Multi-query frequency prompting for physiological signal domain adaptation

This study addresses three key challenges in blood pressure prediction from physiological signals: limited labeled data, domain shifts between source and target populations, and under-representation of hypertensive and hypotensive groups. We propose Multi-Query Frequency Prompting (MQFP), a prompt-learning framework that inserts a small, learnable embedding, formulated in the frequency domain, into a frozen pre-trained model. By capturing key signal characteristics such as variability, periodicity, and locality, the frequency-space prompt enables robustness to both distributional and temporal shifts while mitigating overfitting in few-shot settings. Evaluated on three PPG and two ECG datasets, MQFP achieves up to a 21.3 % reduction in mean absolute error relative to its backbone, while reducing trainable parameters by up to 98.4 %. These results demonstrate MQFP as a lightweight and effective approach to BP estimation in real-world settings.