Prediction of arterial blood pressure waveforms based on Multi-Task learning

Background: Continuous and regular blood pressure (BP) monitoring has great significance for the prevention and treatment of cardiovascular diseases. Arterial blood pressure (ABP) waveforms contain instantaneous changes in BP values. Existing ABP prediction tasks mostly aim at single-target predictions, without considering the correlation between different tasks. There are still many challenges in waveform generation models between individuals. Method: In this paper, a two-stage multi-task learning network (ABPMTL) is proposed to estimate ABP waveforms with the input of photoplethysmogram (PPG) and electrocardiogram (ECG) signals. The first stage is a classification task, and Resnet18 combined domain adversarial network is trained to generate class labels, which is regarded as the auxiliary input in the next stage. The second stage contains two branch tasks:(1) BP value perdition and (2) ABP waveforms generation. A dual attention-based task consistency learning block (TCL) is introduced to ensure hierarchical feature sharing between two tasks while preserving specificity simultaneously. Results: For ABP waveforms generation, the mean absolute error (MAE) of results predicted by ABPMTL reaches 7.10 mmHg in subject-independent manners and 2.89 mmHg after fine-tuning. For isolated BP value prediction, the results achieve Grade A according to the BHS standard. Significance: The proposed method considers the correlation of features between different BP tasks for the first time, and achieves great performance in both BP value prediction and ABP waveform generation.