Abstract
Self-supervised learning (SSL) is promoted as a way to reduce dependence on labeled data in medical machine learning, yet reported gains are rarely attributed to a specific pipeline component; in phonocardiogram (PCG) classification, preprocessing priors, learned representations, and task difficulty are seldom isolated. We audit when SSL adds measurable value with a controlled 2 × 2 factorial design crossing physiologically grounded Butterworth filterbank inputs against raw waveforms, and SSL pre-training against a randomly initialized frozen encoder, all sharing a matched Mamba backbone and evaluated with frozen linear probes. With the filterbank in place, SSL exceeds a seeded random frozen encoder on PhysioNet 2016 by only a small margin (AUROC ≈ 0.930 vs. 0.903; +0.027, disjoint bootstrap CIs), and low-label behaviour is similarly close. Removing the filterbank widens the SSL advantage on raw waveforms (+0.053 AUROC), and on a harder murmur grading task filterbank SSL models exceed the random baseline (n=34 patients, exploratory). We further test whether physiology-informed SSL confers an additional advantage using CycleSSL, an objective that pairs distinct cardiac cycles from the same patient and site; CycleSSL performs comparably to generic SimCLR rather than exceeding it. We propose random-encoder controls, preprocessing ablations, and difficulty-stratified evaluation as a general attribution framework for clinical SSL claims, with direct implications for deployment in low-resource screening settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_035.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=98pHKPNqHX
BibTex
@InProceedings{AmaMon_Preprocessing_MICCAISAT2026,
author = { Aman, Mona AND Uiso, Godbright Nixon AND Fonya, Brandone AND Thuo, John Bosco Gachomba AND Gatimu, Maurine Wanjiku AND Mukamakuza, Carine},
title = { { Preprocessing as Prior: Rethinking Self-Supervised Representation Learning in Cardiac Sound Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}