Separating waveform magnitude, shape and timing variability AI
Different voltages can reflect sensor gain or coupling as well as physiology. The explorer therefore distinguishes recorded magnitude from a unit-range shape. Subtracting a cycle’s minimum and dividing by its peak-to-trough span removes overall offset and scale. It does not remove timing changes, motion or measurement artifacts.
Alignment changes the comparison
Cycle phase divides elapsed time by that ECG cycle’s duration. ECG uses phase −0.2 to +0.8 around the fiducial; the other waveforms use the ECG-to-next-ECG interval. Own-landmark alignment instead centers the ECG fiducial, PPG peak or BCG J. Invalid landmarks leave the corresponding profile unavailable. Curves are not wrapped or extrapolated into adjacent cycles.
The shaded bands show pointwise 10th–90th percentiles across eligible observed beats, supported only where at least 80% of curves provide data. These are empirical variability bands, not confidence intervals. A broad cycle-aligned band may reflect timing jitter, shape changes or both. Realigning may help distinguish them, but it is not a formal physiological decomposition.
Scalar timing variability
SD, sample variance and IQR describe the distribution of beat-level delays. RMSSD describes successive differences between adjacent finite cycles; a missing beat is never bridged. These measures can reflect physiology, detector error or mechanical coupling.
The explorer’s statistics characterize a short excerpt, not a stable personal phenotype. Prediction uses beat-level descriptors with training-only templates and five-beat smoothing, rather than treating every descriptive summary as a model input.