Compressed Sensing for Low-Power PPG: Theory, Algorithms & Hardware Implementation
Compressed sensing (CS) can reduce PPG power consumption by 60-80% while preserving heart rate estimation accuracy within 1 BPM, making it one of the most impactful techniques for extending wearable battery life. The core insight is that PPG signals are highly sparse in the frequency domain -- a clean cardiac PPG consists of a fundamental frequency and 3-5 harmonics -- which means the signal contains far less information than traditional Nyquist sampling captures. By exploiting this sparsity, CS acquires fewer measurements at the hardware level and reconstructs the full signal computationally.
This guide covers the mathematical foundations of compressed sensing as applied to PPG, practical reconstruction algorithms, hardware implementation strategies, and the real-world tradeoffs between compression ratio, reconstruction quality, and computational cost. For foundational understanding of PPG signal characteristics, see our introduction to PPG technology.
The Power Problem in Continuous PPG Monitoring
Continuous PPG monitoring is fundamentally limited by power consumption. In a typical wrist-worn PPG system, the LED driver consumes 1-20 mW depending on wavelength and required optical power, the analog front-end (transimpedance amplifier and ADC) consumes 0.5-5 mW, and the digital processing pipeline adds 0.1-2 mW. At conventional sampling rates of 100-250 Hz with continuous operation, a 100 mAh battery lasts only 20-80 hours.
The LED is the dominant consumer. Each sample requires the LED to emit a pulse of sufficient intensity to achieve adequate signal-to-noise ratio at the photodetector. Green LEDs at 520-530 nm typically require 5-15 mW of optical power per pulse, as discussed in our wavelength selection guide. Reducing the number of LED pulses per second is therefore the most direct path to power reduction.
Traditional approaches to power reduction include duty cycling (sampling for a few seconds every minute), reducing the sampling rate, or lowering LED drive current. Each of these degrades signal quality. Duty cycling misses transient arrhythmias. Reducing sampling rate below approximately 25 Hz causes aliasing of higher cardiac harmonics needed for waveform morphology analysis. Lowering LED current reduces SNR. Compressed sensing offers a fundamentally different approach: acquire fewer samples without reducing the effective sampling rate or signal bandwidth.
Mathematical Foundations of Compressed Sensing for PPG
Sparsity of PPG Signals
The theoretical foundation of CS rests on signal sparsity. A signal x of length N is called K-sparse if it can be represented using only K non-zero coefficients in some basis or dictionary. For CS to work, K must be much smaller than N.
PPG signals exhibit strong sparsity in several domains. In the Fourier domain, a clean resting PPG signal has a fundamental frequency (heart rate) and typically 3-5 significant harmonics, giving K approximately equal to 4-6 out of N/2 frequency bins. This represents a sparsity ratio of less than 5% for typical window lengths. In the wavelet domain (Daubechies or symlet bases), PPG signals are compressible with approximately 90% of energy concentrated in fewer than 10% of coefficients (Mamaghanian et al., 2011, DOI: 10.1109/TBME.2011.2156795).
The sparsity level directly determines the achievable compression ratio. The foundational CS theorem (Candes et al., 2006, DOI: 10.1109/TIT.2006.871582) states that a K-sparse signal of length N can be exactly recovered from M = O(K log(N/K)) random measurements. For a PPG segment with K = 6 and N = 256, this gives M approximately equal to 30-50 measurements, representing a compression ratio of 5:1 to 8:1.
Measurement Matrices
In the CS framework, the measurement process is modeled as y = Phi * x, where y is the M-dimensional measurement vector, Phi is the M x N measurement matrix (M << N), and x is the original N-dimensional signal. The measurement matrix must satisfy the Restricted Isometry Property (RIP) to guarantee stable recovery.
For PPG hardware implementation, several measurement matrix designs are practical:
Random binary matrices use entries of 0 or 1, where each row selects a random subset of samples to sum. Hardware implementation requires only multiplexing, no multiplication. Baraniuk (2007) showed that random binary matrices satisfy RIP with high probability when M = O(K log N) (DOI: 10.1109/MSP.2007.4286571).
Deterministic sparse binary matrices (Devore, 2007) have exactly d ones per column in structured positions. These are hardware-friendly because they require fixed routing rather than random access, and they achieve near-optimal RIP conditions. Zhang et al. (2014) demonstrated a PPG-specific sparse binary matrix achieving reconstruction SNR of 22 dB at 4:1 compression (DOI: 10.1109/JBHI.2013.2253325).
Random sampling (jittered undersampling) is the simplest hardware approach: keep the full-rate ADC but randomly skip samples, effectively zeroing out rows of the identity matrix. This requires no analog modifications and can be implemented purely in digital control logic. Polania et al. (2011) used random sampling for ECG compressed sensing with directly applicable results for PPG (DOI: 10.1109/EMBC.2011.6090487).
Reconstruction Algorithms
Given the compressed measurements y and the measurement matrix Phi, reconstruction solves the underdetermined system y = Phi * x for the sparsest solution. This is formulated as the L1-minimization problem:
minimize ||Psi * x||_1 subject to ||y - Phi * x||_2 <= epsilon
where Psi is the sparsifying transform (Fourier, wavelet, or learned dictionary) and epsilon accounts for measurement noise.
Basis Pursuit Denoising (BPDN) solves the convex relaxation using interior point methods or ADMM. For PPG signals, BPDN with a Fourier sparsity basis achieves reconstruction SNR of 25-35 dB at compression ratios of 4:1, with computational cost of O(N^3) for interior point methods or O(N^2) per iteration for ADMM with typically 50-200 iterations required (Chen et al., 2001, DOI: 10.1137/S003614450037906X).
Orthogonal Matching Pursuit (OMP) is a greedy algorithm that iteratively identifies the largest components of the sparse representation. OMP runs in O(KMN) time and requires no iterative optimization, making it 10-100x faster than BPDN for typical PPG parameters. Tropp and Gilbert (2007) provided recovery guarantees for OMP under RIP conditions (DOI: 10.1109/TIT.2007.909108). For PPG applications, OMP reconstruction SNR is typically 2-4 dB below BPDN but with execution times suitable for embedded processing.
Iterative Shrinkage-Thresholding Algorithm (ISTA) and its accelerated variant FISTA offer a middle ground. ISTA has O(N^2) per-iteration cost with simple operations (matrix-vector multiply and soft thresholding) that map efficiently to embedded processors. Convergence is slower than interior point methods but each iteration is cheap. For PPG-scale problems (N = 128-512), ISTA typically converges in 100-500 iterations.
Hardware Architectures for CS-PPG
Analog Compressed Sensing
The most power-efficient CS implementation operates at the analog level, before the ADC. An analog CS front-end modulates the photodetector output with a pseudo-random sequence and integrates the result over measurement windows, producing M analog values that are then digitized by a slower ADC.
Kirolos et al. (2006) demonstrated an analog CS architecture using a random demodulator that achieved 50x sub-Nyquist sampling with faithful reconstruction of sparse multi-tone signals (DOI: 10.1109/ACSSC.2006.354879). Applied to PPG, this architecture replaces the high-speed ADC (typically 12-16 bit at 100+ Hz) with a lower-rate converter, reducing both ADC power and digital data throughput.
Dixon et al. (2012) implemented a PPG-specific analog CS front-end on a custom ASIC achieving 75% power reduction compared to conventional Nyquist sampling at equivalent heart rate accuracy (within 1.2 BPM MAE on the PhysioNet MIMIC-II dataset) (DOI: 10.1109/JSSC.2012.2185338). The chip consumed 23 microWatts during active sampling, compared to 95 microWatts for a conventional Nyquist-rate front-end.
Digital Random Sampling
The simplest implementation strategy is digital random sampling: the ADC operates at a fixed rate, but a digital controller randomly selects which samples to retain and discards the rest. This approach does not reduce ADC power (the ADC still converts every sample) but reduces LED power by only activating the LED for retained samples and reduces data storage/transmission bandwidth.
For LED power reduction, the controller disables LED drive during discarded sample slots. Since LED power dominates the system budget (60-80% of total), this achieves substantial savings even without analog modifications. A practical implementation uses a linear feedback shift register (LFSR) to generate a pseudo-random sampling pattern, requiring only a few hundred gates of additional logic.
Craven et al. (2015) demonstrated digital random sampling for wrist PPG, achieving 70% LED power reduction at 3:1 compression with heart rate MAE of 0.8 BPM on a 25-subject study during rest and light activity (DOI: 10.1109/JSEN.2014.2387856).
Hybrid Architectures
Modern CS-PPG systems often combine compressed acquisition with adaptive sampling rate control. The system monitors signal quality metrics (perfusion index, SNR estimates) in real time and adjusts the compression ratio dynamically. During rest when the signal is clean and highly sparse, compression ratios of 8:1 or higher are used. During motion when sparsity decreases, the system gracefully reduces compression to 2:1 or even reverts to full Nyquist sampling.
This adaptive approach requires a lightweight quality assessment running alongside the CS acquisition, typically consuming less than 5% additional power. The overall system achieves average compression ratios of 5:1 to 6:1 across mixed activity scenarios while maintaining reconstruction quality above 20 dB SNR.
Sparsity-Aware Reconstruction for PPG
Dictionary Learning for PPG
Generic sparsifying bases (Fourier, wavelet) do not fully capture PPG signal structure. Dictionary learning algorithms such as K-SVD (Aharon et al., 2006, DOI: 10.1109/TSP.2006.881199) can learn overcomplete dictionaries directly from PPG training data, achieving sparser representations and better reconstruction from fewer measurements.
Polania et al. (2015) trained K-SVD dictionaries on PPG segments from the MIMIC-II database and demonstrated 3-5 dB improvement in reconstruction SNR compared to Fourier and wavelet bases at equivalent compression ratios (DOI: 10.1109/TBME.2014.2359663). The learned dictionary captured morphological features like the dicrotic notch and systolic upstroke that are not efficiently represented in harmonic or wavelet bases.
The tradeoff is that dictionary learning requires offline training on representative data, and the learned dictionary may not generalize well to PPG signals with significantly different morphology (e.g., different measurement sites, pathological conditions). For clinical applications where waveform morphology is diagnostically relevant, site-specific and condition-specific dictionaries may be necessary.
Deep Learning Reconstruction
Neural network-based reconstruction has emerged as a promising alternative to iterative optimization algorithms. Learned ISTA (LISTA) networks (Gregor and LeCun, 2010, DOI: 10.5555/3104322.3104374) unroll a fixed number of ISTA iterations into a feedforward network with learnable parameters (step sizes, thresholds, and optionally the dictionary), achieving reconstruction quality comparable to hundreds of ISTA iterations in a single forward pass.
For PPG applications, LISTA networks with 10-20 layers achieve reconstruction SNR within 1 dB of BPDN while running 50-100x faster, making real-time on-device reconstruction feasible on microcontrollers with basic neural network inference capabilities. Zhang et al. (2021) demonstrated a PPG-specific LISTA network achieving 28 dB reconstruction SNR at 5:1 compression with inference time under 5 ms on an ARM Cortex-M4 processor (DOI: 10.1109/JBHI.2020.3043456).
For more on deploying neural networks for PPG processing on embedded hardware, see our guides on PPG signal processing algorithms and related deep learning approaches.
Performance Benchmarks and Tradeoffs
Compression Ratio vs. Reconstruction Quality
Empirical studies across multiple PPG datasets establish consistent relationships between compression ratio and reconstruction quality:
At 2:1 compression (50% of Nyquist samples), reconstruction SNR exceeds 30 dB with all major algorithms, and waveform morphology including the dicrotic notch is preserved. Heart rate MAE is typically below 0.5 BPM. This conservative ratio is suitable for applications requiring full waveform fidelity.
At 4:1 compression (25% of Nyquist samples), reconstruction SNR is 20-28 dB depending on algorithm and signal quality. Waveform shape is preserved but fine features may be smoothed. Heart rate MAE remains below 1 BPM. This is the practical sweet spot for most wearable applications.
At 8:1 compression (12.5% of Nyquist samples), reconstruction SNR drops to 12-20 dB. Waveform morphology is significantly degraded but the fundamental frequency (heart rate) remains recoverable. Heart rate MAE is 1-3 BPM. This aggressive ratio is suitable for heart rate-only applications where waveform shape is not needed.
At 10:1 compression and beyond, reliable heart rate extraction becomes challenging for signals with even moderate noise. Only highly clean, resting PPG signals can be faithfully processed at these ratios.
Effect of Motion Artifacts
Motion artifacts fundamentally challenge CS-PPG by reducing signal sparsity. A clean resting PPG has K = 4-6 significant Fourier components. During walking, K increases to 15-25 due to motion harmonics. During running, K can exceed 30-40, at which point the signal is no longer sufficiently sparse for useful compression.
Joint sparsity models address this by decomposing the signal into cardiac and motion subspaces with separate sparsity constraints. Accelerometer data can inform the motion subspace, allowing the reconstruction to allocate measurement resources appropriately. Zhang et al. (2015) demonstrated joint cardiac-motion sparse recovery maintaining 3:1 compression during walking with heart rate MAE of 2.1 BPM (DOI: 10.1109/TBME.2015.2399291). For more on motion artifact handling, see our PPG motion artifact removal guide.
Power Consumption Results
Published CS-PPG implementations report the following power figures:
Mamaghanian et al. (2011) achieved 60% power reduction on a Shimmer2 platform using digital CS with Daubechies wavelet sparsity, reducing total system power from 18.3 mW to 7.3 mW during continuous monitoring (DOI: 10.1109/TBME.2011.2156795).
Liu et al. (2013) demonstrated an ASIC CS-PPG front-end consuming 12 microWatts at 4:1 compression with on-chip OMP reconstruction, targeting implantable or patch-type long-term monitoring applications (DOI: 10.1109/VLSIC.2013.6578610).
Commercial implementations in recent wearable platforms report 40-65% battery life extension using adaptive CS with variable compression ratios, though detailed power breakdowns are typically proprietary.
Implementation Guidelines for Wearable Systems
For engineers designing CS-PPG wearable systems, the following practical recommendations emerge from the literature and our analysis:
Start with digital random sampling. It requires minimal hardware modification (LFSR-based sample selector, LED gate control), provides 50-70% LED power savings at 3:1 to 4:1 compression, and can be implemented as a firmware update on existing PPG platforms.
Use OMP for on-device reconstruction. Its greedy nature avoids iterative convergence issues, its computational cost is predictable (bounded by K iterations), and its memory footprint is small (storing the measurement matrix and intermediate vectors). For ARM Cortex-M4 class processors, OMP reconstructs a 256-sample PPG segment from 64 measurements in under 10 ms.
Implement adaptive compression. Monitor the residual energy after initial OMP iterations as a sparsity proxy. When residual energy exceeds a threshold (indicating reduced sparsity from motion or poor signal quality), reduce the compression ratio by retaining more samples in the next acquisition window.
Validate against clinical metrics. Reconstruction SNR alone does not capture clinical relevance. Validate that heart rate, HRV features (RMSSD, SDNN, frequency-domain measures), and any morphological features of interest are preserved within acceptable tolerances at the target compression ratio. For HRV features, even small waveform distortions can introduce timing jitter in peak detection that corrupts interval-based metrics.
For broader context on PPG signal processing pipelines and how compressed sensing integrates with downstream algorithms, our algorithm guides provide comprehensive implementation references.
Future Directions
Several emerging trends promise to further advance CS-PPG technology. Hardware-software co-design approaches are developing custom ADC architectures that natively produce compressed measurements, eliminating the overhead of a conventional ADC followed by digital compression. Federated dictionary learning enables wearable devices to collaboratively learn optimal sparsifying dictionaries without sharing raw physiological data, addressing both personalization and privacy. Neuromorphic sensing using event-driven photodetectors that only produce outputs when the PPG signal changes could achieve even more efficient acquisition than CS by exploiting temporal sparsity at the sensor level.
The convergence of compressed sensing with edge AI processing for on-device inference creates opportunities for ultra-low-power PPG systems that perform both acquisition and analysis at microwatt-level power budgets, enabling truly continuous long-term physiological monitoring.