PPG-Bench: A Unified Benchmark for PPG Foundation Models

Arvasu Kulkarni; Mahir Jain; Aditya Ray Mishra; Sandeep Singh; Siddharth Panwar

cs.LGOct 9, 2026

PPG-Bench: A Unified Benchmark for PPG Foundation Models

Authors: Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra, Sandeep Singh, Siddharth Panwar

Organizations: Mannas AI

Abstract

Photoplethysmography (PPG) is increasingly recorded by wearable and clinical devices at enormous scale, and a growing number of PPG foundation models (FMs) have been developed to compute reusable representations for various downstream tasks. These models are difficult to compare: different evaluations use different datasets, task definitions, preprocessing, downstream heads, hyperparameter tuning, and baselines. We present PPG-Bench, a unifying open-source benchmark evaluating five public PPG FMs across nine open-access datasets and 26 tasks spanning cardiovascular state estimation, activity, sleep, screening, and identification. We evaluate models under four downstream regimes of increasing adaptability (linear probing, random forests, frozen MLP probes, and end-to-end finetuning), with model-specific hyperparameter optimization and subject-disjoint cross-validation. We also evaluate against three controls: random/constant baseline, a deterministic pulse-rate-variability (PRV) feature set, and a generic time-series FM (MOMENT). Across all 68 task-regime combinations, 35.3% show no meaningful FM advantage over the random/constant baseline, and in 83.8% no PPG-specific FM outperforms both PRV and MOMENT. We also find that added downstream head capacity generally improves performance, but can cause significant degradation if the model is performing at chance/constant level. Current PPG FMs therefore cannot yet claim to provide generalised and transferrable representations of PPG signals, and future evaluation in this field should include checks against simple, non-learned controls. Code for the benchmark and instructions to evaluate more models can be found here - https://github.com/mannas-ai/ppg-bench.

Figures & tables

Metric	Dataset	Task	AnyPPG	PaPaGei-S	PaPaGei-P	PulsePPG	SigmaPPG	Random	PRV Ctrl	MOMENT
MAE	Dalia	HR	7.39 $\pm$ .18	13.75 $\pm$ .32	11.68 $\pm$ .31	10.00 $\pm$ .27	11.28 $\pm$ .30	18.23 $\pm$ .42	11.19 $\pm$ .34	8.24 $\pm$ .19
	WildPPG	HR	9.76 $\pm$ .14	12.63 $\pm$ .28	11.96 $\pm$ .21	10.83 $\pm$ .19	11.26 $\pm$ .19	12.99 $\pm$ .19	13.00 $\pm$ .18	11.54 $\pm$ .22
	PPG-BP	HR	3.31 $\pm$ .06	4.89 $\pm$ .15	5.22 $\pm$ .16	3.67 $\pm$ .15	4.15 $\pm$ .10	8.63 $\pm$ .02	4.48 $\pm$ .12	3.21 $\pm$ .08
		SBP	13.32 $\pm$ .27	14.92 $\pm$ .26	14.60 $\pm$ .37	13.77 $\pm$ .17	14.95 $\pm$ .29	16.31 $\pm$ .04	16.40 $\pm$ .09	13.89 $\pm$ .33
		DBP	8.21 $\pm$ .36	8.70 $\pm$ .06	8.55 $\pm$ .13	8.49 $\pm$ .18	8.63 $\pm$ .10	8.79 $\pm$ .03	8.77 $\pm$ .03	8.54 $\pm$ .27
		Age	10.76 $\pm$ .24	11.37 $\pm$ .21	11.54 $\pm$ .36	10.31 $\pm$ .06	11.59 $\pm$ .14	12.43 $\pm$ .06	12.24 $\pm$ .06	10.82 $\pm$ .13

Table 1: Benchmark metrics for the 17 primary tasks evaluated under linear probing. Values are mean

\pm

Student-t 95% CI. Lower is better for MAE; higher is better for BAC. Bold entries are the best mean and models whose CI overlaps the best.

Method Transition	Improved %	Mean % Gain
Linear Probe $\rightarrow$ Random Forest	17.1	-6.7
Linear Probe $\rightarrow$ Frozen MLP	58.5	5.9
Linear Probe $\rightarrow$ Finetuning	78.0	9.3
Frozen MLP $\rightarrow$ Finetuning	70.7	4.5

Table 2: Changes between evaluation methods. Improved % is the fraction of model–task pairs that improve; Mean % Gain is their average performance change.

Figure 1: Performance change relative to linear probing for tasks where FMs outperform PRV and the Random Chance / Constant Baseline. Each point is averaged over FMs with above-control performance on that task.

Probe method	Mean	Median	Top- $k$ mean
Linear probe	31.8%	27.5%	40.7%
Random forest	34.2%	36.2%	29.6%
Frozen MLP	23.5%	34.3%	42.2%
Fine-tuning	32.5%	35.2%	32.3%

Table 3: Frequency with which each subject-level aggregation method was selected for non-regression tasks in the final benchmark.

Method	Task	Metric	Subject-level	Window-level
Linear Probes	PPG-BP Age	MAE	11.117	11.714
	PPG-BP BMI		3.006	3.074
	PPG-BP DBP		8.515	8.781
	PPG-BP HR		4.249	5.250
	PPG-BP SBP		14.312	15.315
	PPG-BP Diabetes	BAC	0.509	0.509

Table 4: Subject-level versus window-level evaluation for eligible subject-level tasks. Results are averaged across the PPG FMs. Lower is better for MAE; higher is better for BAC. Bold indicates the better score in each row.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Hyperparameter	Values
Linear (Ridge)	$\alpha$	$\{0.01,0.1,1,10,100,1000,10000\}$
Linear (Ridge)	Solver	auto
Linear (Logistic)	$C$	$\{0.001,0.01,0.1,1,10\}$
	Penalty	L2
	Solver	lbfgs
	Max iterations	1000

Appendix

Table 5: Hyperparameter search spaces used in PPG-Bench. Parameters with a single value are fixed rather than searched.

Method	Hyperparameter	Search range	Lower edge	Upper edge
Linear	Logistic $C$	$[10^{-3},10]$	14.2%	26.9%
Linear	Ridge $\alpha$	$[10^{-2},10^{4}]$	5.8%	50.3%
Random Forest	Max depth	$[5,20]$	43.8%	37.9%
Frozen MLP	Batch size	$[16,128]$	47.0%	13.4%
	Learning rate	$[10^{-5},10^{-2}]$	6.0%	48.6%
	Selected epoch	$[1,25]$	10.2%	4.6%

Appendix

Table 6: Post-hoc edge pressure in the final hyperparameter searches. Lower- and upper-edge percentages indicate how often the selected value lies at the minimum or maximum available setting. Values above 50% are bolded.

Model	Trainable Params
PRV Control	0
Random	0
AnyPPG	4.0M
PaPaGei-P	5.0M
PaPaGei-S	5.8M
PulsePPG	28.5M

Appendix

Table 7: Approximate trainable parameter counts for models used in the benchmark.

Dataset	Primary Tasks	Secondary Tasks	Subjects	Recording duration
Dalia ( Reiss et al., 2019 )	HR; activity	–	15	36.0 h
DREAMT ( Wang et al., 2024 )	Sleep staging	–	100	$\sim$ 700 h
WESAD ( Schmidt et al., 2018 )	Affective state	Valence; Arousal	15	$\sim$ 30 h
ECSMP ( Gao et al., 2021 )	Depression classification	Emotion classification	67	210.5 h
PPG-BP ( Liang et al., 2018 )	Age; BMI; systolic BP; diastolic BP; HR; diabetes; hypertension	Hypertension (binary); cerebral infarction; cerebrovascular disease	219	0.383 h
BIDMC ( Pimentel et al., 2017 )	Respiratory rate; SpO 2	–	53	7.07 h

Appendix

Table 8: Datasets and prediction tasks in the PPG benchmark.

Metric	Dataset	Task	AnyPPG	PaPaGei-S	PaPaGei-P	PulsePPG	SigmaPPG	Random	PRV Ctrl	MOMENT
MAE	Dalia	HR	5.84 $\pm$ .18	14.96 $\pm$ .42	13.12 $\pm$ .42	10.17 $\pm$ .37	13.24 $\pm$ .42	18.23 $\pm$ .42	11.75 $\pm$ .30	8.04 $\pm$ .28
	WildPPG	HR	9.29 $\pm$ .10	12.55 $\pm$ .26	12.45 $\pm$ .23	10.99 $\pm$ .22	12.16 $\pm$ .21	12.99 $\pm$ .19	12.78 $\pm$ .19	12.06 $\pm$ .23
	PPG-BP	HR	3.55 $\pm$ .06	5.92 $\pm$ .13	6.10 $\pm$ .09	5.03 $\pm$ .06	5.13 $\pm$ .05	8.63 $\pm$ .02	4.75 $\pm$ .17	4.20 $\pm$ .06
		SBP	14.60 $\pm$ .16	15.53 $\pm$ .20	15.28 $\pm$ .10	14.87 $\pm$ .16	16.18 $\pm$ .26	16.31 $\pm$ .04	16.36 $\pm$ .17	14.60 $\pm$ .21
		DBP	8.47 $\pm$ .19	8.69 $\pm$ .09	8.63 $\pm$ .14	8.63 $\pm$ .08	8.72 $\pm$ .09	8.79 $\pm$ .03	8.73 $\pm$ .07	8.53 $\pm$ .17
		Age	11.05 $\pm$ .16	11.92 $\pm$ .13	11.83 $\pm$ .07	10.96 $\pm$ .11	12.26 $\pm$ .06	12.43 $\pm$ .06	12.38 $\pm$ .08	10.94 $\pm$ .13

Appendix

Table 9: Benchmark metrics for the 17 primary tasks evaluated with random forests, frozen MLP probes, and finetuning. Values are mean

\pm

Student-t 95% CI. Lower is better for MAE; higher is better for BAC. Bold entries are the best mean and models whose CI overlaps the best.

Metric	Dataset	Task	AnyPPG	PaPaGei-S	PaPaGei-P	PulsePPG	SigmaPPG	Random	PRV Ctrl	MOMENT
MAE	Dalia	HR	6.12 $\pm$ .30	13.74 $\pm$ .68	12.30 $\pm$ 1.22	9.46 $\pm$ 1.11	10.39 $\pm$ .79	18.23 $\pm$ .42	10.35 $\pm$ .80	7.49 $\pm$ .05
	WildPPG	HR	9.89 $\pm$ .61	12.80 $\pm$ 1.01	12.56 $\pm$ 1.44	10.91 $\pm$ .65	11.43 $\pm$ .86	12.99 $\pm$ .19	13.22 $\pm$ 1.09	11.75 $\pm$ 1.06
	PPG-BP	HR	3.96 $\pm$ .11	5.20 $\pm$ .39	5.68 $\pm$ .16	4.30 $\pm$ .36	4.72 $\pm$ .20	8.63 $\pm$ .02	3.94 $\pm$ .17	4.15 $\pm$ .63
		SBP	14.76 $\pm$ 1.16	15.68 $\pm$ .94	16.23 $\pm$ .62	15.01 $\pm$ .54	16.34 $\pm$ .22	16.31 $\pm$ .04	17.24 $\pm$ 1.23	15.51 $\pm$ .87
		DBP	9.08 $\pm$ .30	9.31 $\pm$ .44	9.22 $\pm$ .18	9.12 $\pm$ .67	9.21 $\pm$ .38	8.79 $\pm$ .03	9.13 $\pm$ .14	9.41 $\pm$ .55
		Age	11.22 $\pm$ .18	11.87 $\pm$ .23	12.07 $\pm$ .16	11.03 $\pm$ .72	12.57 $\pm$ .28	12.43 $\pm$ .06	12.59 $\pm$ .67	11.35 $\pm$ .64

Appendix

Table 10: Continued.

Metric	Dataset	Task	AnyPPG	PaPaGei-S	PaPaGei-P	PulsePPG	SigmaPPG	Random	PRV Ctrl	MOMENT
MAE	Dalia	HR	5.07 $\pm$ .29	6.21 $\pm$ .40	6.51 $\pm$ .18	5.10 $\pm$ .22	6.31 $\pm$ .35	18.23 $\pm$ .42	10.35 $\pm$ .80	5.52 $\pm$ .37
	WildPPG	HR	8.40 $\pm$ .79	9.27 $\pm$ .81	10.27 $\pm$ .86	8.51 $\pm$ .46	9.16 $\pm$ .55	12.99 $\pm$ .19	13.22 $\pm$ 1.09	9.39 $\pm$ 1.78
	PPG-BP	HR	8.71 $\pm$ 5.41	9.91 $\pm$ 4.74	9.82 $\pm$ 4.80	14.41 $\pm$ 9.71	8.75 $\pm$ .12	8.63 $\pm$ .02	3.94 $\pm$ .17	15.88 $\pm$ 5.07
		SBP	20.89 $\pm$ 10.13	16.60 $\pm$ .30	22.83 $\pm$ 17.33	18.87 $\pm$ 10.10	16.63 $\pm$ .77	16.31 $\pm$ .04	17.24 $\pm$ 1.23	28.36 $\pm$ 26.14
		DBP	11.43 $\pm$ 5.13	8.77 $\pm$ .20	12.62 $\pm$ 7.20	8.76 $\pm$ .16	11.16 $\pm$ 10.19	8.79 $\pm$ .03	9.13 $\pm$ .14	8.77 $\pm$ .25
		Age	14.01 $\pm$ 3.24	12.68 $\pm$ .69	14.17 $\pm$ 2.45	13.69 $\pm$ 2.83	14.00 $\pm$ 2.90	12.43 $\pm$ .06	12.59 $\pm$ .67	14.70 $\pm$ 5.13

Appendix

Table 11: Continued.

Metric	Dataset	Task	AnyPPG	PaPaGei-S	PaPaGei-P	PulsePPG	SigmaPPG	Random	PRV Ctrl	MOMENT
MAE	BUTPPG	SBP	13.54 $\pm$ .59	12.82 $\pm$ .42	13.06 $\pm$ .47	13.23 $\pm$ .47	12.95 $\pm$ .46	12.72 $\pm$ .48	12.87 $\pm$ .59	13.15 $\pm$ .52
		DBP	6.74 $\pm$ .70	6.34 $\pm$ .48	6.54 $\pm$ .55	6.65 $\pm$ .60	6.46 $\pm$ .57	6.30 $\pm$ .47	6.42 $\pm$ .51	6.49 $\pm$ .56
		SpO 2	1.30 $\pm$ .06	1.30 $\pm$ .05	1.29 $\pm$ .04	1.28 $\pm$ .05	1.29 $\pm$ .06	1.29 $\pm$ .06	1.30 $\pm$ .03	1.29 $\pm$ .03
	WESAD	Valence	0.92 $\pm$ .02	1.12 $\pm$ .02	1.10 $\pm$ .05	1.00 $\pm$ .04	1.08 $\pm$ .03	1.17 $\pm$ .02	1.16 $\pm$ .03	1.00 $\pm$ .05
	WESAD	Arousal	1.29 $\pm$ .03	1.61 $\pm$ .02	1.67 $\pm$ .02	1.46 $\pm$ .03	1.53 $\pm$ .03	1.78 $\pm$ .02	1.78 $\pm$ .01	1.36 $\pm$ .05
BAC	PPG-BP	Hypertension (Binary)	.691 $\pm$ .032	.620 $\pm$ .030	.633 $\pm$ .011	.674 $\pm$ .024	.609 $\pm$ .016	.500 $\pm$ .000	.499 $\pm$ .002	.653 $\pm$ .031

Appendix

Table 12: Metrics for tasks evaluated with linear probes and random forest only. Values are mean

\pm

Student-t 95% CI. Lower is better for MAE; higher is better for BAC. Bold entries are the best mean and models whose CI overlaps the best.

Explore similar work

Jun 5, 2026

A robust PPG foundation model using multimodal physiological supervision

Photoplethysmography (PPG), a non-invasive measure of changes in blood volume, is widely used in both wearable devices and clinical settings. Recent PPG foundation models either use open-source ICU datasets with pretraining paradigms that require curated data and thus complicate generalization to field-like data, or use closed-source field-like PPG data. In contrast, we propose a PPG foundation model that does not require high-quality or field-like pretraining data, and instead leverages accompanying electrocardiogram and respiratory signals in ICU datasets to select contrastive samples during pretraining. Our approach allows the model to retain and learn from noisy PPG segments, improving robustness at inference. Our model, pretrained on 3x fewer subjects than existing state-of-the-art approaches, achieves performance improvements on 14 out of 15 diverse downstream tasks, including field-like daily activity and heart rate prediction. Our results demonstrate that multimodal supervision can integrate complementary physiological information to improve the robustness of PPG foundation models and enhance their generalization to consumer-grade data.

Eloy Geenjaar, Vince Calhoun, Scott Daly +4

Department of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, USA · Tri-Institutional Center for Translational Research in Neuroimaging and Data Science (TReNDS), Georgia State University, Georgia Institute of Technology, Emory University, Atlanta, USA · Dolby Laboratories, San Francisco, USA

Multimodal Robustness Multimodal Pretraining Photoplethysmography Multimodal Contrastive Learning

Ask PDF

Mar 23, 2026

Deriving Health Metrics from the Photoplethysmogram: Benchmarks and Insights from MIMIC-III-Ext-PPG

Photoplethysmography (PPG) is one of the most widely captured biosignals for clinical prediction tasks, yet PPG-based algorithms are typically trained on small-scale datasets of uncertain quality, which hinders meaningful algorithm comparisons. We present a comprehensive benchmark for PPG-based clinical prediction using the \dbname~dataset, establishing baselines across the full spectrum of clinically relevant applications: multi-class heart rhythm classification, and regression of physiological parameters including respiratory rate (RR), heart rate (HR), and blood pressure (BP). Most notably, we provide the first comprehensive assessment of PPG for general arrhythmia detection beyond atrial fibrillation (AF) and atrial flutter (AFLT), with performance stratified by BP, HR, and demographic subgroups. Using established deep learning architectures, we achieved strong performance for AF detection (AUROC = 0.96) and accurate physiological parameter estimation (RR MAE: 2.97 bpm; HR MAE: 1.13 bpm; SBP/DBP MAE: 16.13/8.70 mmHg). Cross-dataset validation demonstrates excellent generalizability for AF detection (AUROC = 0.97), while clinical subgroup analysis reveals marked performance differences across subgroups by BP, HR, and demographic strata. These variations appear to reflect population-specific waveform differences rather than systematic bias in model behavior. This framework establishes the first integrated benchmark for multi-task PPG-based clinical prediction, demonstrating that PPG signals can effectively support multiple simultaneous monitoring tasks and providing essential baselines for future algorithm development.

Mohammad Moulaeifard, Philip J. Aston, Peter H. Charlton +1

AI4Health Department, Oldenburg University, Oldenburg, Germany · Department of Data Science and AI, National Physical Laboratory, Teddington, United Kingdom · School of Mathematics and Physics, University of Surrey, Guildford, United Kingdom +1

Clinical Prediction Photoplethysmography Heart Rate Estimation Cuffless Blood Pressure Estimation

Ask PDF

Sep 27, 2026

PPG-LM: A Photoplethysmography-Language Model with Multi-Level Clinical Alignment

Photoplethysmography (PPG) is widely recorded by clinical monitors and consumer wearables, providing a scalable source of continuous physiological information. These recordings offer an opportunity for physiological assessment at scale, but realizing this potential requires models to learn from both signal-derived physiological supervision and broader clinical context captured in electronic health records (EHRs). This involves aligning information spanning local observations, care events, and entire visits with PPG representations at corresponding temporal scales. However, existing PPG foundation models primarily rely on task-specific prediction heads, while the medical knowledge of large language models does not necessarily translate into waveform understanding. To bridge this gap, we introduce PPG-LM, the first PPG-language model family to learn physiological representations from both signal-derived supervision and broader clinical context captured in EHRs. To construct clinically grounded captions, we develop an automatic captioning pipeline that generates segment-, event-, and visit-level descriptions from signal measurements and structured EHR records. We then learn from these pairs through a two-stage framework that first establishes segment-language correspondence through contrastive learning and waveform-conditioned captioning, then extends alignment to events and visits through time-aware aggregation and temporal statement matching. Pretrained on approximately 73k hours of PPG, PPG-LM supports language-based recognition, cross-modal retrieval, and segment captioning. Experiments on MC-MED, MIMIC-III, and VitalDB show improved retrieval and caption factuality over language-model baselines and gains over PPG and time-series foundation models on multiple clinical prediction tasks.

Xiaoda Wang, Minxiao Wang, Maxwell A Xu +10

Emory University · University of California, Los Angeles · Google +3

Multimodal Pretraining Photoplethysmography

Ask PDF