Can Deep Neural Networks Read Myocardial Fibrosis in the Terminal QRS?

The subendocardial-versus-intramural problem

Subendocardial fibrosis is the signature of ischemic injury — enhancement that begins at the endocardium and follows a coronary territory. In genuinely non-ischemic cardiomyopathy, the dominant pattern is mid-wall or intramural: the classic septal mid-wall stria in dilated cardiomyopathy, and patchy mid-myocardial enhancement in hypertrophic disease.

So a non-ischemic cohort asked to separate intramural septal from subendocardial fibrosis is close to self-contradictory. Subendocardial enhancement in that population is uncommon and, where present, usually flags a mixed or ischemic etiology rather than a second flavor of non-ischemic disease. A classifier trained on such a split risks learning "non-ischemic versus occult ischemic overlap" as much as scar depth itself — a confound no architecture tuning can remove.

Why the last 20 ms is narrower than the substrate

The physiological anchor is real but conventionally defined differently. Late potentials — the low-amplitude, high-frequency terminal signals tied to slow conduction through fibrosis — are classically measured in the terminal 40 ms on a signal-averaged ECG, because they reflect slow, inhomogeneous impulse propagation through fibrotic myocardium that forms a substrate for re-entry.

A 20-millisecond window is narrower than that established substrate signal, and at surface amplitudes it sits largely in the noise floor without signal averaging — precisely where raw, single-beat neural-network input is least reliable.

What invasive mapping actually shows

The strongest real link between terminal QRS behavior and scar geometry comes from the ventricular-tachycardia mapping literature, not from surface-ECG AI. In a 2020 study in the Journal of Cardiovascular Electrophysiology, patients with non-ischemic cardiomyopathy underwent cardiac MRI before ablation; late-enhancement transmurality was quantified and coregistered point-by-point with thousands of endocardial and epicardial electrogram sites.

Voltage tracked transmural signal intensity, and longer electrogram duration and fractionation went with increasing transmurality — but late potentials and local abnormal ventricular activity were actually more prevalent in non-transmural than transmural regions. Critically, exit-site QRS duration was affected by transmurality and by intramural scar location, but not by wall thickness. That is genuine evidence that QRS timing encodes scar depth — yet it is derived from paced beats at anatomically known origins in a small cohort, not from a blind surface classifier.

Which features drive an explainable model

The closest work on feature attribution is explainable scar localization. A 2025 study in Scientific Reports built an interpretable model across hundreds of patients, using late-enhancement MRI within one year of the ECG as ground truth for regional (basal, mid, apical) scar.

Three findings matter here. First, the ground truth was regional presence or absence — not transmurality, and nowhere near an intramural-versus-subendocardial contrast. Second, and most important, it is not a raw-waveform deep network: the authors report that no method trained on raw signals or computed features could separate scar from no-scar effectively, which is why they turned to unsupervised patient clustering plus self-supervised learning over handcrafted, interpretable features. The strongest ECG-detects-fibrosis result in this space is therefore an argument against raw-signal deep learning being sufficient.

Third, on attribution: Shapley values scored each feature per prediction. The features that mapped onto the slope / curvature / extrema question included R-upstroke slope and segment-slope variation, plus QRS fragmentation — fragmentation being the closest analog to a "local extrema count." Curvature as such was not in the feature set. And the depolarization feature that mattered for basal scar was prolonged non-terminal QRS duration (the intrinsicoid deflection), not the terminal deflection — the opposite of where a 20-millisecond terminal window sits.

A more defensible formulation

Two calibration points carry into any design. Detection was near-perfect above roughly 10% scar burden but missed progressively more as burden fell — and even coarse localization degraded fast, with septal detection modest and anterior-wall detection poor. Depth discrimination is harder than either, and no study has attempted it from the surface ECG.

A more defensible reformulation targets transmural-versus-intramural depth in a confirmed non-ischemic cohort, using the terminal 40 ms plus fragmentation features, with Shapley or saliency attribution across leads and time — and likely signal-averaged or high-resolution input rather than raw single beats, so the terminal signal clears the noise floor. Physiology grants the license; methods urge the caution.