Evaluating Reconstruction Methods for Biomedical Signals from Smart Devices

Abstract

Introduction: Continuous monitoring with wearables often yields incomplete biomedical time series because of noise, sensor drop-outs or transmission errors. Reliable reconstruction of the missing information - especially respiratory- and cardiac-waveform morphology - is crucial for downstream clinical interpretation. Methods: Three reconstruction techniques (cubic spline interpolation, Gaussian Process Regression (GPR), autoregressive (AR) modeling) were evaluated on (i) simulated breathing-like signals generated with an undamped oscillator and (ii) real recordings from the piezoelectric VitaLog sensor. Down-sampling factors 1, 2, 4, 8, 16 and 32 simulate sparse sampling. Reconstruction quality was quantified by normalized RMSE percentage (pNRMSE) and the coefficient of determination r ² . Results: Cubic spline interpolation and GPR consistently achieve pNRMSE > 97% and r ² ~ 1.00, whereas AR degrades sharply for DF ≥ 16. Both spline and GPR recover periodic breathing patterns and artefact segments in biomedical piezoelectric data. Conclusion: Cubic spline interpolation and GPR provide robust, low-complexity solutions for biomedical signal reconstruction in telemedical scenarios, enabling reliable analysis even under aggressive sampling reduction.