Empirical mode decomposition (EMD) is an adaptive signal analysis method requiring no predefined basis functions, capable of handling nonlinear and non-stationary signals. However, its high computational cost has long prevented real-time operation on embedded systems, particularly at audio frequencies (8–96 kHz). To the best of our knowledge, this study presents the first real-time audio-frequency EMD system running entirely on a commercial microcontroller (ARM Cortex-M7, STM32H730). The system captures audio via a MEMS microphone, decomposes signals into intrinsic mode functions (IMFs) using fast EMD, applies independent gain to each IMF, and reconstructs the output in real time. Systematic evaluation shows that: (1) the maximum processing rate increases approximately in proportion to the MCU clock frequency, achieving 100 kHz (one IMF) or 40 kHz (four IMFs) at 300 MHz; (2) processing latency is only approximately 1.6 ms (16 kHz, 8 ms window, four IMFs + residual), increasing approximately in proportion to window length; and (3) each IMF corresponds to a distinct frequency range, enabling data-driven frequency-selective processing. Built entirely from commercial off-the-shelf components and powered by two 1.5 V batteries, the system provides initial evidence for transitioning fast EMD from offline analysis to real-time embedded signal processing.

Background: Diabetic retinopathy (DR) is the most common microvascular complication of type 2 diabetes mellitus (T2DM), yet DR screening coverage remains inadequate worldwide. Diabetic autonomic neuropathy typically precedes DR, and heart rate variability (HRV) provides a non-invasive autonomic measure; however, whether HRV can screen for DR remains unexplored. Methods: This pilot study investigated whether sleep-onset minute-scale HRV indices—SDANN1 (standard deviation of 1-min segment means of normal-to-normal intervals) and SDANN2 (2-min segments)—serve as candidate screening biomarkers. Using a public 24-h Holter ECG dataset of T2DM inpatients (N = 39; 11 DR-positive, 28 DR-negative), sleep-onset time-points from cardiopulmonary coupling analysis were used to extract 30 min of post-onset ECG. Results: SDANN1 was significantly lower in the DR group (median 23.6 vs. 44.8 ms, p = 0.028, d = 0.74), with an area under the receiver operating characteristic curve of 0.731 (95% CI: 0.530–0.901). At the optimal threshold of ≤ 25 ms, the odds ratio was 8.1 (95% CI: 1.7–38.4), with DR prevalence of 58% in the high-risk versus 12.5% in the low-risk group. Among 12 candidates, only the SDANN family reached significance. Conclusions: sleep-onset minute-scale HRV shows promise as a candidate risk-stratification biomarker for DR; prospective longitudinal validation in larger, sex-balanced cohorts is required before any screening application.