This is the authors' abstract. We don't add key points for this paper.
Non-contact vital sign monitoring using commodity Wi-Fi Channel State Information (CSI) provides a privacy-preserving and unobtrusive solution for ambient healthcare. However, extracting subtle cardiac micro-vibrations from dominant respiratory movements and environmental multipath noise remains a major technical challenge. Existing studies are limited by small participant cohorts and suffer from temporal data leakage. This article presents an edge-native non-contact heart rate monitoring and state classification framework using ultra-low-cost ESP32 Wi-Fi microcontrollers. Our preprocessing pipeline applies phase sanitation and a 4th-order Butterworth bandpass filter (0.8–2.5 Hz) to isolate cardiac signals. We evaluate nine machine learning and deep learning algorithms across 250 human participants and 24 distinct physical scenarios. These scenarios systematically evaluate physical distance (1 m to 5 m), body postures, through-wall Non-Line-of-Sight (NLoS) barriers, and dual-subject interference. Evaluating on a strict holdout test set of 50 unseen participants yields an average classification accuracy of 88.4% and a Mean Absolute Error of 3.8 bpm for Random Forest, while a hybrid CNN-LSTM model achieves 90.8% accuracy. To achieve true system autonomy, the optimized Random Forest model is translated into C++ arrays using emlearn and compiled directly into ESP32 firmware. The embedded model operates within a 182 KB Flash footprint and 42.4 KB SRAM, achieving a low onboard inference latency of 48.2 ms per window. This work demonstrates that ultra-low-cost commodity microcontrollers can perform continuous, real-time cardiac tracking without server dependency.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Solgun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: