Scale-Separated Concept Bottlenecks for ECG Diagnosis

An electrocardiogram (ECG) diagnosis rests on two kinds of evidence. The first is the shape of individual heartbeats, which is local: a widened wave or a depressed segment is visible within a fraction of a second. The second is rhythm, which is global: the irregular spacing of atrial fibrillation only shows up across many beats. We propose a model that must first name its evidence as human-readable findings, which the literature calls concepts, and must then diagnose from those findings alone. The new constraint is that each finding is computed only from the time span where it can physically live. Beat-shape findings come from a head that sees a short window and reports which window fired. Rhythm findings come from a head that sees the whole record. The diagnosis layer reads nothing else. Training all findings together is the multi-task part, and the time-span restriction is the multi-scale part. The question is whether this structure makes the explanations more trustworthy at little or no cost in accuracy.

The opportunity exists because one public dataset supplies the needed labels. PTB-XL, described in PTB-XL, a large publicly available electrocardiography dataset, annotates 21,837 ten-second 12-lead records with 71 statements split into diagnostic, form (morphology) and rhythm categories. That split lets us supervise local and global findings separately and treat diagnosis as a prediction made from them. No paper in our search does this.

Core Claim

A concept bottleneck model (CBM) routes every prediction through a small layer of named concept scores, following Concept Bottleneck Models. We claim that restricting each concept's receptive field to the timescale of its clinical category has three effects, and that diagnostic accuracy stays within about 0.01 macro AUC of an unconstrained network while they hold.

The first effect is on leakage. Leakage means label information passing through soft concept scores rather than through their meaning, as documented by Hi-CBM. A short-window head can carry only what that window contains, so we expect less leakage than a flat bottleneck with a full view of the record.

The second effect is temporal localization. Each beat-shape concept's evidence reduces to a pooled maximum over windows, so the explanation says which beat drove the finding. This comes from the architecture itself, with no post-hoc attribution and without the many forward passes that Sum-of-Parts Models requires.

The third effect is on intervention. Correcting a few concepts by hand should fix more diagnoses than it does in a flat bottleneck, because each concept means only one thing.

The claim also makes the Resolution Hypothesis of HiMAE falsifiable. HiMAE observed that different health tasks are best read out at different depths of a hierarchical encoder. We impose that observation as a hard constraint and measure what it costs.

Relation to Prior Work

Three existing time-series CBMs motivate this work, and each has a gap we address.

Hi-CBM constrains the design. It shows that soft bottlenecks leak unless tightly restricted, so we audit leakage directly rather than assume a bottleneck is honest.

On the architecture side, several papers show local/global splits help accuracy, but none of their intermediate representations are interpretable:

UniTS and TimeDRL show that jointly training objectives at different granularities helps, but both are fully opaque. We borrow their premise that joint training of fine and coarse targets shares useful structure, and apply it to supervised concepts.

Research Program

We build a shared one-dimensional convolutional encoder, a hierarchical U-Net-style stack similar to HiMAE, on PTB-XL at 100 Hz. The PTB-XL benchmark paper reports that 500 Hz gave no significant diagnostic gain (Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL).

The local head reads a shallow encoder level whose receptive field we cap at roughly one to two beats. It produces per-window scores for each form statement and pools them over time with log-sum-exp. The global head reads the deepest level, which spans the full record, and produces rhythm-statement scores. Diagnostic statements are predicted by a sparse linear layer over the concatenated concept scores only. There is no residual path around the bottleneck.

We train with a weighted sum of concept losses and a diagnosis loss. We compare three training schemes: independent (concept heads first, then diagnosis), sequential, and joint. We also compare soft and hard (thresholded) concept scores, because Hi-CBM's leakage results depend on that choice.

We assign concept tiers in two ways. The first uses PTB-XL's clinical category (form is local, rhythm is global). The second assigns each concept to the encoder depth where a frozen-encoder probe peaks, as in HiMAE. Agreement between the two is itself a result.

A scale-swap control forces rhythm concepts through the local head and form concepts through the global head. If our tier assignment reflects real signal structure, rhythm-concept AUC should collapse under the swap and form-concept AUC should drop.

As a secondary, harder setting, we attach an optional mid-scale tier spanning a few beats. This tests whether a two-tier split is too coarse.

Baselines

The naive reference point is an unconstrained CNN trained directly on diagnostic statements with the same encoder and no bottleneck. It sets the accuracy ceiling for our backbone.

The strongest reported competitor is a CNN ensemble on PTB-XL diagnostic statements, which reaches macro AUC of approximately 0.939 in Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL. WaveSSM reports an overall AUROC of 0.942 on PTB-XL ECG classification, against 0.934 for S4, in WaveSSM. The file does not say which PTB-XL label set that figure uses, so we treat 0.939 as the like-for-like reference. We will retrain one strong CNN under our patient-stratified folds with the same tuning budget as our model, meaning the same number of hyperparameter trials and epochs. Beating it on accuracy is not the goal. Coming within 0.01 macro AUC while winning on leakage, localization and intervention counts as success.

The ablation that isolates our key ingredient is a flat CBM. It has the same encoder, the same 71-statement concept supervision, and the same sparse diagnosis layer, but every concept head reads the full-record representation. This is the soft time-series CBM design without scale restriction. Every interpretability metric is reported as the difference between our model and this flat CBM.

A probe baseline follows HiMAE. We pretrain the encoder by masked reconstruction, freeze it, fit per-concept linear probes at each concept's best depth, and then fit the same sparse diagnosis layer. This tests whether joint supervised training adds anything beyond resolution structure that self-supervision already provides.

Finally, the unconstrained CNN combined with post-hoc integrated gradients serves as the localization and faithfulness baseline.

Why not just…?

Why not just train the black box and explain it with saliency? Post-hoc attributions are not guaranteed to reflect the computation. Sum-of-Parts Models reports that on ImageNet-S its faithful group attributions reach error 0.267 against 0.306 for SHAP-F, and it attributes the gap to correlated inputs. Adjacent ECG samples are exactly that kind of input. Saliency also gives no way to intervene. We test this alternative directly: for each positive form concept, we delete the top-evidence window chosen by our head and, separately, the window chosen by integrated gradients on the black box. We then measure how much the concept or diagnosis score drops. If integrated gradients produce drops as large as ours, the localization claim fails.

Why not just put a standard flat concept bottleneck on a strong backbone? This is the most serious threat. It may match our accuracy, and PTB-XL's concept labels alone may make it look interpretable. The flat-CBM ablation settles the question on three measures. The first is a Hi-CBM-style stress test: we remove the top-k concepts and check how much diagnostic AUC survives, where a leak-free model should degrade. The second is a probe that predicts held-out diagnostic labels from concept scores beyond what the ground-truth concept values explain. The third is an intervention curve: diagnostic AUC as we replace k predicted concepts with ground truth. If the flat CBM matches on all three, scale separation adds nothing, and we will report that.

Why not just probe a pretrained hierarchical encoder at the right depth? HiMAE already shows that fine layers carry cardiac morphology and coarse layers carry slower structure. A shallow-layer probe with pooling is close to our local head. If the frozen-probe baseline matches our concept and diagnosis AUC, the contribution shrinks to an audit that the resolution structure comes for free. That would still be a useful finding, and we would reframe the paper around it.

Why not just train separate single-concept models? This removes negative transfer between tasks but loses shared features for rare statements. We compare per-concept AUC on the rarest decile of form and rhythm statements under joint versus separate training, with matched total compute.

Decisive Checks

All results use PTB-XL's patient-stratified folds, with five seeds and confidence intervals from a bootstrap over patients.

The proposal succeeds if four things hold together. Diagnostic macro AUC stays within 0.01 of the retrained unconstrained CNN. Accuracy degrades clearly faster than the flat CBM's under top-k concept removal, meaning less leakage. Deletion-based faithfulness beats integrated gradients. The scale-swap control collapses rhythm-concept AUC, confirming that the tier assignment reflects signal structure.

Two outcomes would falsify the core claim. One is that the scale-separated model costs more than 0.02 AUC. The other is that the flat CBM matches it on leakage and intervention.

A known limitation is that ten-second records make "global" mean record-level rhythm, not multi-hour dynamics. Extending the approach to long recordings is left for later work, because no concept-labeled long-record benchmark appears in our sources.