itsSHAS/dora-nano-ecg

DoraNano ECG: An Embedded Brain-Inspired Neural Model for Real-Time Cardiac Monitoring

DoraNano ECG is an embedded inference model for real-time electrocardiogram analysis, developed within the DORA (Dual Ontogenic Resonance Architecture) neural framework.

The model is designed for resource-constrained microcontrollers and implantable medical hardware. It executes deterministically in 18,280 FLOPs per cardiac cycle with a compiled footprint of 18.5 KB and zero dynamic heap allocations (#![no_std]).

• GitHub Repository & Evaluation Harness: https://github.com/Ashixi/dora-nano-ecg
• Author: Andrii Shumko (Prague, Czech Republic)

---

Architecture Overview

• D — Dual: Coupled baseline tracking and sparse resonance memory dynamics.
• O — Ontogenic: Dynamic topological self-organization starting from a single somatic unit.
• R — Resonance: Phase-space waveform alignment in bounded state space $[-1, 1]$ without dense matrix operations.
• A — Architecture: Unified neural framework for continuous temporal signals.

---

Technical Specifications

Parameter	Specification	Verification Context
Binary Footprint	18.5 KB	Measured (LLVM/MSVC .text + .rdata, zero heap allocations)
Inference Latency	3.33 µs / beat	Measured on host x86_64 @ 3.4 GHz
Algorithmic Complexity	18,280 FLOPs / beat	Deterministic arithmetic operation count
MCU Execution Timing	~500 µs (~32,000 cycles)	Analytical model: ARM Cortex-M4 @ 64 MHz with FPU
CPU Energy Budget	~10.62 µJ / cycle	Analytical model: enables 99.94% Deep Sleep / WFI duty cycle

---

Benchmark Results (MIT-BIH DS2 Protocol)

Evaluated under the inter-patient standard across 22 unseen test patients (44,168 beats):

• Ventricular Ectopic Recall (Sensitivity): 98.01% (2,855 / 2,913 events)
• Normal Sinus Rhythm Specificity: 98.27% (38,555 / 39,234 normal beats)
• Inappropriate Defibrillator Shocks on Normal Heart: 0 / 39,234 (0.0000%)
• Normal Heart ICD Safety Score: 100.0000%