With the continuous advancement of wearable devices, mobile health pulsative physiological signals (mHealth PPS) have transformed real-time medical monitoring. However, signal loss caused by improper device use and system instability remains a critical challenge. To address this issue, we propose DiPPSI (Diffusion-based Pulsative Physiological Signal Imputation), a novel framework that combines personalized pulse pattern learning with activity-aware signal reconstruction. DiPPSI introduces three key innovations: (1) a dual-phase pulse pattern function that extracts individual-specific templates and generates rhythmic sequences, (2) seamless integration of tri-axial accelerometer data to capture user activity context, and (3) a conditional diffusion model that leverages both pulse pattern and activity information for robust imputation. Extensive experiments on three real-world datasets, MIMIC-III, ScientISST MOVE, and WSAM, demonstrate DiPPSI’s superiority over state-of-the-art methods. The framework achieves up to a 43.56% reduction in Mean Squared Error (MSE) for long-duration missing segments, a 33.12% improvement in F1 score for pattern reconstruction, and consistently superior performance across both ECG and PPG signals. These results validate DiPPSI’s effectiveness in addressing real-world physiological signal imputation challenges, particularly for extended missing periods and varying activity conditions. This marks a significant advancement in continuous health monitoring technology.