Limitations and intended use

This software is research and engineering software. It is not a medical device and is not intended to diagnose, treat, cure, or prevent disease. Outputs require independent validation for the intended use and must not be used as a substitute for professional medical judgment.

Records carry intended_use = "research_and_engineering_only"; no record is medically cleared, approved or clinically validated.

  • In schema 1.1.0 only the record-level P duration, PR, QRS and QT are benchmark_validated, for resting 10 s, 500 Hz, 12-lead ECGs like LUDB; the heart rate and the P, QRS and T frontal axes are indirectly_validated (agreement with GE 12SL, another algorithm); every other field is unvalidated. Of the per-lead fiducials only t_offset meets its CSE delineation tolerance on LUDB (P and QRS boundaries do not, by a factor of about two), and median-beat amplitudes agree with 12SL less closely than a second commercial algorithm does (Validation).
  • The engine is designed for resting recordings of about 10 s. Longer recordings are measured window by window (ecg_record_segments), without beat continuity across windows.
  • In atrial fibrillation the record can still report P-wave quantities (PR, P axis) found by the atrial detector; check organized_p_fraction, the feature states and the rhythm plot.
  • The bundled example ECGs are PTB-XL records (CC BY 4.0); cite PTB-XL when you use them.
  • P-wave delineation is the weakest endpoint in the repository benchmarks (per-lead P onset/offset SD about 20 ms on LUDB against a tolerance of 10–13 ms), especially at short RR intervals or when the P wave overlaps the previous T wave.
  • LUDB distributes amplitude-normalized signals, so LUDB cannot validate amplitudes.
  • The four-class ST morphology was never validated and is anti-correlated with ischemia; it is not published and must not be used as ischemia evidence.
  • The package measures; it does not diagnose. Every record states capabilities.diagnosis = not_applicable (measurement_package_boundary).
  • Measurements come from automated delineation. They can be wrong on noisy signals, paced rhythms, arrhythmias with irregular or absent P waves, and morphologies unlike the benchmark datasets. Check quality, the absence states and the plots before relying on a value.
  • The pipeline is inspired by publicly documented practice; it is not a reproduction of any proprietary product.