ICASSP 2026 · Biomedical audio

Hear a heartbeat at three temporal scales.

TopSeg turns persistent topological structure into a data-efficient representation for segmenting S1, systole, S2, and diastole from phonocardiograms.

Peihong Zhang · Zhixin Li · Yuxuan Liu · Rui Sang · Yiqiang Cai · Yizhou Tan · Shengchen Li

71.9
macro-F1 · 10% labels · 60 ms tolerance
+6.2
F1 over log-mel TCN · 10% labels
2 datasets
training benchmark + external validation
Method in motion

One heart sound becomes three topological views.

Animated reading: nested temporal windows scan the PCG, then topology persists across scale before the decoder emits a physiologically ordered state sequence.

TopSeg multi-scale topology pipelineA heart-sound waveform is analyzed at global, mesoscopic, and fine temporal scales, converted to topological descriptors, decoded by a TCN, and refined into four physiological states.PCG · MULTI-SCALE OBSERVATIONGLOBAL · 2/4/8 sMESO · ≈500 msFINE · ≈100 msDELAY EMBEDDINGPERSISTENT HOMOLOGY · H₀ / H₁Persistence landscapesTCN + CONVEX REFINEMENTS1SYSS2DIANovelty: topology encodes rhythm at three physiological scales before sequence refinement.
Figure · TopSeg representation and decoding pathThe moving windows denote scale-specific analysis; the four-state band shows the constrained output order.
01 · Question

Can structural signal replace some annotation?

Heart-sound boundaries are expensive to annotate, while conventional time-frequency representations do not explicitly encode the repeating topology of a cardiac cycle.

Scarcity

Expert labels are limited.

The controlled protocol evaluates 5%, 10%, 25%, 50%, and 100% subject-level label budgets.

Noise

Envelope cues can break.

Simple amplitude envelopes can be sensitive to recording quality and do not expose multi-scale structure.

Physiology

Order is not arbitrary.

The four states follow a physiological sequence with characteristic durations that can constrain inference.

02 · Method

Represent shape, then enforce plausible state transitions.

Embed

Time-delay geometry

Each temporal scale is embedded into a point cloud so repeating acoustic dynamics become geometric structure.

Describe

Persistent homology

H0 and H1 persistence landscapes summarize connected components and loops across filtration scales.

Decode

TCN + refinement

A lightweight TCN predicts states; inference-only convex refinement encourages valid order, duration, and topology alignment.

03 · Evidence

Topology helps most when labels are scarce.

5% labels

66.7 F1

Full TopSeg at 60 ms tolerance.

TopSeg
66.7
Topo TCN
64.1
10% labels

71.9 F1

The log-mel TCN reports 64.2, while the topology-only TCN reports 70.4.

100% labels

85.3 F1

The advantage persists when the complete labeled training set is used.

Macro-F1 values at 60 ms tolerance, reported in Tables 3 and 4 of the paper.

04 · Boundary

Designed around data efficiency and physiological plausibility.

Evaluation frame

  • PhysioNet/CinC 2016: 3,153 recordings from 764 subjects.
  • External validation on CirCor: 5,272 recordings from 1,568 subjects.
  • Subject-level splits prevent recording leakage across budgets.

Claim boundary

  • Topological features add preprocessing and design choices.
  • The convex refinement is used only at inference.
  • Clinical deployment requires prospective validation beyond benchmark segmentation.
Method figure · original paper

Read the heart at the scale of its physiology.

The published framework is deliberately representation-first: three time horizons produce topology-aware descriptors, then a lightweight temporal decoder and inference-only convex refinement recover the ordered cardiac states.

Published TopSeg Figure 3 showing multi-scale topological encoding, temporal convolutional decoding, convex refinement, and final heart-sound segmentation.
Figure 3 · TopSeg framework, reproduced from the paper.The refinement layer is used at inference to preserve physiological sequence structure.
Three scales global, meso, and fineTransform embed → homology → landscapesDecode TCN + constrained refinement

Read the topology construction and full ablation.

Open paperPublication list