MultitaskBP: Reliable PPG-Based Blood Pressure Measurement Under Motion Artifact Corruption

Photoplethysmography (PPG)-based continuous non-invasive blood pressure (NIBP) measurement under wearable health monitoring scenario suffers from two major challenges. First, low reliability due to motion artifact (MA) and missing data due to loose contact between sensor and skin interface. Secondly, often long-term feature dependencies over cardiac cycles cannot be addressed due to limited computational resources. This work mitigates these concerns using a new transformer encoder architecture, named, MultitaskBP, with hierarchical attention for beat-to-beat (BtB) NIBP measurement. The model simultaneously performs three key tasks: signal quality assessment (SQA), reconstruction of missing or corrupted signal segments, and estimation of BtB systolic and diastolic BP (SBP and DBP). By integrating both local intrabeat and global interbeat features, the model effectively captures contextual and temporal dependencies across multiple PPG beats. Unlike conventional deep networks like CNN, RNN, or TCN- based methods that depend on predefined feature extraction or suffer from vanishing gradients, this model learns task-relevant features directly from raw data without manual intervention. It was evaluated on 250 records (approx. 25,000 cardiac cycles for blind test) from the PhysioNet MIMIC-II/III waveform database, achieving mean absolute errors (MAE) of 2.17 mmHg for SBP and 1.20 mmHg for DBP. To ensure low resource embedding, the algorithm was implemented in ARM Cortex A72 controller with a 25-s PPG window, resulting end-to- end task latency of 0.93 s, confirming suitability for deployment in real-time NIBP measurement. The model shows reliable performance even on noisy or inconsistent data, reinforcing its robustness and potential for continuous monitoring in wearable and mobile healthcare systems.