distribution in dataset, given TFNet as feature learner. (4) The fusion of information of different modalities of PPG signal can significantly improve the generalization ability of the prediction model. (5) The proposed model has achieved superior performance over several representative methods/systems, while using only PPG signal and no any calibration procedure is required.