SignalMC-MED: A Multimodal Benchmark for Evaluating Biosignal Foundation Models on Single-Lead ECG and PPG

Fredrik K. Gustafsson; Xiao Gu; Mattia Carletti; Patitapaban Palo; David W. Eyre; David A. Clifton

cs.LGMar 10, 2026

SignalMC-MED: A Multimodal Benchmark for Evaluating Biosignal Foundation Models on Single-Lead ECG and PPG

Authors: Fredrik K. Gustafsson, Xiao Gu, Mattia Carletti, Patitapaban Palo, David W. Eyre, David A. Clifton

Organizations: Department of Engineering Science, University of Oxford · NIHR Health Protection Research Unit in Healthcare Associated Infections and Antimicrobial Resistance · Big Data Institute, Nuffield Department of Population Health, University of Oxford · NIHR Oxford Biomedical Research Centre · Oxford Suzhou Centre for Advanced Research, University of Oxford

Abstract

Recent biosignal foundation models (FMs) have demonstrated promising performance across diverse clinical prediction tasks, yet systematic evaluation on long-duration multimodal data remains limited. We introduce SignalMC-MED, a benchmark for evaluating biosignal FMs on synchronized single-lead electrocardiogram (ECG) and photoplethysmogram (PPG) data. Derived from the MC-MED dataset, SignalMC-MED comprises 22,256 visits with 10-minute overlapping ECG and PPG signals, and includes 20 clinically relevant tasks spanning prediction of demographics, emergency department disposition, laboratory value regression, and detection of prior ICD-10 diagnoses. Using this benchmark, we perform a systematic evaluation of representative time-series and biosignal FMs across ECG-only, PPG-only, and ECG + PPG settings. We find that domain-specific biosignal FMs consistently outperform general time-series models, and that multimodal ECG + PPG fusion yields robust improvements over unimodal inputs. Moreover, using the full 10-minute signal consistently outperforms shorter segments, and larger model variants do not reliably outperform smaller ones. Hand-crafted ECG domain features provide a strong baseline and offer complementary value when combined with learned FM representations. Together, these results establish SignalMC-MED as a standardized benchmark and provide practical guidance for evaluating and deploying biosignal FMs.

Figures & tables

Panel of Figure 1 — Figure 1: (a) Overview of the SignalMC-MED benchmark and evaluation framework. For each of the 22,256 visits, synchronized 10-minute single-lead ECG ( black ) and PPG ( red ) signals are divided into non-overlapping 10-second segments. A frozen FM ( blue ) extracts a feature vector for each 10-second segment. The resulting segment-level features are aggregated via mean pooling ( $\Sigma$ ) to form a visit-level representation, which is used to train a linear prediction model ( green ) across 20 downstream tasks: age regression, sex classification, emergency department disposition classification, laboratory value regression (8 tasks), and prior ICD-10 diagnosis classification (9 tasks). (b) Examples of synchronized ECG and PPG signals from three different visits. For each visit, five 3-second segments are shown, evenly sampled from the full 10-minute signal (segment start times indicated above each panel). The signals illustrate inter-patient variability in rhythm, morphology, amplitude and noise characteristics. Additional examples are provided in Figure S1 in the supplementary material.

Model Name	Size	Feature Dimension	Sampling Frequency	Pretraining Data
MOMENT-small [ 10 ]	35M	512	Mixed	13M time series; 1.2B timestamps
MOMENT-base	110M	768	—	—
MOMENT-large	341M	1024	—	—
Chronos-Bolt-tiny [ 4 ]	9M	256	Mixed	100B time series observations
Chronos-Bolt-mini	21M	384	—	—
Chronos-Bolt-small	48M	512	—	—

Table 1: Overview of all evaluated FMs , including model size (number of parameters) and expected input sampling frequency.

Model Name	Modality	Age reg	Sex cls	ED_dispo cls	Labs reg	ICD-10 cls
Domain Features	ECG-only	0.678 $\pm$ 0.002	0.652 $\pm$ 0.004	0.711 $\pm$ 0.003	0.319 $\pm$ 0.002	0.748 $\pm$ 0.003
Domain Features	PPG-only	0.684 $\pm$ 0.052	0.737 $\pm$ 0.002	0.676 $\pm$ 0.002	0.213 $\pm$ 0.018	0.677 $\pm$ 0.002
Domain Features (concat)	ECG + PPG	0.751 $\pm$ 0.044	0.769 $\pm$ 0.002	0.713 $\pm$ 0.001	0.299 $\pm$ 0.015	0.736 $\pm$ 0.005
MOMENT-base	ECG-only	0.641 $\pm$ 0.002	0.741 $\pm$ 0.002	0.662 $\pm$ 0.004	0.239 $\pm$ 0.005	0.671 $\pm$ 0.005
MOMENT-base	PPG-only	0.702 $\pm$ 0.002	0.724 $\pm$ 0.003	0.676 $\pm$ 0.002	0.240 $\pm$ 0.002	0.670 $\pm$ 0.002
MOMENT-base	ECG + PPG	0.757 $\pm$ 0.002	0.776 $\pm$ 0.001	0.678 $\pm$ 0.003	0.283 $\pm$ 0.004	0.692 $\pm$ 0.004

Table 2: Main model comparison on the test set , with results aggregated across training using 10%, 25%, 50%, and 100% of the train visits. Regression tasks are evaluated using Pearson correlation (

\uparrow

), and classification tasks using AUROC (

\uparrow

). Columns correspond to aggregated performance for: age regression, sex classification, ED disposition classification, laboratory value regression (mean across eight laboratory tests), and prior ICD-10 diagnosis classification (mean across nine diagnosis groups). For each train percentage, training visits are sampled with replacement, and both downstream training and hyperparameter selection are repeated five times. Each table entry represents the mean

\pm

standard deviation across resampling repetitions, after aggregation across tasks within each category and across train visit percentages. Bold red marks the best mean value in each column, bold marks the second-best, and underline marks the third-best.

Table 4

Task	Label / Description	Label Source	ECG-only	PPG-only	ECG + PPG
Age reg	“Age”	visits.csv	0.785 $\pm$ 0.002	0.799 $\pm$ 0.002	0.842 $\pm$ 0.001
Sex cls	“M” vs “F”	visits.csv	0.816 $\pm$ 0.002	0.767 $\pm$ 0.003	0.840 $\pm$ 0.002
ED_dispo cls	“Discharge” vs {“Inpatient”, “Observation”, “ICU”}	visits.csv	0.744 $\pm$ 0.001	0.726 $\pm$ 0.001	0.750 $\pm$ 0.002
Potassium reg	“POTASSIUM”	labs.csv	0.386 $\pm$ 0.007	0.190 $\pm$ 0.006	0.386 $\pm$ 0.006
Calcium reg	“CALCIUM”		0.209 $\pm$ 0.010	0.183 $\pm$ 0.015	0.265 $\pm$ 0.010
eGFR reg	“EGFR REFIT WITHOUT RACE (2021)”		0.591 $\pm$ 0.006	0.567 $\pm$ 0.005	0.613 $\pm$ 0.003

Table 4: SignalMC-MED benchmark task summary and CSFM-base reference performance . Overview of the 20 included tasks, grouped into two demographics tasks, ED disposition classification, eight laboratory value regression tasks, and nine prior ICD-10 diagnosis classification tasks. CSFM-base performance across modalities on the val set when using 100% of the train set visits is shown as a reference for relative task difficulty. Training visits are sampled with replacement, and both downstream training and hyperparameter selection are repeated five times; results are reported as mean

\pm

standard deviation across resampling repetitions. Bold red marks the best mean value in each row. Full results with 100%, 50%, 25%, and 10% of the train visits for all eight evaluated models are provided in Table S4 - S15 .

Panel of Figure 2 — Figure 2: Main model comparison on the test set , with average classification task performance (eleven tasks, left column) and regression task performance (nine tasks, right column) shown as a function of the percentage of train visits used (10%, 25%, 50%, 100%). Rows correspond to ECG-only (top), PPG-only (middle), and ECG + PPG (bottom) inputs. Each curve represents one evaluated model and shows the mean performance across five resampling repetitions (standard deviations omitted for visual clarity).

Panel of Figure 3 — Figure 3: Multimodal fusion comparison on the val set , comparing the ECG-only, PPG-only, and ECG + PPG performance for three representative models (MOMENT-base, xECG-10min, CSFM-base). Average classification (left) and regression (right) task performance is shown as a function of the percentage of train visits used. Results for the five remaining models are provided in Figure S2 .

Panel of Figure 4 — Figure 4: Signal length comparison on the val set , showing how the ECG + PPG performance is affected by increasing the signal length from 10 seconds to 10 minutes for three representative models (MOMENT-base, xECG, CSFM-base). Results for the five remaining models are provided in Figure S3 , and for ECG and PPG domain features in Figure S4 .

Panel of Figure 5 — Figure 5: Multimodal fusion comparison across signal lengths on the val set , comparing the ECG-only (dashed lines) and ECG + PPG (solid lines) performance with multiple signal lengths (10 seconds, 1 minute, 10 minutes) for three representative models (MOMENT-base, xECG, CSFM-base). Results for the five remaining models are provided in Figure S5 .

Panel of Figure 6 — Figure 6: Effect of concatenating ECG domain features with learned FM representations on the val set . Comparison of ECG-only performance before (solid lines) and after (dashed lines) concatenation of hand-crafted ECG domain features to the frozen FM feature vectors, shown for each of the seven evaluated FMs. The same comparison for PPG domain features is provided in Figure S8 .

Panel of Figure 7 — Figure 7: UMAP visualizations colored by age group . UMAP [ 20 ] projections of the train set ECG + PPG feature vectors for six selected models, colored according to three age groups ( $<30$ years, $30-85$ years, $>85$ years). CSFM-base, xECG-10min and ECGFounder exhibit clearer age-related structure compared to general time-series FMs (MOMENT-base, Chronos-Bolt-small) and PaPaGei.

Figure 8: Detailed evaluation framework for SignalMC-MED. A 10-minute synchronized single-lead ECG ( black ) and PPG ( red ) signal is divided into non-overlapping 10-second segments. ECG and PPG segments are processed separately by a frozen FM ( blue ) to extract a feature vector for each 10-second segment. The evaluated FM is applied to both ECG and PPG inputs, regardless of its pretraining modality. The resulting segment-level features are aggregated via mean pooling ( $\Sigma$ ) to form ECG-only and PPG-only visit-level feature vectors, which are then averaged to obtain an ECG + PPG feature vector of the same dimensionality. Separate linear models are trained on the ECG-only, PPG-only, and ECG + PPG representations to generate task-specific predictions across the 20 benchmark tasks.

Appendix figures & tables28 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1 : Additional synchronized ECG and PPG signals in SignalMC-MED . An extension of Figure 1(b) , with seven more examples of synchronized single-lead ECG ( black ) and PPG ( red ) signals from visits in the train set.

Panel of Figure 14 — Figure S2 : Multimodal fusion comparison on val for the remaining models . An extension of Figure 3 , comparing the ECG-only, PPG-only, and ECG + PPG performance for the remaining models (PaPaGei, Chronos-Bolt-small, D-BETA, ECGFounder, Domain Features).

Panel of Figure 15 — Figure S3 : Signal length comparison on val for the remaining models . An extension of Figure 4 , showing how the ECG + PPG performance is affected by increasing the signal length from 10 seconds to 10 minutes for the remaining models (PaPaGei, Chronos-Bolt-small, D-BETA, ECGFounder, Domain Features).

Panel of Figure 16 — Figure S4 : Signal length comparison on val for hand-crafted ECG and PPG domain features , showing how the performance is affected by increasing the signal length from 10 seconds to 10 minutes for ECG, and from 1 minute to 10 minutes for PPG. This analysis complements Figure 4 & S3 .

Panel of Figure 17 — Figure S5 : Multimodal fusion comparison across signal lengths on val for the remaining models . An extension of Figure 5 , comparing the PPG-only and ECG + PPG performance for PaPaGei, and comparing the ECG-only and ECG + PPG performance for Chronos-Bolt-small, D-BETA, ECGFounder, and Domain Features.

Panel of Figure 18 — Figure S6 : Model size comparison on the val set , comparing the ECG + PPG performance of small/base/large model variants for MOMENT, of tiny/mini/small/base variants for Chronos-Bolt, and of tiny/base/large variants for CSFM.

Panel of Figure 19 — Figure S7 : Model size comparison with PCA on the val set . The same analysis as Figure S6 , but with all visit-level feature vectors projected down to 256 dimensions using PCA prior to downstream linear probing. Performance trends remain similar: larger model variants do not reliably outperform smaller ones. This indicates that differences in feature dimensionality are not the primary explanation for the occasional underperformance of larger models.

Panel of Figure 20 — Figure S8 : Effect of concatenating PPG domain features with learned FM representations on the val set . Comparison of PPG-only performance before (solid lines) and after (dashed lines) concatenation of hand-crafted PPG domain features to the frozen FM feature vectors, shown for each of the seven evaluated FMs. Identical analysis as in Figure 6 , but for PPG domain features.

Panel of Figure 21 — Figure S9 : Multimodal fusion comparison for CSFM on the val set . Comparison between feature-level fusion (solid lines), obtained by averaging independently extracted ECG-only and PPG-only features, and CSFM’s internal multimodal fusion mechanism (dashed lines), which jointly processes ECG + PPG as a two-channel input. Results are shown for multiple signal lengths (10 seconds, 1 minute, 10 minutes) and all three CSFM model variants (tiny, base, large). Feature-level fusion consistently achieves slightly better performance.

Panel of Figure 22 — Figure S10 : Comparison of CSFM-base and Domain Features across signal lengths on the val set , showing ECG-only, PPG-only, and ECG + PPG performance of CSFM-base (solid lines) and Domain Features (dashed lines) with multiple signal lengths. Especially for ECG-only, the performance gap narrows with longer signals.

Panel of Figure 23 — Figure S11 : UMAP visualizations colored by atrial fibrillation (I48) label . UMAP projections of the train set ECG + PPG feature vectors for six selected models. CSFM-base, xECG-10min and ECGFounder exhibit clearer separation between positive and negative cases compared to general time-series FMs (MOMENT-base, Chronos-Bolt-small) and PaPaGei.

Model Name	Modality	Age reg	Sex cls	ED_dispo cls	Labs reg	ICD-10 cls
MOMENT-small	ECG-only	0.668	0.750	0.696	0.239	0.685
MOMENT-small	PPG-only	0.676	0.721	0.665	0.224	0.652
MOMENT-small	ECG + PPG	0.747	0.781	0.706	0.272	0.701
MOMENT-base ∗	ECG-only	0.680	0.757	0.695	0.253	0.683
MOMENT-base ∗	PPG-only	0.691	0.725	0.680	0.233	0.660
MOMENT-base ∗	ECG + PPG	0.762	0.791	0.703	0.284	0.700

Appendix

Table S1 : Comparison of all model variants on the val set , with results averaged across training with 10%, 25%, 50%, and 100% of the train visits (five resampling repeats per train fraction). Regression tasks are evaluated using Pearson correlation (

\uparrow

), and classification tasks using AUROC (

\uparrow

). Bold red marks the best result in each column, bold marks the second-best, and underline marks the third-best. Based on these results and the corresponding ranking in Table S2 , one variant each for MOMENT, Chronos-Bolt, xECG, CSFM, and the PPG domain features is selected for inclusion in the main model comparison on the test set. These selected variants are here indicated with italics and a superscript ∗ next to the model name.

Rank	Model Name	Modality	Model Ranks	Mean Model Rank ( $\downarrow$ )
1	CSFM-base	ECG + PPG	1,4,1,1,1	1.6
2	CSFM-large	ECG + PPG	2,1,4,3,5	3.0
3	xECG-10min	ECG + PPG	3,2,3,2,7	3.4
4	xECG	ECG + PPG	4,3,5,4,8	4.8
5	CSFM-base	ECG-only	9,9,2,5,2	5.4
6	xECG-10min	ECG-only	11,6,6,6,10	7.8

Appendix

Table S2 : Ranking of all model variants on the val set , computed jointly across ECG-only, PPG-only, and ECG + PPG. Mean model rank (

\downarrow

) across the five aggregated task categories based on Table S1 . Based on this ranking, one variant each for MOMENT, Chronos-Bolt, xECG, CSFM, and the PPG domain features is selected for inclusion in the main model comparison on the test set.

Rank	Model Name	Modality	Model Ranks	Mean Model Rank ( $\downarrow$ )
1	CSFM-base	ECG + PPG	1,1,1,1,3	1.4
2	xECG-10min	ECG + PPG	2,2,6,3,6	3.8
3	CSFM-base	ECG-only	7,5,5,2,4	4.6
4	ECGFounder	ECG + PPG	3,4,8,5,5	5.0
5	Domain Features (concat)	ECG + PPG	8,7,2,9,2	5.6
6	xECG-1omin	ECG-only	10,3,7,4,8	6.4

Appendix

Table S3 : Main model ranking on the test set , computed jointly across ECG-only, PPG-only, and ECG + PPG. Mean model rank (

\downarrow

) across the five aggregated task categories based on Table 2 . This table complements the modality-specific rankings shown in Table 3 .

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.696 $\pm$ 0.001	0.694 $\pm$ 0.001	0.710 $\pm$ 0.001	0.559 $\pm$ 0.001	0.688 $\pm$ 0.005	0.778 $\pm$ 0.003	0.785 $\pm$ 0.001	0.785 $\pm$ 0.002
Sex cls	0.676 $\pm$ 0.001	0.774 $\pm$ 0.002	0.770 $\pm$ 0.003	0.675 $\pm$ 0.002	0.730 $\pm$ 0.003	0.788 $\pm$ 0.003	0.822 $\pm$ 0.002	0.816 $\pm$ 0.002
ED_dispo cls	0.729 $\pm$ 0.001	0.703 $\pm$ 0.003	0.712 $\pm$ 0.005	0.655 $\pm$ 0.003	0.695 $\pm$ 0.004	0.725 $\pm$ 0.003	0.745 $\pm$ 0.001	0.744 $\pm$ 0.001
Potassium reg	0.268 $\pm$ 0.004	0.322 $\pm$ 0.008	0.304 $\pm$ 0.010	0.194 $\pm$ 0.016	0.188 $\pm$ 0.023	0.315 $\pm$ 0.005	0.383 $\pm$ 0.007	0.386 $\pm$ 0.007
Calcium reg	0.155 $\pm$ 0.011	0.150 $\pm$ 0.007	0.127 $\pm$ 0.018	0.077 $\pm$ 0.012	0.098 $\pm$ 0.016	0.155 $\pm$ 0.014	0.205 $\pm$ 0.010	0.209 $\pm$ 0.010
eGFR reg	0.546 $\pm$ 0.001	0.494 $\pm$ 0.007	0.509 $\pm$ 0.005	0.369 $\pm$ 0.006	0.481 $\pm$ 0.015	0.579 $\pm$ 0.007	0.585 $\pm$ 0.003	0.591 $\pm$ 0.006

Appendix

Table S4 : Model comparison of ECG-only val performance on all 20 benchmark tasks when using 100% of the train visits . Tasks are grouped into two demographics tasks, ED disposition classification, eight laboratory value regression tasks, and nine prior ICD-10 diagnosis classification tasks. Performance on the val set when using 100% of the train set visits. Training visit are sampled with replacement and both downstream training and hyperparameter selection are repeated five times, results are mean

\pm

std across repetitions. Regression tasks are evaluated using Pearson correlation (

\uparrow

), and classification tasks using AUROC (

\uparrow

). Bold red marks the best mean value in each row.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.734 $\pm$ 0.003	0.704 $\pm$ 0.002	0.726 $\pm$ 0.002	0.673 $\pm$ 0.003	0.663 $\pm$ 0.007	0.703 $\pm$ 0.005	0.778 $\pm$ 0.002	0.799 $\pm$ 0.002
Sex cls	0.744 $\pm$ 0.004	0.731 $\pm$ 0.004	0.733 $\pm$ 0.005	0.691 $\pm$ 0.002	0.715 $\pm$ 0.004	0.735 $\pm$ 0.004	0.767 $\pm$ 0.003	0.767 $\pm$ 0.003
ED_dispo cls	0.706 $\pm$ 0.003	0.688 $\pm$ 0.005	0.689 $\pm$ 0.001	0.673 $\pm$ 0.005	0.671 $\pm$ 0.006	0.686 $\pm$ 0.003	0.707 $\pm$ 0.004	0.726 $\pm$ 0.001
Potassium reg	0.152 $\pm$ 0.016	0.175 $\pm$ 0.016	0.178 $\pm$ 0.008	0.164 $\pm$ 0.014	0.115 $\pm$ 0.025	0.135 $\pm$ 0.025	0.145 $\pm$ 0.007	0.190 $\pm$ 0.006
Calcium reg	0.149 $\pm$ 0.010	0.122 $\pm$ 0.009	0.156 $\pm$ 0.015	0.117 $\pm$ 0.012	0.100 $\pm$ 0.009	0.121 $\pm$ 0.040	0.165 $\pm$ 0.016	0.183 $\pm$ 0.015
eGFR reg	0.485 $\pm$ 0.005	0.439 $\pm$ 0.004	0.479 $\pm$ 0.004	0.436 $\pm$ 0.007	0.399 $\pm$ 0.014	0.419 $\pm$ 0.063	0.513 $\pm$ 0.007	0.567 $\pm$ 0.005

Appendix

Table S5 : Model comparison of PPG-only val performance on all 20 benchmark tasks when using 100% of the train visits . Identical format and content as Table S4 , but for PPG-only . Performance on the val set when using 100% of the train set visits.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.803 $\pm$ 0.003	0.776 $\pm$ 0.001	0.781 $\pm$ 0.002	0.697 $\pm$ 0.002	0.747 $\pm$ 0.003	0.822 $\pm$ 0.002	0.831 $\pm$ 0.001	0.842 $\pm$ 0.001
Sex cls	0.771 $\pm$ 0.004	0.810 $\pm$ 0.001	0.808 $\pm$ 0.003	0.728 $\pm$ 0.002	0.770 $\pm$ 0.001	0.819 $\pm$ 0.002	0.848 $\pm$ 0.001	0.840 $\pm$ 0.002
ED_dispo cls	0.734 $\pm$ 0.003	0.710 $\pm$ 0.006	0.720 $\pm$ 0.003	0.676 $\pm$ 0.003	0.702 $\pm$ 0.005	0.732 $\pm$ 0.002	0.750 $\pm$ 0.002	0.750 $\pm$ 0.002
Potassium reg	0.238 $\pm$ 0.017	0.333 $\pm$ 0.009	0.326 $\pm$ 0.005	0.216 $\pm$ 0.010	0.192 $\pm$ 0.019	0.308 $\pm$ 0.010	0.389 $\pm$ 0.005	0.386 $\pm$ 0.006
Calcium reg	0.203 $\pm$ 0.010	0.181 $\pm$ 0.011	0.180 $\pm$ 0.009	0.096 $\pm$ 0.003	0.130 $\pm$ 0.021	0.189 $\pm$ 0.007	0.231 $\pm$ 0.015	0.265 $\pm$ 0.010
eGFR reg	0.573 $\pm$ 0.008	0.515 $\pm$ 0.008	0.541 $\pm$ 0.007	0.448 $\pm$ 0.008	0.488 $\pm$ 0.017	0.585 $\pm$ 0.006	0.596 $\pm$ 0.003	0.613 $\pm$ 0.003

Appendix

Table S6 : Model comparison of ECG + PPG val performance on all 20 benchmark tasks when using 100% of the train visits . Identical format and content as Table S4 , but for ECG + PPG . Performance on the val set when using 100% of the train set visits.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.694 $\pm$ 0.001	0.685 $\pm$ 0.003	0.699 $\pm$ 0.004	0.554 $\pm$ 0.003	0.672 $\pm$ 0.009	0.775 $\pm$ 0.002	0.776 $\pm$ 0.002	0.780 $\pm$ 0.002
Sex cls	0.674 $\pm$ 0.005	0.766 $\pm$ 0.006	0.757 $\pm$ 0.006	0.673 $\pm$ 0.006	0.725 $\pm$ 0.002	0.779 $\pm$ 0.003	0.817 $\pm$ 0.002	0.811 $\pm$ 0.006
ED_dispo cls	0.728 $\pm$ 0.002	0.699 $\pm$ 0.008	0.704 $\pm$ 0.003	0.653 $\pm$ 0.003	0.691 $\pm$ 0.005	0.720 $\pm$ 0.004	0.733 $\pm$ 0.004	0.737 $\pm$ 0.006
Potassium reg	0.268 $\pm$ 0.009	0.302 $\pm$ 0.017	0.279 $\pm$ 0.008	0.179 $\pm$ 0.005	0.178 $\pm$ 0.022	0.294 $\pm$ 0.026	0.370 $\pm$ 0.010	0.377 $\pm$ 0.005
Calcium reg	0.150 $\pm$ 0.016	0.133 $\pm$ 0.016	0.109 $\pm$ 0.012	0.088 $\pm$ 0.012	0.077 $\pm$ 0.025	0.142 $\pm$ 0.011	0.199 $\pm$ 0.016	0.195 $\pm$ 0.005
eGFR reg	0.540 $\pm$ 0.005	0.493 $\pm$ 0.004	0.500 $\pm$ 0.006	0.352 $\pm$ 0.009	0.459 $\pm$ 0.019	0.559 $\pm$ 0.007	0.581 $\pm$ 0.003	0.580 $\pm$ 0.005

Appendix

Table S7 : Model comparison of ECG-only val performance on all 20 benchmark tasks when using 50% of the train visits . Tasks are grouped into two demographics tasks, ED disposition classification, eight laboratory value regression tasks, and nine prior ICD-10 diagnosis classification tasks. Performance on the val set when using 50% of the train set visits. Training visit are sampled with replacement and both downstream training and hyperparameter selection are repeated five times, results are mean

\pm

std across repetitions. Regression tasks are evaluated using Pearson correlation (

\uparrow

), and classification tasks using AUROC (

\uparrow

). Bold red marks the best mean value in each row.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.718 $\pm$ 0.013	0.701 $\pm$ 0.004	0.716 $\pm$ 0.002	0.674 $\pm$ 0.003	0.604 $\pm$ 0.068	0.655 $\pm$ 0.083	0.768 $\pm$ 0.003	0.793 $\pm$ 0.002
Sex cls	0.740 $\pm$ 0.004	0.729 $\pm$ 0.002	0.726 $\pm$ 0.004	0.692 $\pm$ 0.001	0.711 $\pm$ 0.008	0.734 $\pm$ 0.006	0.764 $\pm$ 0.006	0.765 $\pm$ 0.004
ED_dispo cls	0.704 $\pm$ 0.004	0.687 $\pm$ 0.004	0.682 $\pm$ 0.002	0.670 $\pm$ 0.004	0.664 $\pm$ 0.003	0.677 $\pm$ 0.008	0.703 $\pm$ 0.002	0.721 $\pm$ 0.004
Potassium reg	0.136 $\pm$ 0.008	0.173 $\pm$ 0.010	0.167 $\pm$ 0.018	0.172 $\pm$ 0.007	0.095 $\pm$ 0.036	0.136 $\pm$ 0.019	0.169 $\pm$ 0.014	0.175 $\pm$ 0.008
Calcium reg	0.133 $\pm$ 0.018	0.125 $\pm$ 0.016	0.157 $\pm$ 0.010	0.101 $\pm$ 0.006	0.100 $\pm$ 0.015	0.119 $\pm$ 0.026	0.171 $\pm$ 0.015	0.188 $\pm$ 0.018
eGFR reg	0.467 $\pm$ 0.009	0.425 $\pm$ 0.005	0.458 $\pm$ 0.006	0.424 $\pm$ 0.002	0.322 $\pm$ 0.020	0.381 $\pm$ 0.041	0.503 $\pm$ 0.005	0.552 $\pm$ 0.006

Appendix

Table S8 : Model comparison of PPG-only val performance on all 20 benchmark tasks when using 50% of the train visits . Identical format and content as Table S7 , but for PPG-only . Performance on the val set when using 50% of the train set visits.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.797 $\pm$ 0.003	0.768 $\pm$ 0.003	0.774 $\pm$ 0.002	0.687 $\pm$ 0.005	0.730 $\pm$ 0.011	0.818 $\pm$ 0.003	0.823 $\pm$ 0.002	0.839 $\pm$ 0.002
Sex cls	0.767 $\pm$ 0.003	0.798 $\pm$ 0.004	0.798 $\pm$ 0.005	0.727 $\pm$ 0.004	0.765 $\pm$ 0.002	0.815 $\pm$ 0.003	0.842 $\pm$ 0.002	0.835 $\pm$ 0.002
ED_dispo cls	0.731 $\pm$ 0.005	0.708 $\pm$ 0.004	0.715 $\pm$ 0.005	0.673 $\pm$ 0.005	0.695 $\pm$ 0.005	0.724 $\pm$ 0.003	0.736 $\pm$ 0.004	0.745 $\pm$ 0.006
Potassium reg	0.223 $\pm$ 0.017	0.315 $\pm$ 0.012	0.297 $\pm$ 0.007	0.199 $\pm$ 0.007	0.189 $\pm$ 0.012	0.269 $\pm$ 0.035	0.384 $\pm$ 0.012	0.376 $\pm$ 0.006
Calcium reg	0.192 $\pm$ 0.020	0.165 $\pm$ 0.017	0.165 $\pm$ 0.014	0.104 $\pm$ 0.015	0.100 $\pm$ 0.022	0.164 $\pm$ 0.008	0.220 $\pm$ 0.017	0.254 $\pm$ 0.010
eGFR reg	0.568 $\pm$ 0.004	0.508 $\pm$ 0.008	0.529 $\pm$ 0.004	0.423 $\pm$ 0.011	0.467 $\pm$ 0.025	0.564 $\pm$ 0.008	0.588 $\pm$ 0.003	0.598 $\pm$ 0.004

Appendix

Table S9 : Model comparison of ECG + PPG val performance on all 20 benchmark tasks when using 50% of the train visits . Identical format and content as Table S7 , but for ECG + PPG . Performance on the val set when using 50% of the train set visits.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.689 $\pm$ 0.004	0.677 $\pm$ 0.005	0.681 $\pm$ 0.003	0.538 $\pm$ 0.007	0.645 $\pm$ 0.026	0.762 $\pm$ 0.013	0.769 $\pm$ 0.002	0.775 $\pm$ 0.004
Sex cls	0.665 $\pm$ 0.006	0.753 $\pm$ 0.003	0.737 $\pm$ 0.004	0.668 $\pm$ 0.005	0.713 $\pm$ 0.004	0.764 $\pm$ 0.004	0.806 $\pm$ 0.003	0.802 $\pm$ 0.004
ED_dispo cls	0.725 $\pm$ 0.002	0.691 $\pm$ 0.004	0.690 $\pm$ 0.007	0.644 $\pm$ 0.003	0.693 $\pm$ 0.003	0.711 $\pm$ 0.006	0.724 $\pm$ 0.002	0.730 $\pm$ 0.005
Potassium reg	0.239 $\pm$ 0.018	0.291 $\pm$ 0.012	0.262 $\pm$ 0.018	0.175 $\pm$ 0.025	0.187 $\pm$ 0.020	0.285 $\pm$ 0.008	0.345 $\pm$ 0.014	0.362 $\pm$ 0.009
Calcium reg	0.141 $\pm$ 0.016	0.111 $\pm$ 0.011	0.101 $\pm$ 0.014	0.069 $\pm$ 0.010	0.050 $\pm$ 0.039	0.133 $\pm$ 0.013	0.183 $\pm$ 0.008	0.189 $\pm$ 0.012
eGFR reg	0.536 $\pm$ 0.005	0.468 $\pm$ 0.005	0.474 $\pm$ 0.010	0.352 $\pm$ 0.010	0.467 $\pm$ 0.021	0.554 $\pm$ 0.014	0.558 $\pm$ 0.007	0.567 $\pm$ 0.005

Appendix

Table S10 : Model comparison of ECG-only val performance on all 20 benchmark tasks when using 25% of the train visits . Tasks are grouped into two demographics tasks, ED disposition classification, eight laboratory value regression tasks, and nine prior ICD-10 diagnosis classification tasks. Performance on the val set when using 25% of the train set visits. Training visit are sampled with replacement and both downstream training and hyperparameter selection are repeated five times, results are mean

\pm

std across repetitions. Regression tasks are evaluated using Pearson correlation (

\uparrow

), and classification tasks using AUROC (

\uparrow

). Bold red marks the best mean value in each row.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.711 $\pm$ 0.007	0.690 $\pm$ 0.003	0.704 $\pm$ 0.003	0.661 $\pm$ 0.003	0.580 $\pm$ 0.082	0.696 $\pm$ 0.015	0.754 $\pm$ 0.003	0.783 $\pm$ 0.004
Sex cls	0.729 $\pm$ 0.005	0.724 $\pm$ 0.004	0.717 $\pm$ 0.004	0.688 $\pm$ 0.003	0.699 $\pm$ 0.006	0.723 $\pm$ 0.004	0.754 $\pm$ 0.004	0.757 $\pm$ 0.003
ED_dispo cls	0.689 $\pm$ 0.007	0.678 $\pm$ 0.005	0.676 $\pm$ 0.007	0.651 $\pm$ 0.008	0.660 $\pm$ 0.007	0.674 $\pm$ 0.009	0.696 $\pm$ 0.005	0.720 $\pm$ 0.002
Potassium reg	0.092 $\pm$ 0.036	0.156 $\pm$ 0.018	0.144 $\pm$ 0.018	0.148 $\pm$ 0.008	0.088 $\pm$ 0.024	0.099 $\pm$ 0.026	0.146 $\pm$ 0.022	0.154 $\pm$ 0.019
Calcium reg	0.107 $\pm$ 0.021	0.106 $\pm$ 0.010	0.130 $\pm$ 0.017	0.081 $\pm$ 0.020	0.085 $\pm$ 0.018	0.079 $\pm$ 0.033	0.168 $\pm$ 0.017	0.188 $\pm$ 0.027
eGFR reg	0.412 $\pm$ 0.021	0.401 $\pm$ 0.008	0.422 $\pm$ 0.013	0.419 $\pm$ 0.006	0.318 $\pm$ 0.027	0.407 $\pm$ 0.015	0.476 $\pm$ 0.008	0.539 $\pm$ 0.002

Appendix

Table S11 : Model comparison of PPG-only val performance on all 20 benchmark tasks when using 25% of the train visits . Identical format and content as Table S10 , but for PPG-only . Performance on the val set when using 25% of the train set visits.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.789 $\pm$ 0.003	0.759 $\pm$ 0.004	0.762 $\pm$ 0.003	0.669 $\pm$ 0.002	0.696 $\pm$ 0.026	0.798 $\pm$ 0.028	0.816 $\pm$ 0.002	0.833 $\pm$ 0.002
Sex cls	0.757 $\pm$ 0.006	0.785 $\pm$ 0.002	0.781 $\pm$ 0.002	0.720 $\pm$ 0.004	0.748 $\pm$ 0.004	0.800 $\pm$ 0.003	0.829 $\pm$ 0.002	0.826 $\pm$ 0.004
ED_dispo cls	0.722 $\pm$ 0.007	0.697 $\pm$ 0.005	0.701 $\pm$ 0.004	0.660 $\pm$ 0.003	0.697 $\pm$ 0.003	0.719 $\pm$ 0.005	0.728 $\pm$ 0.003	0.737 $\pm$ 0.004
Potassium reg	0.157 $\pm$ 0.034	0.297 $\pm$ 0.007	0.274 $\pm$ 0.014	0.187 $\pm$ 0.030	0.191 $\pm$ 0.029	0.258 $\pm$ 0.014	0.360 $\pm$ 0.014	0.361 $\pm$ 0.004
Calcium reg	0.167 $\pm$ 0.012	0.141 $\pm$ 0.011	0.144 $\pm$ 0.023	0.082 $\pm$ 0.012	0.074 $\pm$ 0.044	0.152 $\pm$ 0.012	0.203 $\pm$ 0.012	0.248 $\pm$ 0.012
eGFR reg	0.523 $\pm$ 0.014	0.492 $\pm$ 0.005	0.508 $\pm$ 0.013	0.418 $\pm$ 0.007	0.469 $\pm$ 0.031	0.549 $\pm$ 0.025	0.570 $\pm$ 0.004	0.586 $\pm$ 0.002

Appendix

Table S12 : Model comparison of ECG + PPG val performance on all 20 benchmark tasks when using 25% of the train visits . Identical format and content as Table S10 , but for ECG + PPG . Performance on the val set when using 25% of the train set visits.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.681 $\pm$ 0.003	0.663 $\pm$ 0.002	0.657 $\pm$ 0.004	0.517 $\pm$ 0.008	0.521 $\pm$ 0.139	0.726 $\pm$ 0.042	0.755 $\pm$ 0.004	0.761 $\pm$ 0.003
Sex cls	0.656 $\pm$ 0.011	0.735 $\pm$ 0.006	0.715 $\pm$ 0.004	0.665 $\pm$ 0.009	0.706 $\pm$ 0.004	0.745 $\pm$ 0.007	0.791 $\pm$ 0.004	0.788 $\pm$ 0.003
ED_dispo cls	0.712 $\pm$ 0.011	0.687 $\pm$ 0.005	0.677 $\pm$ 0.014	0.638 $\pm$ 0.007	0.690 $\pm$ 0.010	0.708 $\pm$ 0.015	0.718 $\pm$ 0.004	0.726 $\pm$ 0.009
Potassium reg	0.220 $\pm$ 0.022	0.249 $\pm$ 0.012	0.229 $\pm$ 0.014	0.141 $\pm$ 0.020	0.124 $\pm$ 0.064	0.232 $\pm$ 0.021	0.310 $\pm$ 0.014	0.329 $\pm$ 0.011
Calcium reg	0.125 $\pm$ 0.013	0.094 $\pm$ 0.012	0.100 $\pm$ 0.019	0.062 $\pm$ 0.016	0.060 $\pm$ 0.010	0.111 $\pm$ 0.005	0.163 $\pm$ 0.006	0.163 $\pm$ 0.014
eGFR reg	0.519 $\pm$ 0.010	0.440 $\pm$ 0.015	0.440 $\pm$ 0.007	0.333 $\pm$ 0.016	0.423 $\pm$ 0.038	0.483 $\pm$ 0.060	0.541 $\pm$ 0.005	0.558 $\pm$ 0.011

Appendix

Table S13 : Model comparison of ECG-only val performance on all 20 benchmark tasks when using 10% of the train visits . Tasks are grouped into two demographics tasks, ED disposition classification, eight laboratory value regression tasks, and nine prior ICD-10 diagnosis classification tasks. Performance on the val set when using 10% of the train set visits. Training visit are sampled with replacement and both downstream training and hyperparameter selection are repeated five times, results are mean

\pm

std across repetitions. Regression tasks are evaluated using Pearson correlation (

\uparrow

), and classification tasks using AUROC (

\uparrow

). Bold red marks the best mean value in each row.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.633 $\pm$ 0.035	0.668 $\pm$ 0.006	0.673 $\pm$ 0.006	0.649 $\pm$ 0.003	0.549 $\pm$ 0.049	0.636 $\pm$ 0.023	0.726 $\pm$ 0.003	0.763 $\pm$ 0.005
Sex cls	0.716 $\pm$ 0.004	0.715 $\pm$ 0.005	0.698 $\pm$ 0.004	0.682 $\pm$ 0.006	0.673 $\pm$ 0.013	0.704 $\pm$ 0.008	0.734 $\pm$ 0.004	0.744 $\pm$ 0.006
ED_dispo cls	0.676 $\pm$ 0.007	0.667 $\pm$ 0.006	0.657 $\pm$ 0.005	0.632 $\pm$ 0.007	0.635 $\pm$ 0.008	0.658 $\pm$ 0.005	0.672 $\pm$ 0.007	0.712 $\pm$ 0.005
Potassium reg	0.064 $\pm$ 0.023	0.124 $\pm$ 0.026	0.122 $\pm$ 0.026	0.129 $\pm$ 0.020	0.048 $\pm$ 0.028	0.103 $\pm$ 0.025	0.112 $\pm$ 0.020	0.127 $\pm$ 0.017
Calcium reg	0.076 $\pm$ 0.020	0.085 $\pm$ 0.027	0.076 $\pm$ 0.028	0.072 $\pm$ 0.006	0.030 $\pm$ 0.010	0.075 $\pm$ 0.014	0.118 $\pm$ 0.009	0.160 $\pm$ 0.013
eGFR reg	0.207 $\pm$ 0.029	0.374 $\pm$ 0.008	0.395 $\pm$ 0.019	0.401 $\pm$ 0.004	0.214 $\pm$ 0.097	0.318 $\pm$ 0.082	0.425 $\pm$ 0.018	0.510 $\pm$ 0.015

Appendix

Table S14 : Model comparison of PPG-only val performance on all 20 benchmark tasks when using 10% of the train visits . Identical format and content as Table S13 , but for PPG-only . Performance on the val set when using 10% of the train set visits.

Task	Domain Features	MOMENT-base	Chronos-Bolt-small	PaPaGei	D-BETA	ECGFounder	xECG-10min	CSFM-base
Age reg	0.746 $\pm$ 0.022	0.747 $\pm$ 0.003	0.742 $\pm$ 0.005	0.650 $\pm$ 0.010	0.528 $\pm$ 0.198	0.793 $\pm$ 0.004	0.801 $\pm$ 0.002	0.823 $\pm$ 0.003
Sex cls	0.746 $\pm$ 0.003	0.772 $\pm$ 0.006	0.755 $\pm$ 0.005	0.714 $\pm$ 0.009	0.736 $\pm$ 0.007	0.782 $\pm$ 0.004	0.813 $\pm$ 0.004	0.814 $\pm$ 0.004
ED_dispo cls	0.712 $\pm$ 0.004	0.696 $\pm$ 0.007	0.691 $\pm$ 0.009	0.652 $\pm$ 0.006	0.686 $\pm$ 0.015	0.711 $\pm$ 0.013	0.718 $\pm$ 0.001	0.728 $\pm$ 0.006
Potassium reg	0.124 $\pm$ 0.053	0.260 $\pm$ 0.026	0.226 $\pm$ 0.019	0.166 $\pm$ 0.021	0.126 $\pm$ 0.067	0.199 $\pm$ 0.037	0.316 $\pm$ 0.027	0.323 $\pm$ 0.017
Calcium reg	0.113 $\pm$ 0.030	0.115 $\pm$ 0.012	0.121 $\pm$ 0.010	0.075 $\pm$ 0.011	0.087 $\pm$ 0.009	0.124 $\pm$ 0.014	0.180 $\pm$ 0.011	0.210 $\pm$ 0.007
eGFR reg	0.334 $\pm$ 0.072	0.469 $\pm$ 0.011	0.479 $\pm$ 0.008	0.381 $\pm$ 0.019	0.422 $\pm$ 0.031	0.514 $\pm$ 0.037	0.547 $\pm$ 0.006	0.572 $\pm$ 0.010

Appendix

Table S15 : Model comparison of ECG + PPG val performance on all 20 benchmark tasks when using 10% of the train visits . Identical format and content as Table S13 , but for ECG + PPG . Performance on the val set when using 10% of the train set visits.

Task	Train	Val	Test
Potassium reg (“POTASSIUM”)	13,481	1,700	1,683
Calcium reg (“CALCIUM”)	13,464	1,700	1,677
eGFR reg (“EGFR REFIT WITHOUT RACE (2021)”)	13,407	1,694	1,675
Glucose reg (“GLUCOSE”)	13,100	1,658	1,644
Hemoglobin reg (“HEMOGLOBIN (HGB)”)	13,854	1,772	1,724
Albumin reg (“ALBUMIN”)	13,538	1,729	1,707

Appendix

Table S16 : Number of train , val , and test visits for each of the eight laboratory regression tasks . To ensure unambiguous supervision, only visits with exactly one recorded measurement of the corresponding laboratory test are included.

Task	Label / Description	Train	Val	Test
Sex cls	“M” vs “F”	52.2%	52.4%	52.5%
ED_dispo cls	“Discharge” vs {“Inpatient”, “Observation”, “ICU”}	50.5%	51.0%	51.2%
I48 cls	Atrial Fibrillation Detection	10.5%	8.5%	7.6%
I50 cls	Heart Failure Detection	9.8%	7.2%	6.5%
{Z95.0, Z95.2, Z95.3} cls	Cardiac Device Detection	3.6%	2.6%	2.4%
N18 cls	Chronic Kidney Disease Detection	10.5%	7.0%	7.4%

Appendix

Table S17 : Positive class prevalence for each of the eleven binary classification tasks . Percentage of visits labeled as the positive class in the train , val , and test splits. Values are computed at the visit level and are provided to contextualize reported AUROC performance.

Explore similar work

Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model
May 10, 2026Zhangdaihong Liu, Chang Liu, Fenglin Liu +4Cross-Modal Representation Learning Masked Autoencoders
Department of Engineering Science, University of Oxford, Oxford OX3 7DQ, UK · Oxford Suzhou Centre for Advanced Research, Suzhou, 215123, China · School of Public Health, Shanghai Jiao Tong University, Shanghai, China
MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning
Oct 6, 2026Robert A. Lewis, I-Min Chiu, Kyle Verrier +7Electrocardiogram Foundation Models Electrocardiogram Classification
Massachusetts Institute of Technology (work done while at Apple). · Apple, Inc. · Division of Cardiology, Johns Hopkins Medicine. +1
Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning
May 1, 2026Hao Zhou, Simon A. Lee, Cyrus Tanade +12Representation Learning Cross-Modal Representation Learning
Samsung Research America · Algorithm Team, Digital Health Lab, Samsung Research America · The Pennsylvania State University +1

May 10, 2026J/K move · Enter open · S save

Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model

Zhangdaihong Liu, Chang Liu, Fenglin Liu, Yixuan Chen+3

Department of Engineering Science, University of Oxford, Oxford OX3 7DQ, UK · Oxford Suzhou Centre for Advanced Research, Suzhou, 215123, China · School of Public Health, Shanghai Jiao Tong University, Shanghai, China

Cross-Modal Representation Learning Masked Autoencoders Electrocardiography Photoplethysmography Wearable Health Monitoring Time-Series Foundation Models

Open paper AskPDF arXiv

Cardiovascular disease remains the leading cause of global mortality, yet scalable cardiac monitoring is hindered by the gap between diagnostic-rich ECG and ubiquitous wearable PPG. Bridging this gap requires representations that are compact, transferable across modalities and devices, and deployable without task-specific retraining. Here we introduce biosignal fingerprints: compact latent representations of cardiovascular state derived from a cross-modal foundation model, the Multi-modal Masked Autoencoder (M2AE), trained on over 3.4 million paired ECG and PPG signals. M2AE integrates modality-specific encoders with a shared bottleneck and dual decoders, jointly optimized using reconstruction and cross-modal contrastive objectives, yielding generalizable fingerprints that retain intra- and inter-modality features. Like a biometric fingerprint, these representations uniquely encode an individual's cardiovascular state in a modality-agnostic, privacy-preserving form reusable across clinical tasks without exposing raw waveform data or requiring model retraining. Across 7 downstream tasks, spanning cross-modal reconstruction, cardiovascular disease classification, hypertension detection, mortality prediction, and demographic inference, biosignal fingerprints achieve competitive or superior performance compared to leading domain-specialist foundation models in frozen settings, including an AUROC of 0.974 for five-class CVD classification and 0.877 for hypertension detection, with a maximum improvement of 27.7% in AUROC across 5 classification tasks. Critically, strong performance is maintained with only a single modality, enabling deployment in resource-constrained, single-sensor environments typical of real-world wearable monitoring, with direct implications for continuous cardiovascular monitoring across clinical and consumer health settings.

Abstract

Explore similar work

Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model

MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning

Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning