ECG/EKG Defibrillator Shock Model for Embedded

An ECG/EKG (electrocardiogram) signal analyzer for detecting ventricular arrhythmias and providing defibrillator shock recommendations optimized for embedded systems achieving 96.0% F2-score.

• Optimized for embedded systems (C99, no external dependencies, <10 KB RAM, <10 KB Flash)
• Hardware floating-point support for ARM Cortex-M33 FPU
• Real-time processing at 100 Hz sample rate, 3-10s windows
• Training pipeline with random/local/adaptive grid search over 30+ parameters
• Pan-Tompkins QRS detection for identifying ventricular depolarizations
• Frequency domain analysis using Goertzel algorithm for VF detection
• Time domain analysis for rhythm characterization

THIS IS AN EDUCATIONAL/PROTOTYPE IMPLEMENTATION ONLY

• NOT a medical device
• NOT certified for patient care
• NOT compliant with IEC 60601 or other medical device standards
• DO NOT use this system for actual patient care

This software is intended solely for educational purposes, algorithm development, and prototype testing.

This analyzer processes EKG signals in real-time on an ARM Cortex M33 microcontroller to detect:

• Ventricular Fibrillation (VF) - chaotic, life-threatening arrhythmia
• Ventricular Tachycardia (VT) - rapid ventricular rhythm

The system recommends whether a defibrillator shock should be administered based on the detected rhythm.

• Signal Type: Single-lead EKG (Lead II equivalent)
• Data Type: 16-bit signed integers (int16_t)
• Sample Rate: 100 Hz
• Duration: 3-10 seconds (300-1000 samples)
• Pre-filtering: Baseline wander removed, noise filtered

The analyzer provides:

• Shock recommendation (yes/no)
• Rhythm classification (normal, VT, VF, other)
• Confidence level (0-100%)
• Heart rate (if detectable)
• Diagnostic information (QRS count, dominant frequency)

6286 labeled single-lead ECG windows as headerless CSV files at 100 Hz, 5s (500 measurements) or 8s (800 measurements):

• Shockable (3123): ventricular fibrillation (coarse, fine)
• Non-shockable (3163): normal sinus rhythm, ventricular (non-VF) arrhythmias, premature beats, supraventricular arrhythmias, and conduction abnormalities
• Labels: encoded in file names (shock-123059-vfib=coarse.csv and no-shock-084252-nsr=060.csv)

See datasets/README.md for the full description.

datasets/ ──► ekg_train ──► outputs/best_config.h ──► ekg_eval ──► src/ (embedded library)
 labeled CSV   grid search     tuned parameters        metrics      ekg_analyze()
               (F2 objective)                          per rhythm

Evaluated with ekg_eval on the 5s dataset (6286 windows):

Missing a shockable rhythm is far more dangerous than a false alarm, so training maximizes F2 (recall weighted over precision) and trades specificity for sensitivity. Results are measured on the same data used for tuning, without a held-out split.

• GCC compiler with C99 support
• Standard math library (-lm)

make          # Default: host build
make host     # Explicit host build
make test     # Build and run all tests
make ekg_csv  # Build CSV analyzer tool
make ekg_train # Build training/optimization tool
make ekg_eval # Build evaluation tool
make clean    # Clean build artifacts
make help     # Show all available targets

• ARM GCC cross-compiler (gcc-arm-none-eabi)
• ARM Cortex-M33 target support

make arm      # Cross-compile for ARM

make test     # Build and run all tests

.
├── src/                    # Core algorithm implementation
│   ├── ekg_analyzer.c/h    # Main analysis API
│   ├── ekg_config.h        # Configuration constants
│   ├── signal_processing.c/h  # Filters and preprocessing
│   ├── qrs_detector.c/h    # Pan-Tompkins QRS detection
│   ├── heart_rate.c/h      # Heart rate analysis
│   ├── vt_detector.c/h     # Ventricular Tachycardia detection
│   ├── vf_detector.c/h     # Ventricular Fibrillation detection
│   └── signal_utils.c/h    # Signal quality and Bayesian helpers
├── tests/                  # Test suite
│   ├── test_signal_processing.c  # Unit tests
│   ├── test_ekg_analyzer.c       # Integration tests
│   └── test_vf_detector.c        # VF detector tests
├── tools/                  # Development utilities
│   ├── ekg_csv.c           # CSV signal analyzer
│   ├── ekg_train.c         # Training/optimization tool
│   ├── ekg_eval.c          # Evaluation tool
│   ├── csv_utils.c/h       # CSV parsing utilities
│   └── eval_utils.c/h      # Evaluation helpers
├── docs/                   # Documentation
│   ├── architecture.md     # Signal processing pipeline
│   ├── development.md      # Development guidelines
│   ├── tools.md            # Tool usage details
│   └── testing.md          # Test suite details
├── Makefile                # Build configuration
└── README.md               # This file

#include "ekg_analyzer.h"

int main(void) {
    // Sample EKG data (from ADC)
    int16_t ekg_samples[500];

    // ... acquire samples ...

    // Analyze
    ekg_analysis_result_t result;
    int ret = ekg_analyze(ekg_samples, 500, &result, NULL);

    if (ret == EKG_SUCCESS) {
        if (result.shock_advised) {
            // Trigger defibrillator
            printf("SHOCK ADVISED - Rhythm: %s\\n",
                   result.rhythm_type == 2 ? "VT" : "VF");
        } else {
            printf("No shock advised\\n");
        }
    }

    return 0;
}

Analyze individual EKG signal files:

make ekg_csv
./bin/ekg_csv data/path/to/signal.csv

Optimize algorithm parameters using grid search:

make ekg_train
./bin/ekg_train dataset/

make ekg_eval_diag
./bin/ekg_eval_diag dataset/

See docs/tools.md for detailed usage information.

• docs/architecture.md - Signal processing pipeline details
• docs/development.md - Development guidelines and constraints
• docs/testing.md - Test suite documentation
• docs/tools.md - CLI tools for training and evaluation
• datasets/README.md - Dataset description

1. Pan, J., & Tompkins, W. J. (1985) - "A Real-Time QRS Detection Algorithm"
2. AHA/ACC/HRS (2017) - "Guideline for Management of Patients With Ventricular Arrhythmias"
3. Goertzel Algorithm - Efficient frequency domain analysis
4. Shannon Entropy - Signal complexity measurement
5. Spectral Flatness (Wiener Entropy) - Noise-like vs tonal signal characterization

Copyright © 2026 gw0 <gw.2026@ena.one>. All rights reserved.

If my work saves you time, money, or tokens, consider sponsoring the work that keeps it maintained — and tell whoever else benefits, human or agent.