PRECISE annotation protocol:
notice everything

New benchmark in ECG interpretation.

All diagnostic criteria that describe abnormal conditions seen on the ECG, are written in terms of intervals, amplitudes, waves (e.g. “prolonged QT”, “inverted T wave”). But any ECG recording is basically nothing more than a wavy line. Recognition of elements in an ECG recording is called ECG annotation.
What makes this task so complicated? The fact that the real ECG never looks that way. It has different kind of noise, baseline wandering and other artifacts. Also an ECG often combines different abnormalities, making it non-periodical, with uneven waves hard to recognize.

Why correctness of annotations means a lot? Although automated ECG interpretation standard ISO – 11073:91064 requires a medical professional to re-think and validate an interpretation provided by a software, an annotation mistake makes the whole interpretation useless: - undetected waves mean respective intervals calculated wrong
Undetected P waves VS correctly detected in PRECISE

- non-existent complexes extracted out of noise corrupt e.g. heart rate values, rhythm characteristics etc.

Exta QRS in baseline noise VS noise treated as noise in PRECISE

In fact, annotation is the most technically challenging part of ECG interpretation – it’s foundation stone. Thus, having a precise annotation algorithm is essential for correct ECG interpretation.
Most of the current algorithms use approach of a “moving window”: a particular wave is expected to be found within an interval, so any peak found in this interval recorded as this wave, repeated for the duration of an ECG recording. Such an approach often fails with skipped beats, irregular abnormal rhythm, fast tachycardia (when those “windows” for R and P overlap), inverted waves and other conditions.

In PRECISE we use a completely different approach. First, we carefully remove noise, moving from basic to a more specialized filtering, calculating noise levels and components for each particular ECG. Then we remove baseline wandering. Then – detect each feature on a signal: each upward and downward peak. So we leave no blind spots – waves that remain unrecognized in other approaches because “windows” didn’t “expect” them to appear where they actually are. The next level is annotating itself – naming detected waves/peaks.