Abstract
Cardiovascular disease disproportionately affects low- and middle-income countries (LMICs), where specialist cardiac care is scarce and AI-assisted digital stethoscopes are often cost-prohibitive or cloud-dependent. We study murmur screening on a ~150USD offline edge platform, asking: when compute is fixed, what model should be deployed? Using the CirCor DigiScope phonocardiogram dataset, we benchmark six CNN and transformer architectures (1.5M-27.6M parameters) using a unified 30s mel-spectrogram pipeline and patient-level five-fold cross-validation. Diagnostic performance is largely independent of architecture: all models achieve Challenge Scores between 0.875 and 0.896 despite an 18x difference in parameter count. In contrast, deployment efficiency differs substantially. On a $36 ARM board, EfficientNet-B0 with static INT8 runs in 210 ms using 4.3 MB, whereas the audio transformer is 7x slower and 25x larger without improving diagnostic performance. Post-training INT8 quantization reduces patient-level accuracy by 12-65%, and quantization-aware training does not recover this loss. On hardware without native INT8 acceleration, selecting an inherently compact FP32 model is therefore more effective than compressing a larger one. We release the bill of materials together with hardware configuration, data splits, and inference benchmarks to support reproducible deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_023.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MIRASOL_023_supp.pdf
Link to Open Review
Open Review Page: https://openreview.net/forum?id=M745SDYJPk
BibTex
@InProceedings{LudDom_ALowCost_MICCAISAT2026,
author = { Ludera, Dominik AND Kersten-Oertel, Marta},
title = { { A Low-Cost AI Stethoscope for Resource-Constrained Heart Murmur Screening } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17265},
month = {pending},
page = {pending}
}