Improving decoders for ECG compressed sensing by means of time-prediction

The proliferation of low-power sensor networks and the growing demand for data-driven intelligence have made efficient, reliable transmission of timeseries data a central challenge. Since wireless communication is typically the most energy-expensive task of a sensor node, reducing the amount of transmitted information through compression is crucial to improve energy efficiency and scalability. Compressive Sensing (CS) offers a lightweight encoding strategy suitable for such energy-constrained devices, but standard CS treats each time frame independently and does not exploit transferable temporal structure learned from data. This thesis introduces predictive regularization, a principled framework that integrates pretrained foundational forecasters into the CS decoding step to improve reconstruction accuracy and stability in streaming settings. We adopt a large, pretrained forecasting model (Chronos) as a data-driven prior and incorporate its prediction into the recovery objective as a soft regularizer. By biasing the decoder toward temporally coherent explanations of the measurements while preserving the ability to override poor predictions, predictive regularization improves robustness in low-information and high-noise regimes and naturally accommodates nonlinear, nonstationary dynamics. The approach is implemented and evaluated on two complementary benchmarks: controlled synthetic signals (to probe fundamental properties and failure modes) and a realistic ECG waveform testbed (to validate practical utility). Reconstruction quality is measured using Average Reconstruction Signal-to-Noise Ratio (ARSNR) across a variety of compression ratios and noise levels. The main contribution of this work is to demonstrate that forecasting models can serve as effective temporal priors in compressed sensing. We develop and analyze predictive-regularization objectives, explore practical aspects and provide extensive empirical evidence of their benefits on synthetic and biomedical signals. The results highlight consistent improvements—particularly under extreme compression—showing that data-driven forecasting priors open a promising path toward more accurate and efficient dynamic compressed sensing, with applications in biomedical monitoring, structural health assessment, and other sensorized domains.