At the Edge of the Heart: ULP FPGA-Based CNN for On-Device Cardiac Feature Extraction in Smart Health Sensors for Astronauts
Kazi Mohammad Abidur Rahman 1 , Davis Rakhshan 1 , Philipp Lütke 1 , Laura Harms 3,2 , Ulf Kulau 1 · 2026
Read At the Edge of the Heart: ULP FPGA-Based CNN for On-Device Cardiac Feature Extraction in Smart Health Sensors for Astronauts by Kazi Mohammad Abidur Rahman 1 , Davis Rakhshan 1 , Philipp Lütke 1 , Laura Harms 3,2 , Ulf Kulau 1 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-25799.txt.
Introduction
Rahman and four colleagues build an ultra-low-power FPGA system that extracts cardiac features on-device from seismocardiography signals. Both spaceflight and ground monitoring demand real-time features on tiny battery-powered devices. Their design pairs training that accounts for quantization with a systolic-array accelerator for integer-only inference on a Lattice chip. It reaches 98 percent validation accuracy, with each inference under a tenth of a second. The work shows autonomous cardiac monitoring is feasible on very constrained hardware.
Catalog record
| CREATOR | Kazi Mohammad Abidur Rahman 1 , Davis Rakhshan 1 , Philipp Lütke 1 , Laura Harms 3,2 , Ulf Kulau 1 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38354 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.25799
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
At the Edge of the Heart: ULP FPGA-Based CNN for On-Device Cardiac Feature Extraction in Smart Health Sensors for Astronauts Thanks: The biomedical data used in this study were collected with approval from the Ethics Committee of Hamburg University of Technology. Thanks: © 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. Kazi Mohammad Abidur Rahman 1 , Davis Rakhshan 1 , Philipp Lütke 1 , Laura Harms 3,2 , Ulf Kulau 1 Abstract The convergence of accelerating human spaceflight ambitions and critical terrestrial health monitoring demands is driving the unprecedented requirements for reliable and real-time feature extraction capabilities on extremely resource-constraint wearable health sensors. We present a ULP FPGA -based solution for real-time SCG ( SCG ) feature classification using CNN . Our approach combines quantization-aware training with a systolic-array accelerator to enable efficient integer-only inference on the Lattice iCE40UP5K FPGA , which offers an ideal platform for battery-powered deployments – particularly in space environments – for their power efficiency and radiation resilience. The implementation achieves a validation accuracy of 98% while consuming only 8.55 mW of power, completing inference in 95.5 ms with minimal hardware resources (2,861 LUT and 7 DSP blocks). These results demonstrate that fully on-device SCG -based cardiac feature extraction is feasible on resource-constrained hardware, enabling energy-efficient, autonomous health monitoring for astronauts in long-duration space missions. I Introduction As humanity pushes toward sustained exploration of the Moon and Mars, safeguarding astronaut health, one of the most critical priorities in space missions, demands a new generation of intelligent, wearable health monitoring systems capable of operating far beyond LEO ( LEO ). Unlike terrestrial or near-Earth deployments, deep-space missions impose severe communication constraints, where latency to ground stations can extend to minutes, rendering continuous remote supervision infeasible. Consequently, physiological monitoring systems must achieve unprecedented levels of autonomy, reliability, and real-time decision-making to enable timely health assessment and intervention. Building upon insights from previous cardiovascular monitoring experiments conducted aboard the ISS ( ISS ) [ 1 , 2 , 3 ] , this work fosters the development of smart, space-grade wearable sensors specifically designed for continuous cardiac monitoring. SCG ( SCG ), which captures the mechanical vibrations of the heart using inertial sensors, offers a non-invasive and compact sensing modality well suited for long-duration missions. However, robust extraction of clinically meaningful cardiac features from SCG signals, particularly systolic and diastolic intervals, remains challenging due to motion artifacts, physiological variability, and altered cardiac mechanics in microgravity. Recent advances in deep learning have demonstrated that CNN can robustly extract cardiac features from SCG signals by learning hierarchical temporal-spectral representations directly from raw data [ 4 , 5 ] . While CNN enable accurate systolic-diastolic segmentation and feature extraction, their computational complexity presents a significant challenge for deployment on wearable, spaceborne platforms. Deep-space health monitoring systems must operate under extreme constraints: they must be highly miniaturized to meet spacecraft mass and volume limitations and exhibit exceptional energy efficiency to support prolonged autonomous operation. In addition the radiation environment limits available processing components that could be used within this harsh environment. These combined requirements preclude the use of conventional radiation-hardened components, which are typically bulky and power-hungry. To address these constraints, this work adopts a carefully selected COTS ( COTS ) design strategy. A key enabler of this approach is the Lattice iCE40UP-series ULP ( ULP ) FPGA ( FPGA ), which offers milliwatt-level power consumption alongside demonstrated resilience to radiation-induced effects [ 6 ] . This ULP FPGA provides a unique balance between energy efficiency, flexibility, and reliability, making it well suited for continuous, real-time physiological monitoring in space. However, deploying CNN inference on such resource-constrained hardware requires algorithm-architecture co-design to overcome limitations in logic resources, memory capacity, and arithmetic precision. This paper demonstrates that ULP FPGA using co-designed CNN models, quantization-aware training, and a systolic-array accelerator are capable of classifying systolic and diastolic phases in wearable SCG signals accurately and in real-time. The outline of this paper is as follows: Section II reviews related work, followed by short introduction to the space-grade SCG health sensor hardware architecture in Section III . Section IV describes SCG data acquisition and labeling, while Section V presents the CNN design and its optimization for resource-constrained devices. Section VI details the ULP FPGA inference implementation 1 1 1 Implementation code available at: https://github.com/Smart-Sensors-Group/NN_SysArray.git . Section VII evaluates accuracy and performance, and Section VIII concludes the paper. II Related Works Continuous cardiovascular monitoring is essential for astronaut health, as microgravity induces cardiac deconditioning and altered hemodynamics [ 7 ] . Wearable SCG offers a non-invasive, lightweight modality for monitoring cardiac mechanical activity in space; however, reliable analysis beyond HR ( HR ) and HRV ( HRV ) remains challenging due to motion artifacts, inter-cycle variability, sensor placement sensitivity, and microgravity-induced morphological changes [ 8 , 9 ] . Accurate systolic-diastolic segmentation is therefore critical, as key mechanical events such as mitral and aortic valve motions occur in distinct cardiac phases and enable extraction of physiologically meaningful metrics including the SDR ( SDR ), which reflects ventricular ejection-filling balance and is sensitive to myocardial relaxation, coronary perfusion, and hemodynamic load [ 10 , 11 ] . Traditional SCG analysis based on handcrafted features and classical classifiers such as SVM ( SVM ) and KNN ( KNN ) degrades under physiological variability, whereas recent studies show that CNN consistently outperform these approaches by learning hierarchical temporal-spectral representations directly from raw SCG signals [ 4 , 5 ] . Despite their effectiveness, CNN are computationally intensive, motivating hardware acceleration for real-time, on-device inference. Existing FPGA -based CNN accelerators primarily target high-end platforms such as Virtex-7, Zynq, Stratix-V, and Cyclone-V [ 12 , 13 ] . These devices provide extensive DSP ( DSP ) resources, large on-chip memories, and power budgets ranging from 18\text{\,}\mathrm{W} to 50\text{\,}\mathrm{W} [ 14 , 15 ] . While suitable for high-throughput inference, such platforms are incompatible with wearable and ULP health monitoring systems. In contrast, resource-constrained FPGA such as the Lattice iCE40 family operate in the milliwatt range but offer limited logic resources, minimal DSP support, and small memory capacities, rendering conventional CNN mapping strategies impractical. For CNN inference on ultra-constrained devices, hardware capabilities and on-chip storage must be carefully considered, as memory is primarily allocated to model parameters, including weights and biases. Floating-point inference is generally infeasible on ULP FPGA ; therefore, quantization is essential. While post-training quantization can reduce model size, QAT ( QAT ) explicitly models reduced-precision arithmetic during training, enabling accurate integer-only inference at low bit widths. This approach significantly reduces memory footprint and aligns computation with resource-constrained hardware datapaths [ 16 , 17 ] , making it particularly suitable for ULP FPGA -based CNN deployment. Systolic array architectures, introduced by H.T. Kung [ 18 ] , offer an efficient solution for CNN inference on resource-constrained FPGA . Composed of two-dimensional grids of PE with local registers, they enable pipelined data movement