Organizations: School of Computer Science and Engineering, Changchun University of Technology, Jilin, Changchun, 130102, China · Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Jilin, Changchun, 130012, China · Smart Health Joint Innovation Laboratory for the New Generation of AI, Jilin, Changchun, 130102, China
Aligning physiological parameter labels with large-scale photoplethysmographic (PPG) data for deep learning is challenging and resource-intensive. While self-supervised representation learning (SSRL) can handle limited annotated data, the challenge lies in learning robust shared representations from vast unlabeled data and integrating contextual cues to learn distinctive representations. To alleviate these challenges, a generative SSRL framework TS2TC is proposed to utilize the temporal, spectrogram, and temporal-spectrogram mixed domains to explore and incorporate the unique features of PPG for universal and noninvasive physiological parameter estimation. A pretext task named Cross-Temporal Fusion Generative Anchor (CTFGA) is designed, modeling temporal dependencies and reconstructing independent segments at a coarse level to provide robust global feature extraction and local contextual representation. The framework includes sub-signals from PPG with diverse frequency scales and order derivatives reflecting hemodynamics to facilitate learning shared representations at varying semantic levels. Secondly, a cognitive-inspired dual-process transfer (DPT) strategy is formulated, consisting of prior-dependent autonomous processes and posterior observation reasoning processes, to leverage the independent and integrated advantages of shared and specific representations. TS2TC introduces a bilinear temporal-spectrogram fusion method in the mixed domain, aligning latent representations from different domains and establishing fine-grained contextual interactions across multiple sources of information. Extensive experiments on physiological parameter estimation tasks showed that the joint performance of CTFGA and DPT outperforms standard generative learning significantly. TS2TC achieved an average 2.49% improvement in RMSE over state-of-the-art estimation methods with only 10% training data.
Figures & tables
Figure 1: Overview of photoplethysmography technology.
Figure 2: Overview of the TS2TC framework. In the figure, x(t) represents the input PPG signal, xs denotes the corresponding spectrogram, and y signifies the labeled physiological parameters. x^B′(t) and y^1 represent the signals and physiological parameters estimated in the temporal domain. Similarly, x^s and y^2 are the spectrogram and physiological parameters estimated in the spectrogram domain, while y^3 represents the physiological parameters estimated in the mixed domain.
Figure 3: Contrasting supervised, unsupervised, and self-supervised learning paradigms. Unsupervised learning involves learning from unlabeled data by constructing generative models or density estimators. Here, we illustrate this with the classical method known as the autoencoder.
Figure 4: Workflow of self-supervised representation learning. The term ’(Update)’ signifies that parameter updates are optional. When not updated, the method follows a linear probing paradigm; when updated, it adheres to a fine-tuning paradigm. (SBP is systolic blood pressure, DBP is diastolic blood pressure, HR is heart rate, SpO2 is blood oxygen saturation)
Figure 5: Generative-based SSRL for time series data. (Repainted from ( Zhang et al., 2024 ) )
Table 1: Statistical distribution of the PPG databases.
Figure 7: Transfer results under different pre-training methods and data amounts. (Standardized)
Figure 8: Pre-training results under different transfer methods and data amounts ( P : proportion of training data, SBP: systolic blood pressure, DBP: diastolic blood pressure).
Component
Temporal
Spectrogram
Mixed
Database
MAE
MAPE
MAE
MAPE
MAE
MAPE
SpO2 (%)
BIDMC
1.007±0.0485
0.010
2.297±0.218
0.024
0.996±0.085
0.010
UniGreifswald
0.947±0.0215
0.010
0.463±0.041
0.010
0.514±0.042
0.011
UniSydney
0.467±0.0175
0.005
0.273±0.011
0.004
0.427±0.024
0.006
HR (bpm)
BIDMC
1.879±0.074
0.022
10.41±1.105
0.104
8.001±0.687
0.080
IEEEPPG
15.293±0.2405
0.149
22.525±0.669
0.239
16.439±1.482
0.174
Table 2: Performance of components in each composition of TS2TC. (bold represents the best results in different domains)
SpO2 (%)
HR (bpm)
RR (bpm)
DBP (mmHg)
SBP (mmHg)
Average
Models
Metric
BIDMC
UniGreifswald
UniSydney
BIDMC
IEEEPPG
UniSydney
BIDMC
PWDB
CapnoBase
UniSydney
PPGBP
NBPE
UniSydney
PPGBP
NBPE
TFTC
MAE
1.217±0.103
0.608±0.051
0.428±0.035
4.094±0.172
16.305±0.613
3.950±0.236
1.445±0.095
4.832±0.171
4.707±0.141
8.571±0.367
9.576±0.303
3.426±0.149
5.504±0.550
12.836±1.347
9.352±0.827
5.790
MAPE
0.022
0.006
0.006
0.041
0.173
0.053
0.136
0.338
0.292
0.035
0.088
0.033
0.027
0.178
0.072
0.100
PatchTST
MAE
2.537±0.101
4.749±0.390
11.261±0.472
3.097±0.300
27.797±2.348
8.180±0.550
2.029±0.057
14.016±0.598
6.033±0.365
57.325±5.869
19.729±0.956
6.230±0.551
9.300±0.871
17.437±1.014
14.582±0.884
13.620
MAPE
0.027
0.049
0.115
0.035
0.262
0.108
0.159
0.909
0.364
0.757
0.281
0.089
0.080
0.143
0.106
0.232
TSSequencer
MAE
2.456±0.229
0.907±0.046
0.532±0.023
1.834±0.147
24.746±1.230
7.839±0.398
1.817±0.187
14.018±0.898
6.107±0.444
56.396±5.521
9.473±0.980
5.985±0.534
9.222±0.941
17.526±1.138
12.829±0.625
11.446
Table 3: Comparison of results of TS2TC framework under sample scarcity situation (MAE and MAPE ± Std), where bold and underline represent the best performance among the comparison methods, and underline only represents the second best performance.
Figure 9: The results of the TS2TC method on IEEEPPG (P=10%).
Models
CTFG-DPT
PatchTST
TSSequencer
TSPerceiver
XCM
MultiRocket
TSiTransformer
TST
OmniScale
InceptionTime
Metric
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
RMSE
MAPE
SpO2 (%)
BIDMC
10%
1.444±0.156
0.010
3.384±0.142
0.027
3.276±0.156
0.026
3.245±0.251
0.026
2.814±0.243
0.022
2.445±0.149
0.019
3.277±0.115
0.026
5.506±0.268
0.050
2.615±0.277
0.019
2.963±0.188
0.022
50%
0.800±0.065
0.005
3.253±0.263
0.025
3.277±0.126
0.026
3.233±0.155
0.025
2.053±0.122
0.015
1.236±0.078
0.009
3.276±0.328
0.026
3.839±0.164
0.027
1.225±0.119
0.009
1.506±0.115
0.011
100%
0.603±0.061
0.004
3.275±0.246
0.026
3.276±0.285
0.026
3.202±0.239
0.025
1.514±0.077
0.011
1.138±0.120
0.008
3.276±0.217
0.026
3.265±0.512
0.026
1.006±0.099
0.007
1.100±0.092
0.008
UniGreifswald
10%
1.422±0.102
0.010
19.515±2.024
0.049
1.328±0.098
0.009
1.590±0.092
0.011
2.058±0.062
0.014
5.373±0.307
0.038
1.306±0.092
0.009
10.367±0.794
0.101
1.647±0.165
0.014
8.572±0.449
0.024
50%
1.333±0.110
0.009
19.497±0.866
0.046
1.328±0.035
0.009
1.330±0.100
0.009
1.306±0.091
0.009
1.679±0.047
0.013
1.309±0.992
0.009
1.597±0.098
0.010
1.323±0.110
0.010
1.392±0.068
0.009
Table 4: Comparison of results based on CTFGA and DPT (RMSE±Std and MAPE), where bold indicates the best performance of the method, and underline indicates the second-best performance method.
Figure 10: Pretraining Validation Loss Results at Different γ and T .
Figure 11: Performance of different γ and T values in downstream tasks, with asterisks indicating the best performance.
Model
OCTFG
OCTFG+VMD
OCTFG+VMD+MSDI
OCTFG+VMD+MSDI+ZDL
Database
MAE
MAPE
MAE
MAPE
MAE
MAPE
MAE
MAPE
SpO2
BIDMC
1.698±0.252
0.018
1.695±0.111
0.018
1.614±0.047
0.016
1.007±0.049
0.010
UniGreifswald
1.170±0.030
0.012
1.144±0.099
0.012
1.085±0.069
0.012
0.947±0.022
0.010
UniSydney
8.201±1.602
0.084
5.929±0.407
0.061
4.931±0.348
0.056
0.467±0.018
0.005
HR
BIDMC
2.345±0.040
0.028
2.205±0.039
0.026
2.109±0.189
0.026
1.879±0.074
0.022
IEEEPPG
16.491±0.484
0.145
16.280±0.111
0.146
16.245±1.404
0.146
15.293±0.241
0.149
Table 5: Ablation of VMD, MSDI, and ZDL.
Database
R-squared
ϕ1
ϕ2
ϕ3
SpO2
BIDMC
0.916±0.034
0.755±0.014
-0.106±0.012
0.755±0.099
UniGreifswald
0.906±0.027
0.853±0.012
-0.190±0.016
0.853±0.063
UniSydney
0.918±0.038
0.848±0.026
0.170±0.028
0.848±0.081
HR
BIDMC
0.953±0.036
0.997±0.031
0.129±0.066
0.997±0.072
IEEEPPG
0.969±0.022
0.943±0.033
-0.115±0.062
0.943±0.126
UniSydney
0.934±0.029
0.932±0.014
-0.296±0.036
0.932±0.090
Table 6: Parameters of the TS2TC framework based on OLS fusion across different tasks.
Figure 12: Analysis of the contributions of components in the TS2TC framework on IEEEPPG.
Figure 13: Clarke error grid analysis of the proposed TS2TC.
Method
MAE (mmol/L)
MAPE (%)
Zone A (%)
Zone B (%)
PatchTST
2.522±0.122
35.93
42.55
53.19
TSSequencer
1.312±0.072
19.08
63.82
34.04
TSPerceiver
2.261±0.257
31.76
27.65
70.21
XCM
1.509±0.086
22.31
55.31
42.55
MultiRocket
1.699±0.102
25.24
40.42
57.44
TSiTransformer
1.387±0.019
20.91
44.68
53.19
Table 7: Comparison results of different methods on the BGDB dataset.
Self-supervised time series representation learning uses unlabeled sequences to learn features that transfer to forecasting, classification, and anomaly detection.
Remote Photoplethysmography (rPPG) enables convenient non-contact physiological measurement. Existing Self-Supervised Learning (SSL) methods commonly fall into a correlation trap: they tend to learn the most dominant periodic signals in the data, such as high-energy motion or illumination noise, rather than the faint, true rPPG signal, leading to poor model generalization. To address this, we propose a new SSL paradigm, Physiological Causal Probing (PCP), which treats the latent rPPG signal as the underlying physical source and the resulting pixel chrominance variations as its visual manifestation. Its core idea is to shift from passive correlation learning to active, precise intervention: it intervenes on the video based on a proposed rPPG hypothesis, and verifies whether the post-intervention changes match physical expectations. We propose the Interv-rPPG framework to implement PCP: an rPPG extractor named PhysMambaFormer hypothesizes the rPPG signal, while a Controllable Physiological Signal Editor conducts precise chrominance-domain interventions on videos based on this hypothesis. Interv-rPPG validates the physical realism of the hypothesis through Falsifiability via Nulling' and Axiomatic Equivariance'. Our editor achieves precise editing of the rPPG signal by intervening in the low-frequency chrominance components of the video. Our method improves both in-domain and cross-domain performance on challenging datasets such as VIPL-HR and MMPD. Furthermore, it surpasses the supervised baseline in complex cross-dataset settings, while remaining competitive on clean datasets where the intervention mechanism may introduce slight residual chrominance noise. Extensive experiments, including diagnostic analysis of nuisance sensitivity, demonstrate that the PCP paradigm effectively resists motion and illumination artifacts.
Zhiyi Niu, Xiaoguang Tu, Bo Zhao +3
School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230601, PR China. · College of Aviation Electronic and Electrical Engineering, Civil Aviation Flight University of China, Guanghan 618307, PR China. · School of Computing and Information Technology, Great Bay University, Dongguan 523000, PR China.
Photoplethysmography (PPG) enables low-cost and non-invasive heart rate (HR) monitoring but remains sensitive to waveform distortion, physiological variability, and measurement interference. Temporal waveform characteristics and frequency-domain periodicity provide complementary information for HR estimation; however, effectively exploring and integrating these representations within a compact model remains challenging. This paper proposes TF-IG, a parameter-efficient interaction-gated time-frequency learning framework for PPG-based HR estimation. TF-IG employs a dual-branch architecture in which a temporal branch captures pulse morphology via dilated depthwise-separable convolutions, while a frequency branch extracts cardiac periodicity information from bounded log-spectrogram representations. To combine the two domains, we introduce an interaction-gated fusion module that explicitly models feature agreement and discrepancy, while also learning feature-wise weights to balance temporal and spectral information. A parameter-shared regression head further maps the temporal, spectral, and fused representations into the final HR prediction space. Experiments on the benchmark BIDMC PPG and Respiration Dataset using a subject-exclusive training, validation, and test protocol show that TF-IG achieves a mean absolute error (MAE) of 1.227 BPM, a root mean square error (RMSE) of 2.060 BPM, a Pearson correlation coefficient of 0.9934, and 92.28 percent of estimates within 3 BPM, while remaining parameter-efficient with only 84.6K parameters. Ablation results further demonstrate the complementary contributions of the temporal and frequency branches, interaction-gated fusion, and shared regression head. These results indicate that explicitly modeling cross-domain interactions provides an effective accuracy-complexity trade-off for lightweight PPG-based HR estimation.
Anh Khoi Pham, Phat Lam, Thien Truong Nguyen Ly +1
International University, Vietnam National University HCMC, Ho Chi Minh City, Vietnam · University of New Brunswick, Fredericton, Canada · Ho Chi Minh City University of Technology, Ho Chi Minh City, Vietnam +1
Remote photoplethysmography (rPPG) enables non-contact extraction of blood volume pulse (BVP) signals using consumer-grade cameras. Recent unsupervised rPPG methods learn BVP representations without requiring ground-truth physiological annotations, yet their optimization is often hindered by noisy and unstable gradients, resulting in slow convergence and limited cross-domain generalization. In this paper, we propose FCUS-rPPG, a fast-converging unsupervised rPPG framework with strong generalization capability. Motivated by the observation that BVP representations exhibit both multi-spectral covariation and low-dimensional manifold structure, we design a spectrally shared backbone that facilitates BVP feature disentanglement while improving optimization efficiency. To jointly enhance convergence stability and generalization performance, we further develop a unified optimization framework operating at the gradient, loss-landscape, and feature-representation levels. Specifically, a post-verification masking mechanism filters out misleading gradients according to the weak-amplitude physiological prior of BVP signals; a perturbation-based loss landscape smoothing strategy steers optimization toward more generalizable flat minima; and a noise-aware null-space regularization constrains feature updates to the orthogonal complement of the noise subspace, thereby mitigating noise-induced representation drift. Extensive experiments on five datasets demonstrate that FCUS-rPPG requires only one training epoch, whereas existing methods typically require tens to hundreds of epochs. Notably, FCUS-rPPG consistently achieves state-of-the-art (SOTA) performance in cross-dataset evaluations. This study provides an efficient and robust solution to the real-world deployment of unsupervised rPPG. The source code will be publicly available at https://github.com/JiaJieLee/FCUS-rPPG.
Jiajie Li, Yu Liu, Rencheng Song +2
Department of Biomedical Engineering, and also with Anhui Province Key Laboratory of Measuring Theory and Precision Instrument, Hefei University of Technology, Hefei 230009, China · Department of Electronic Engineering and Information Science, University of Science and Technology of China, Hefei 230026, China