ECG Signal to Spectrogram HEX Conversion

Proposed :

Electrocardiography (ECG) is the principal clinical tool used to record the electrical activity of the heart over time. ECG-based arrhythmia classification has historically relied on time-domain analysis, requiring accurate detection of the PQRST complexes. Recent literature, however, has demonstrated that converting an ECG segment into its time-frequency representation (the spectrogram) and treating arrhythmia classification as an image-classification problem can yield significantly higher accuracy (over 99%) using Convolutional Neural Networks (CNN). This image-based approach removes the need for explicit peak detection and is therefore better suited to hardware (VLSI) realization, where a fixed-size 2D image can be processed by a CNN pipeline with deterministic memory access.

This work develops a MATLAB GUI that bridges the gap between the raw ECG time-series and its hardware-ready representation. The GUI accepts an ECG .txt file (in MIT-BIH Arrhythmia Database format), allows the user to scroll through the waveform and visually select a region of interest, generates a 256×256 RGB spectrogram of the selected region, and exports the image to three HEX files corresponding to the R, G, and B channels. These HEX files are directly usable as ROM/RAM initialization data in a Verilog/VHDL testbench. The GUI also provides a reverse capability: given the three HEX files, the spectrogram image is rebuilt and a pseudo-ECG signal is synthesized using the Fast Griffin-Lim phase reconstruction algorithm followed by a custom Inverse Short-Time Fourier Transform.

The reverse operation is included for verification, debugging, and demonstration purposes: it allows a hardware designer to confirm that the HEX data written to memory still carries the expected time-frequency information by visually inspecting the synthesized signal. It is important to note that exact ECG reconstruction is mathematically impossible from a magnitude-only spectrogram image because the Fourier phase is discarded; the Griffin-Lim algorithm provides only an approximation.