Towards Real-Time ECG and EMG Modeling on μ \mu NPUs DOI: 10.1145/3774906.3802787 ISBN: 979-8-4007-2309-4/26/05 Conference: ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems; May 11–14, 2026; Saint Malo, France ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys ’26), May 11–14, 2026, Saint Malo, France CCS: Computer systems organization Embedded systems CCS: Applied computing Consumer health CCS: Human-centered computing Human computer interaction (HCI)
Josh Millar, Ashok Samraj Thangarajan, Soumyajit Chatterjee, Hamed Haddadi · 2026
Read Towards Real-Time ECG and EMG Modeling on μ \mu NPUs DOI: 10.1145/3774906.3802787 ISBN: 979-8-4007-2309-4/26/05 Conference: ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems; May 11–14, 2026; Saint Malo, France ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys ’26), May 11–14, 2026, Saint Malo, France CCS: Computer systems organization Embedded systems CCS: Applied computing Consumer health CCS: Human-centered computing Human computer interaction (HCI) by Josh Millar, Ashok Samraj Thangarajan, Soumyajit Chatterjee, Hamed Haddadi (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-2604-18067.txt.
Introduction
Millar and three colleagues introduce PhysioLite, a lightweight model for ECG and EMG analysis on microcontroller-scale neural processing units. Recent Transformer models perform well but are too large and power-hungry for such hardware, and their dynamic attention clashes with micro-NPU constraints. PhysioLite uses learnable wavelet filters, hardware-aware layer design, and CPU-offloaded positional encoding. It matches Transformer models on ECG and EMG benchmarks at under 10 percent of their size, about 370 kilobytes with 8-bit quantization. Tests on two micro-NPUs confirm viability for wearables.
Catalog record
| CREATOR | Josh Millar, Ashok Samraj Thangarajan, Soumyajit Chatterjee, Hamed Haddadi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45670 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.18067
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Towards Real-Time ECG and EMG Modeling on μ \mu NPUs DOI: 10.1145/3774906.3802787 ISBN: 979-8-4007-2309-4/26/05 Conference: ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems; May 11–14, 2026; Saint Malo, France ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys ’26), May 11–14, 2026, Saint Malo, France CCS: Computer systems organization Embedded systems CCS: Applied computing Consumer health CCS: Human-centered computing Human computer interaction (HCI) Josh Millar, Ashok Samraj Thangarajan, Soumyajit Chatterjee, Hamed Haddadi Abstract . The miniaturisation of neural processing units (NPUs) and other low-power accelerators has enabled their integration into microcontroller-scale wearable hardware, supporting near-real-time, offline, and privacy-preserving inference. Yet physiological signal analysis has remained infeasible on such hardware; recent Transformer-based models show state-of-the-art performance but are prohibitively large for resource- and power-constrained hardware and incompatible with µNPUs due to their dynamic attention operations. We introduce PhysioLite , a lightweight, NPU-compatible model architecture and training framework for ECG/EMG signal analysis. Using learnable wavelet filter banks, CPU-offloaded positional encoding, and hardware-aware layer design, PhysioLite reaches performance comparable to state-of-the-art Transformer-based foundation models on ECG and EMG benchmarks, while being <10% of the size ( ∼ \sim 370KB with 8-bit quantization). We also profile its component-wise latency and resource consumption on both the MAX78000 and HX6538 WE2 µNPUs, demonstrating its viability for signal analysis on constrained, battery-powered hardware. We release our model(s) and training framework at: https://github.com/j0shmillar/physiolite . † † cc-license: by 1. Introduction Physiological time-series signals, such as electrocardiography (ECG) and electromyography (EMG), are rich markers of cardiovascular and neuromuscular function, and underpin a wide range of applications, including continuous health monitoring ( 39 ) , early disease detection ( 6 ) , prosthetic control ( 48 ) , and human–computer interaction ( 16 ) . The growing interest in personalized and ubiquitous healthcare has motivated the development of wearable systems capable of directly modeling physiological signals, enabling low-latency feedback, user privacy guarantees, and robust deployment in settings with limited or intermittent connectivity. However, physiological signals acquired in real-world environments are often non-stationary, subject-specific ( 11 ) , and susceptible to noise and motion artifacts ( 23 ) ; effective physiological models must extract and integrate subtle, multi-scale temporal dependencies. Designing such models while satisfying the compute and memory constraints of wearable ( e.g. , microcontroller-class) hardware remains arduous. Recent foundation models for physiological signal modeling have adopted Transformer-based architectures, achieving state-of-the-art performance on ECG and EMG benchmarks by capturing long-range temporal dependencies and multi-scale structure ( 9 ; 27 ; 34 ) . Importantly, foundation-scale models exhibit strong generalization across downstream applications; large-scale pretraining enables the learning of broadly applicable features, which can then be adapted to specific clinical or application-driven tasks via fine-tuning. Other recent work has investigated specialized embedded platforms for biosignal acquisition and processing. BioGAP and BioGAP-Ultra, for example, are modular, NPU-enabled wearable systems designed for multimodal physiological signal processing and inference across signals including ECG, EMG, and EEG ( 18 ; 19 ) . Additionally, open-source RISC-V–based platforms have been studied for medical-grade EMG processing, showing that dedicated custom instruction extensions and tightly integrated accelerators can substantially improve efficiency for embedded biosignal processing ( 24 ) . Such works demonstrate the growing viability of embedded hardware for biosignal workloads, and motivate the use of hardware-efficient model architectures. However, such models typically operate on the scale of tens of millions of parameters and rely on memory- and compute-intensive attention mechanisms, making them impractical to execute on resource-constrained wearables ( 28 ) . Even microcontroller-class neural processing units (µNPUs), while offering hardware acceleration, provide only limited on-chip memory and support for a constrained set of static computation kernels. They do not directly support dynamic attention or large intermediate activation buffers. Consequently, state-of-the-art physiological foundation models rely on server- or mobile-grade hardware, limiting their applicability for continuous, private, and real-time inference on wearable devices. Figure 1. Comparison of avg. F 1 on ECG benchmarks (§ 5.1 ) with PhysioLite and a number of ECG foundation models. We introduce PhysioLite , a lightweight architecture and training paradigm designed to bridge the gap between large ECG/EMG foundation models and microcontroller-class hardware. The key idea is to retain the multi-scale temporal modeling capacity and contextual awareness of wavelet-Transformer models ( 9 ) , while replacing hardware-incompatible, or accelerator-unfriendly, attention mechanisms with structured convolutional alternatives and lightweight positional encoding. Specifically, PhysioLite approximates learnable wavelet decomposition using parallel convolution branches to capture frequency-specific dynamics, and substitutes attention blocks with compact feature-mixing modules that rely on pointwise convolutions. Positional information is externally incorporated on the CPU prior to accelerator execution, ensuring temporal context is maintained without requiring dynamic operations on the accelerator. This design mirrors the functional behavior of attention-based foundation models while remaining compatible with µNPU operator sets and strict memory constraints, enabling local inference under tight memory and power budgets. To assess the effectiveness of our design, we evaluate PhysioLite across a suite of ECG and EMG benchmark datasets commonly used for supervised classification and arrhythmia or gesture recognition tasks. Compared to state-of-the-art ECG and EMG foundation models ( 9 ) , PhysioLite remains competitive while staying under a 512KB memory budget ( \sim 370KB with 8-bit quantization, <0.3M parameters), as shown in Figure 1 . On PTB-XL ( 52 ) , CPSC ( 30 ) , and Chapman–Shaoxing ( 54 ) ECG benchmarks, PhysioLite achieves AUROC of 0.893–0.994 and F 1 of 0.600–0.942, nearing foundation-scale models ( 27 ; 34 ) . On Ninapro DB5 ( 40 ) , EPN-612 ( 7 ) , and UCI ( 26 ) sEMG gesture benchmarks, PhysioLite achieves F 1 scores of 0.738–0.949 and AUROC of 0.994–0.996, matching or exceeding prior foundation models while using substantially fewer parameters, demonstrating that key multi-scale physiological structures can be captured without RoPE attention ( 46 ) , or other attention mechanisms. Finally, through end-to-end deployment and component-wise profiling on the MAX78000 ( 33 ) and Himax HX6538 WE2 ( 50 ) µNPUs, we show that PhysioLite fits within on-chip memory limits and achieves near real-time latency (sub \sim 20 ms), confirming its suitability for ECG and EMG monitoring on such hardware. To summarize, our key contributions are: (1) We detail PhysioLite , a lightweight and µNPU-compatible model architecture and training framework for general-purpose ECG and EMG processing, designed to preserve the representational strengths of Transformer-based foundation models, and release our pretrained model(s) and framework as open-source: https://github.com/j0shmillar/physiolite . (2) We conduct a comprehensive evaluation of our method across multiple ECG and EMG benchmarks, showing that PhysioLite achieves performance comparable to state-of-the-art foundation models while reducing model size and compute footprint by approximately an order of magnitude. (3) We perform detailed system-level profiling on the MAX78000 and Himax WE2, µNPU-integrated microcontrollers designed for battery-powered wearables. We quantify latency, memory usage, and power consumption, across various processing stages, and demonstrate near-real-time inference latency (sub \sim 20 ms) suitable for continuous monitoring. 2. Related Work and Background 2.1. µNPU Hardware Constraints NPUs are specialized hardware accelerators designed for executing neural networks. In the last few years, low-power microcontroller-scale variants, µNPUs, have emerged, supporting neural acceleration on battery-powered and wearable hardware. Fig. 2 , below, outlines the general design and layout of a µNPU, composed of a systolic array