The Front End Holds the Line: What Heart-Sound CNNs Lose When Models Shrink

TL;DR for operators

A team shrinking an audio classifier for a low-cost screening device has more than one place to spend scarce compute. The network can stay larger, or the input representation can do more work before the signal reaches the network.

That trade-off became visible when this heart-sound CNN was reduced from three convolutional blocks to two. Using a plain log-mel spectrogram, modified accuracy—a score that gives equal weight to abnormal-case sensitivity and normal-case specificity—fell from 0.910 to 0.826. With a front-end that normalizes each frequency band against its recent energy so locally unusual sounds stand out more clearly, called PCEN, it fell only from 0.915 to 0.894. A front-end representing the same sound at several time-frequency resolutions, called multi-resolution log-mel, similarly fell from 0.916 to 0.894.

For edge or low-cost screening systems, preprocessing is therefore not merely a secondary implementation choice once model capacity becomes constrained. The engineering decision is whether a richer front-end can preserve enough detection quality—and limit false-positive burden—to justify shrinking the downstream network. The experiment supports that trade-off as a design hypothesis, not a universal front-end ranking or evidence of clinical readiness.

Model compression can expose weaknesses upstream

Suppose a team has an audio classifier that performs adequately in development but needs to run on cheaper hardware. Reducing network depth is an obvious place to recover compute. The assumption behind that choice is often that preprocessing will remain a secondary concern: keep the same spectrogram, shrink the neural network, and accept some degradation.

The experiment by Abhinav Pala and Dhanush Pala challenges that assumption in a specific heart-sound setting.¹ When their CNN was reduced from three convolutional blocks to two, modified accuracy fell from 0.910 to 0.826 with the vanilla log-mel representation. With two enhanced front-ends, it fell only from 0.915 to 0.894 for PCEN and from 0.916 to 0.894 for multi-resolution log-mel.

That difference is larger than the headline advantage seen with the original network. The representation choice becomes more consequential as downstream capacity is removed.

Modified accuracy matters here because the dataset contains substantially more normal than abnormal recordings. It is defined as

so abnormal-case recall and normal-case recall receive equal weight rather than letting the majority class dominate the score.

With the full CNN, the improvement is real but small

The main experiment holds the three-block CNN, optimizer, learning rate, training duration, class weighting, data split, and seed constant. Only the spectrogram front-end changes. This is the paper’s strongest evidence for attributing the observed differences to representation choice within this setup.

The three representations are:

• Vanilla log-mel: a conventional 64-bin spectrogram representation.
• PCEN: per-channel energy normalization, which adjusts each frequency band relative to its recent energy level so locally unusual signals become more prominent.
• Multi-resolution log-mel: three versions of the signal computed with different FFT sizes, exposing several time-frequency trade-offs to the CNN simultaneously.

On the main architecture, the numerical differences are modest.

Calling 0.916 versus 0.910 a decisive benchmark victory would overstate the evidence. Vanilla log-mel is already strong, particularly on sensitivity. PCEN and multi-resolution improve F1 and specificity while changing abnormal-case recall very little.

The more informative question is therefore not which front-end wins by six thousandths on one aggregate score. It is what kinds of errors change, and whether those differences persist when the network has fewer resources.

The gain is mostly fewer false alarms

The confusion matrices make the main-model result easier to interpret.

Vanilla log-mel produced 197 false positives on normal test clips. PCEN reduced that to 179, while multi-resolution produced 182. False negatives remained tightly clustered: 24 for vanilla, 25 for PCEN, and 23 for multi-resolution.

So the enhanced representations do not produce their advantage through a large increase in abnormal-case detection. Sensitivity remains around 0.95 for all three. Their main contribution is better recognition of normal clips.

That matters operationally in a screening workflow. If an abnormal prediction leads to additional examination or clinician review, false positives consume follow-up capacity. The paper does not quantify that downstream cost, so a financial ROI claim would exceed the evidence. But it does identify the error category a product team would need to price: fewer unnecessary alerts rather than dramatically fewer missed abnormalities.

The smaller CNN is a robustness test, not a second benchmark win

The two-block experiment serves a different purpose from the main comparison. It asks whether the apparent advantage of the enhanced representations survives a change in classifier capacity.

It does, and the difference widens.

On the smaller model, PCEN records higher accuracy, specificity, abnormal F1, and macro-F1 than multi-resolution. Multi-resolution instead has the best sensitivity and only 14 false negatives, compared with 26 for PCEN and 39 for vanilla.

This is stronger evidence for a representation-versus-capacity trade-off than for a universal front-end ranking. PCEN and multi-resolution appear to provide structure that the smaller network would otherwise have to learn for itself.

The mechanism is plausible but should remain framed as interpretation. PCEN adaptively normalizes energy within frequency channels. Multi-resolution processing gives the model several views of the same signal at different time-frequency resolutions. In a limited-data setting, either can supply useful inductive bias before the CNN begins learning.

For an edge implementation, that creates an architectural choice: allocate scarce computation to a larger learned model, to a more structured front-end, or to some combination of the two.

Grad-CAM supports a signal-level interpretation, not a diagnosis

The paper also applies Grad-CAM to inspect which time-frequency regions influence abnormal predictions. This analysis is diagnostic evidence about model attention, not an independent validation of classification performance.

Across the models, attention is concentrated mainly in the lower-frequency heart-sound region associated with S1/S2 and murmurs. Vanilla log-mel shows relatively more high-frequency attention. PCEN concentrates more attention around and immediately after heartbeat spikes, while multi-resolution attention is spread more broadly across the waveform.

These patterns are consistent with the idea that changing the front-end changes what information the CNN finds salient. They also reduce one concern: the models do not appear to concentrate primarily on obvious high-frequency artifacts.

They do not demonstrate that the CNN has learned a medically validated disease mechanism. The authors explicitly identify medical-professional verification as necessary before interpreting these attention maps clinically.

There is no established universal winner

The small differences between PCEN and multi-resolution need to be read against the paper’s variance evidence. In reruns of the smaller-architecture experiment, modified accuracy moved by roughly one to two percentage points between seeds. Several front-end margins are of comparable size.

That makes statements such as “multi-resolution is the best representation” too strong. Its 0.916 modified accuracy is the highest main-model result, but only narrowly. On the smaller network, it ties PCEN on modified accuracy while losing on several other metrics and winning on sensitivity.

Complexity also differs. Multi-resolution uses three input channels rather than one, increasing front-end and input-processing cost for a small measured advantage. The study does not provide a deployment-level latency or energy analysis, so it cannot determine whether that additional complexity pays for itself.

PCEN therefore deserves attention as an engineering option rather than as a benchmark champion: in this experiment it preserves the smaller CNN’s performance almost as well as multi-resolution while avoiding the latter’s three-resolution input representation.

The deployment implication is architectural, not clinical

For a team building a resource-constrained digital-stethoscope or screening system, the paper supports a concrete design hypothesis: preprocessing can be tested as a substitute for some downstream model capacity.

The evidence is not yet sufficient to convert that hypothesis into a clinical deployment rule. The study uses one benchmark dataset, one primary CNN, one additional smaller architecture, and limited repeated-seed evidence. Cross-validation or systematic multi-seed averaging would strengthen confidence in the size and ordering of the effects. The recordings are also largely described as resting-state heartbeats, and condition-specific diagnostic evidence is limited.

The appropriate next experiment for a product team would therefore compare end-to-end system configurations rather than front-ends in isolation: model size, preprocessing cost, latency, energy use, sensitivity, specificity, and false-alarm workload under the intended deployment conditions.

The paper’s strongest result is narrower and more useful than declaring a new best spectrogram. Representation choices that look minor when compute is abundant can determine how gracefully a classifier degrades when capacity is constrained.

Cognaptus: Automate the Present, Incubate the Future.