Objective This study proposes a task-oriented signal compression paradigm for non-invasive blood pressure (BP) prediction, eliminating manual feature engineering through resampling to achieve a high-performance lightweight model suitable for embedded systems.

Methods The PDA-1 pulse diagnostic device equipped with a piezoelectric sensor was used to collect radial artery pulse waves from healthy volunteers at Shuguang Hospital Affiliated to Shanghai University of Traditional Chinese Medicine between November 20, 2021 and November 19, 2022. The most stable 8 s segment containing multiple complete pulse cycles was extracted from each 30 s recording. Original signals (200 Hz) were resampled to 40 Hz using Fourier transform, wavelet transform (db9), and cubic spline interpolation. Three feature sets were constructed: baseline (resampled data and demographics), fusion (baseline plus handcrafted time-domain features), and handcrafted (handcrafted features and demographics) feature sets. XGBoost and random forest (RF) models were developed with Bayesian optimization for hyperparameter tuning, and evaluated using 10-fold cross-validation. Prediction accuracy was assessed against British Hypertension Society (BHS) and Association for the Advancement of Medical Instrumentation (AAMI) standards, while reconstruction quality was quantified by percentage root-mean-square difference (PRD). Model evaluation and multiple comparisons among different configurations were conducted using the corresponding statistical methods. Visual and interpretability analyses were performed using SHapley Additive exPlanations (SHAP), frequency-energy distribution analysis, and wavelet multi-resolution decomposition.

Results A total of 348 healthy volunteers were enrolled. XGBoost and RF models built on resampled data achieved clinically competitive performance, with XGBoost demonstrating superior accuracy and computational efficiency. The XGBoost model employing Fourier transform resampling and the baseline feature set achieved a mean absolute error (MAE) of (6.01 ± 0.94) mmHg for systolic blood pressure (SBP) and (4.60 ± 0.34) mmHg for diastolic blood pressure (DBP) on the test set, meeting BHS Grade B for SBP and BHS Grade A for DBP, as well as AAMI standards. No statistically significant differences in predictive accuracy were found among the three resampling methods (P > 0.05); however, their reconstruction errors differed markedly: wavelet transform exhibited the highest error (PRD = 35.52%), followed by Fourier transform (PRD = 8.87%), and interpolation the lowest (PRD = 1.33%). Compared with the baseline feature set, the fused feature set reduced SBP MAE by only 0.09 mmHg (corrected P > 0.05) while increasing XGBoost training time by 92.5% − 247.5% (P < 0.001). The handcrafted feature set achieved comparable predictive performance to the baseline, indicating that manual feature engineering is redundant when resampled signals are employed.

Conclusion This study validated the sufficiency of compressed signals for clinical-grade prediction and the redundancy of handcrafted feature engineering. Integrating task-oriented compression with machine learning may support the development of lightweight wearable continuous BP monitoring devices.