Misdetections of `ecg_delineate()` with CWT on signals with T-wave inversions

[laiamp]

Hi!

I wanted to perform an ML pipeline for cardiac diseases using hand-crafted features computed from the output of ecg_delineate(), so I need a delineation that works with ECGs from patients with a disease.

• Do all the algorithms operate under the assumption that R-waves and T-waves are concave?

From prior discussions, I learned that if the signal is inverted, ecg_invert() should be applied. However, there are some scenarios where despite having concave P-waves and R-waves, the T-wave is convex (i.e. inverted).

Example

Here you have an example from Lobachevsky University Electrocardiography Database. In the lead II of the record 11, the ground truth T-wave is inverted. However, when applying ecg_delineate() using CWT it detects a maxima as the ECG_T_Peak (red line):

[welcome]

Hi 👋 Thanks for reaching out and opening your first issue here! We'll try to come back to you as soon as possible. ❤️

[atpage]

I am working on this problem now. My first impression is that it would be nice to detect P, QRS, and T wave direction/inversion separately, instead of considering the entire signal inverted or not. I'll post here if I come up with anything.

[stale]

Is this still relevant? If so, what is blocking it? Is there anything you can do to help move it forward?

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.

[atpage]

My current method to get annotations when a lead isn't fully inverted looks like:

original_anns = ecg_process(signal)
inverted_anns = ecg_process(inverted version of signal)
clean_signal = signal with baseline wander removed
clean_inverted_signal = inverted signal with baseline wander removed
for wave_type in ['P', 'T']:
    original_aucs = areas under all original_anns[wave_type] on clean_signal
    inverted_aucs = areas under all inverted_anns[wave_type] on clean_inverted_signal
    if median(inverted_aucs) > median(original_aucs):
        use inverted_anns for wave_type
    else:
        use original_anns for wave_type

This seems to work pretty well in most cases, but it's still a work in progress.

[DerAndereJohannes]

Hey, sorry for never getting back to this. Can you test out the changes that I made here: #1186 ? I would greatly appreciate your feedback if you deal with this type of data frequently.

[stale]

Is this still relevant? If so, what is blocking it? Is there anything you can do to help move it forward?

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.

[extreme-coder]

I dug into this and CWT delineator only accepts neg->pos wavelet crossings (upright waves) and then always does argmax, so an inverted T is doubly unreachable, because its crossing is pos->neg and argmax can't land on a trough. Explains near-baseline "T" on record 11.

I tried updating _find_tppeaks to accept both crossing directions (using argmax for upright waves and argmin for inverted waves) and selecting the T wave based on the largest deflection within an early post-R window. When I tested (LUDB lead II across 60 records) the T-wave hit rate at 150 ms improved from 79% to 88%. P-wave performance remained unchanged. Timing error record 11 decreased from 182 ms to 2 ms. Branch here

@atpage , I like your AUC + whole-signal method. I went per beat + amplitude to keep the change small. You think its worth trying this branch on your data?

Ready to open a PR if everyone's cool. First time contributing here so feedback is appreciated :)