Hardware implementation of a Pantompkins-style ECG processing pipeline with R-peak detection, RR estimation, and morphological feature extraction. The top-level module is pantompkins.sv.

• rtl/ SystemVerilog RTL blocks (filters, R-peak detection, morph features/descriptor).
• tb/ Testbench top wrapper.
• simulation/ Simulation scripts and Makefile targets for Questa and Verilator.
• synth/ Quartus synthesis scripts/targets.
• utils/ Helper scripts (e.g., dataset conversion).
• utils/verif/ Verification scripts for comparing R-peak outputs vs MIT-BIH annotations.
• data/ Input datasets (e.g., MIT-BIH).
• Makefile Top-level unified make targets.
• requirements.txt Python dependencies for data conversion.

rtl/pantompkins.sv

Inputs:

• clk System clock.
• clk_enable Sample enable.
• reset Active-high reset.
• filter_in[11:0] ECG sample input.

Outputs:

• morph_p_q8_23[31:0], morph_q_q8_23[31:0], morph_s_q8_23[31:0], morph_t_q8_23[31:0] Normalized morph descriptor outputs.
• morph_r_time[31:0] R-peak time associated with the morph descriptor.
• pre_rr_interval[31:0], post_rr_interval[31:0], local_rr_interval[31:0], global_rr_interval[31:0] RR interval features.
• feat_valid Valid strobe when RR + morph outputs are aligned.

Internal pipeline (high level):

1. Bandpass/filtering (filter)
2. Differentiation (differentiator)
3. Squaring
4. Moving average (lowp6)
5. MWI smoothing (MWI)
6. R-peak detection (r_peak_detector_full)
7. RR calculation (rr_hr_block)
8. R-peak refinement (r_peak_align_stream)
9. Morphological features (morph_features)
10. Morphological descriptor (morph_descriptor)

ECG (12b) --> filter (FIR 181) --> differentiator (FIR 5) --> square
          --> moving avg (lowp6) --> MWI (FIR 51) --> r_peak_detector_full
                                                       |-> rr_hr_block
                                                       |-> r_peak_align_stream --> morph_features --> morph_descriptor

• rtl/filter.sv : 181-tap FIR bandpass/conditioning filter (MATLAB HDL Coder output).
• rtl/differentiator.sv : 5-tap FIR differentiator.
• rtl/lowp6.sv : moving average / smoothing over 23 samples (sum of input + 22 delays).
• rtl/MWI.sv : 51-tap FIR used as MWI smoothing.
• rtl/r_peak_detector_full.sv : adaptive thresholding, refractory, T-wave rejection, and search-back.
• rtl/rr_hr_block.sv : RR interval calculation (assumes FS_HZ=360 by default).
• rtl/r_peak_align_stream.sv : aligns detected R-peak with raw ECG using a delayed stream and a local search.
• rtl/morph_features.sv : extracts P/Q/S/T times and amplitudes around each R-peak.
• rtl/morph_descriptor.sv : normalizes R-peak location and amplitude within P/Q/R/S/T extrema.
• rtl/int_div.sv : iterative integer divider used by morph_descriptor.

• Sample rate assumptions in control logic:
  • r_peak_detector_full uses constants for ~200 ms and ~2 s windows at 360 Hz.
  • rr_hr_block defaults to FS_HZ=360.
• morph_features windows:
  • PRE_R_SAMPLES=43 (~120 ms at 360 Hz)
  • POST_R_SAMPLES=72 (~200 ms at 360 Hz)
  • SB_PRE_SAMPLES=PRE_R_SAMPLES+245 (extended pre-window for search-back)
• r_peak_align_stream:
  • TOTAL_LATENCY controls alignment between processed stream and raw ECG.
  • Top-level sets TOTAL_LATENCY=220 in rtl/pantompkins.sv (sum of filter delays: 181+5+22+51).
  • Top-level sets SEARCH_BACK=9 and SEARCH_FWD=9 (±9 sample refinement window).

• The pipeline is streaming and gated by clk_enable.
• The filters have inherent group delay; the alignment stage uses TOTAL_LATENCY to compensate before morphological extraction.
• R-peak detection operates on the MWI output; refined R-peaks are aligned back to raw ECG by r_peak_align_stream.
• If you change filter lengths or sample rate, re-evaluate:
  • TOTAL_LATENCY in r_peak_align_stream
  • r_peak_detector_full timing constants (refractory, init window)
  • rr_hr_block.FS_HZ

The top-level Makefile delegates to simulation/Makefile and also generates input data before simulation.

Questa:

make questa

Verilator:

make verilator

Examples:

make verilator TRACE=0
make verilator TRACE=1 RUN_FILE=data/100.txt
make verilator SAMPLES=5000

Notes:

• simulation/src.args enumerates RTL sources.
• The Verilator flow builds Vtb and runs it. If TRACE!=0 and RUN_FILE is set, GTKWave opens the matching VCD with simulation/sig.gtkw.
• simulation/verilator/main.cpp reads all files in simulation/data/ unless RUN_FILE is set.
• The Verilator run writes VCDs to simulation/trace/ and R-peak indices to simulation/rpeaks/.
• TRACE=0 disables VCD dumping; RUN_FILE=<path> runs a single input file.
• SAMPLES=<n> controls how many samples utils/cvt.py extracts into simulation/data/.

1. Ensure the required simulator is installed (Questa or Verilator + GTKWave).
2. Run one of the make targets above.
3. Inspect waveforms and R-peak outputs in the generated traces.

Input files are generated by utils/cvt.py from MIT-BIH data under data/MIT-BIH. The simulation make targets run this automatically, but you can run it directly:

python3 utils/cvt.py --samples 20000

Install dependencies for utils/cvt.py and utils/verif/check_ver.py:

pip install -r requirements.txt

utils/verif/check_ver.py compares the simulation r-peak CSVs (simulation/rpeaks/*.csv) against MIT-BIH annotations and writes summary CSV/XLSX files.

Top-level make:

make verify

Simulation make:

make -C simulation verify

Overrides:

make verify VERIFY_ECG_DB=data/MIT-BIH VERIFY_RPEAKS=simulation/rpeaks VERIFY_RESULTS=simulation/results

Each input file in simulation/data/ should be plain text with one signed integer sample per line. Values are read as 12-bit signed ECG samples (the testbench drives filter_in[11:0]).

Example (simulation/data/example.txt):

12
15
18
20
17
12
8
5
3
1
-2
-6
-9
-7
-3
0
4
10
15

• feat_valid pulses once per beat when RR and morph descriptors are aligned.
• When feat_valid is high, the top-level outputs (morph_*, morph_r_time, and RR intervals) are valid for that beat.
• Internally, you can probe:
  • r_peak_align_stream.r_peak_refined_valid / r_peak_refined for refined peak timing.
  • r_peak_detector_full.candidate_valid and candidate_time for raw detections.
  • rr_hr_block.rr_feat_valid with pre_rr_interval / post_rr_interval / local_rr_interval.
  • morph_features.valid with p/q/r/s/t time and amplitude outputs.
  • morph_descriptor.morph_done for descriptor-valid strobe.

Evaluated against MIT-BIH annotations with ±50 ms tolerance.

Notable difficult records (>5% failed detection):

Fixed-point hardware SVM (L1-RBF kernel) trained on the DS1/DS2 AAMI split from MIT-BIH. Features are extracted directly by the RTL pipeline and match the hardware exactly. RR features are normalised by global_rr (rolling 5-min mean), matching the hardware int_div.

Best configuration — morph_1, morph_2, pre_RR, post_RR (4 features):

AAMI targets (Sensitivity ≥ 75%, Positive Predictivity ≥ 75%) — both met on DS2.

SVM: L1-RBF kernel, C=0.5, gamma=0.30, weight_V=6 (threshold calibrated on DS1). Hardware: 822 support vectors — 3,288 MACs per beat.

To reproduce:

python -m utils.classification.train \
  --hw-C 0.5 --hw-gamma 0.30 --hw-weight-v 6 \
  --hw-features morph_1,morph_2,pre_RR,post_RR \
  --hw-train-n 99999 \
  --q-weight-scale 4096 \
  --hw-export

Add your preferred license information here.