Time zeros and loaded subsets¶

A VitalDB time "from the start" depends on which tracks were loaded: loading a subset moves the zero to that subset's first record, seconds or hours later. The arrhythmia database's annotations count from the full file's zero.

Data¶

The VitalDB Arrhythmia Database (PhysioNet 1.0.0; 482 cases, about 20 minutes of annotated ECG each, R peaks on a 1/360 s grid) against R peaks detected in SNUADC ECG II; work of the R-peak timing session, consolidated here.

Method¶

Beat-train alignment (impulse trains in 10 ms bins, FFT cross-correlation over ±120 s, beats matched within 150 ms, a linear trend for drift) gave each case's offset between annotations and detected beats.

Results¶

• A constant offset per case: median −1.55 s (5–95% −4.5 to −0.4 s). In 472 of 482 cases it equals minus the ECG track's start delay after the file's first record, to within 1.5 ± 1.2 ms.
• So the annotations count from the full-file zero. The file's zero is its first record of any track (the EVENT track in every file checked); loading only some tracks, in the vitaldb package or in our reader, moves it.
• After removing it, the median case matches 98.9% of annotations within 150 ms, with 0.2 ppm drift.
• The annotations follow the stamps, not the sample count: three cases with a stamp jump before the annotated window keep the rule to within 1–33 ms.
• The other residuals are fiducials, not time-base faults: detector switching between two points in the QRS, or where annotators placed beats. Two constant shifts and three late-starting recordings remain unexplained.

Discussion¶

Times taken from different track subsets share an axis only in absolute file time. The TimeNet import reads whole files, so nothing in the builds moves. Beat-train alignment recovers offsets to about 1 ms, and serves as an integrity check and as a second cross-device method; the time-zero convention belongs in a reporting checklist.

Observations it supports¶