Source-linked AI summary
Pulseflow: PPG Counterfactual Generation Via Latent Transport
Hung Manh Pham, Dong Ma, Bin Zhu, Pan Zhou
TL;DR
AF labels can be scarce, while existing PPG generators do not explicitly transform a particular source recording toward a target rhythm. PulseFlow combines conditional representation learning with invertible latent transport to retain source information during rhythm editing. Across clinical cohorts, it shows rhythm transformation and source correspondence, with improved AF classification under limited labels.
Problem
AF labels are limited in many clinical settings, and existing generators do not explicitly model how a particular source recording should be transformed while retaining source-associated information.
Source: §1 · p. 1Method
PulseFlow combines conditional PPG representation learning with invertible latent transport, changing rhythm while retaining source-derived information for counterfactual generation.
Source: §1 · p. 1, §2.2 · p. 2Results
Across two clinical cohorts, PulseFlow demonstrated effective rhythm transformation and measurable source correspondence, with the highest mean AUROC across all three scarce-label settings.
Source: §4.2 · p. 4, §5 · p. 5Takeaways & Limitations
The generated signals provide practical value for AF classification when labeled data are scarce.
Source: §5 · p. 5The generated waveform is a model-implied counterfactual under the assumed structural model, not a uniquely identified biological outcome.
Source: §2.2 · p. 2
Abstract
from arXiv · showhide
Photoplethysmography (PPG) has become an important modality for continuous cardiovascular monitoring, including atrial fibrillation (AF) detection. However, labeled AF recordings remain limited in many clinical settings, making model adaptation difficult when only limited target data are available. Generative modeling offers a natural way to alleviate this scarcity by synthesizing additional AF signals. Existing approaches, however, mainly generate samples that match the target condition without explicitly modeling how an observed source recording should be transformed, making it difficult to leverage abundant source recordings from a specific population or cohort for targeted augmentation. We introduce PulseFlow, a source-conditioned counterfactual generation framework that combines conditional representation learning with invertible latent transport to edit cardiac rhythm while retaining information from the source. Experiments across two clinical cohorts demonstrate effective rhythm transformation, measurable source correspondence, and improved AF classification under limited labels.
1. INTRODUCTION
PPG can support continuous cardiovascular monitoring, but adapting models to cohorts with scarce AF labels remains difficult. PulseFlow addresses a gap in existing generation by transforming a particular source recording toward a target rhythm while retaining source information.
1. INTRODUCTION
- Local labels can be costly to acquire, and AF may be under-represented, motivating augmentation for adaptation to new clinical cohorts.PPG is used for cardiovascular tasks including AF detection.Source: §1 · p. 1
- Existing PPG generators match target-condition populations but do not explicitly model how a particular observed recording should be transformed.This distinction matters because PPG waveforms vary substantially between individuals and are associated with age and sex.Source: §1 · p. 1
- PulseFlow combines conditional PPG representation learning and invertible latent transport to change rhythm while retaining source-derived information, with weakly paired SR-AF recordings guiding correspondence.Its inference follows abduction, rhythm intervention, and counterfactual prediction.Source: §1 · p. 1
2. METHOD
PulseFlow learns a conditional PPG representation and an invertible latent transport to generate source-conditioned rhythm counterfactuals. Its method proceeds from representation learning through conditional flow modeling to inference, under a structural model that treats rhythm and demographics as given.
2.1. Overview
- Stage 1 learns a compact PPG representation with a conditional decoder, then Stage 2 freezes that representation and learns an invertible transport conditioned on rhythm and demographics.At inference, the source latent state is transported under the changed rhythm context and decoded into a counterfactual waveform.Source: §2.1 · p. 2
2.2. Structural Setting
- The structural model treats rhythm and demographics as given and models their relations to latent state and waveform, including direct context effects on waveform expression.Generated waveforms are model-implied counterfactuals under this assumed structure, not uniquely identified biological outcomes.Source: §2.2 · p. 2
2.3. Stage 1: Representation Learning
- Stage 1 uses an encoder and conditional decoder to learn a waveform representation, with reconstruction, Gaussian-posterior regularization, and mean-RR prediction objectives.The RR objective encourages beat-timing information in the latent representation; the encoder and decoder are frozen before Stage 2.Source: §2.3 · p. 2
2.4. Stage 2: Conditional Flow Modeling
- Stage 2 fits context-conditioned latent distributions with an invertible flow, but likelihood fitting alone cannot determine which AF state corresponds to a particular SR state.This motivates an additional correspondence objective.Source: §2.4 · pp. 2–3
- Same-patient recordings observed under both rhythms guide cross-rhythm correspondence as weak supervision, not as exact counterfactual targets.The recordings may differ in beat phase and clinical state because they were acquired at different times.Source: §2.4 · p. 3
2.5. Counterfactual Inference
- At inference, the encoder and flow infer source state, while only rhythm is changed and source demographics, shared base state, and reconstruction residual are retained for decoding.The residual restores waveform information not captured through the VAE bottleneck; VAE encoding remains approximate.Source: §2.5 · p. 3
3. EXPERIMENTAL SETUP
The experiments use patient-separated MIMIC-III-Ext-PPG data and compare PPG generators using rhythm, distributional, and patient-level correspondence measures. They also test scarce-label augmentation on unseen VitalDB.
3.1. Datasets and baselines
- The MIMIC-III-Ext-PPG setup uses 30 s, 125 Hz segments, patient-level 80/10/10 splits, and paired SR-AF data from 400 training patients for Lcf.Stage 1 uses approximately one million segments; Stage 2 is fitted to their latent states.Source: §3.1 · p. 3
3.2. Evaluation protocol
- Evaluation measures AF conversion, distributional similarity to real AF, beat-interval statistics, and patient-level Wasserstein-1 distance between generated and observed AF recordings.Distributional measures include FD in AnyPPG and CardioState spaces, SDNN, and RMSSD.Source: §3.2 · p. 3
3.3. Augmentation with scarce labels
- On unseen VitalDB, the scarce-label setup trains with 1, 2, or 4 AF patients and a fixed 60 SR patients, then tests on 8 AF and 60 SR patients.Generation sources are SR recordings from 60 further patients; each baseline adds the same number of synthetic AF windows.Source: §3.3 · pp. 3–4
4. RESULTS
PulseFlow transforms SR recordings toward AF while retaining measurable source correspondence, and achieves the highest mean AUROC in three scarce-label settings on VitalDB. Ablations indicate contributions from latent transport, residual source information, cross-rhythm supervision, and rhythm-related latent information.
4.1. Rhythm conversion and source correspondence
- PulseFlow achieves 93.1% AF conversion versus 92.6% for real AF, the lowest FD among generation methods in both representation spaces, and 17.2% top-1 patient retrieval versus 5% chance.DenoisePPG-DM has slightly lower patient-level W1 than PulseFlow: 112.3 versus 113.6 ms.Source: §4.1 · p. 4
4.2. Augmentation performance
- Table 2 reports the highest mean AUROC for PulseFlow across all three settings; with one AF patient, it reaches 83.6% versus 67.6% with real data alone and 78.0% for DenoisePPG-DM.PulseFlow's margin over DenoisePPG-DM narrows from 5.6 to 2.9 and 0.4 percentage points as AF labels increase.Source: Table 2 · p. 4, §4.2 · p. 4
4.3. Component ablations
- Ablations show that decoding the source latent under AF without Stage 2 yields 5.5% AF conversion; removing Lcf lowers conversion from 93.6% to 87.9% and raises W1 from 112.8 to 142.3 ms.Removing residual ε lowers augmentation AUROC from 83.6% to 75.0%, while removing Lvital increases FD from 45.5 to 67.4.Source: Table 3 · p. 4, §4.3 · p. 4
5. CONCLUSION
PulseFlow combines conditional representation learning with invertible latent transport for source-conditioned PPG counterfactual generation.
5. CONCLUSION
- PulseFlow uses conditional representation learning with invertible latent transport to generate source-conditioned PPG counterfactuals, with reported rhythm transformation, source correspondence, and AF classification value under scarce labels.Future work will extend the framework to broader physiological factors and more explicit modeling of inter-individual PPG variation.Source: §5 · p. 5