RAE-PPG: Duration-Grounded Retain-and-Extend Pretraining for PPG Foundation Models

Suyeong Lee; Hochang Lee; Seokyong Sheem; Daekyum Kim

cs.LGSep 29, 2026

RAE-PPG: Duration-Grounded Retain-and-Extend Pretraining for PPG Foundation Models

Authors: Suyeong Lee, Hochang Lee, Seokyong Sheem, Daekyum Kim

Organizations: Korea University

Abstract

Signal features derived from photoplethysmography (PPG) require different signal durations to characterize. Existing PPG foundation models treat duration as a pretraining or evaluation condition rather than using the different durations required by PPG features to organize self-supervision. We hypothesize that self-supervision should expand with signal duration, allowing a single encoder to progressively acquire additional features while preserving and reusing earlier learning. We introduce Retain-and-Extend PPG (RAE-PPG), which trains a single Transformer encoder successively on 10 s, 30 s, and 240 s inputs, adding supervision for signal features supported by each longer observation. The encoder is partitioned into duration-specific parameter groups, allowing later stages to reuse earlier groups while updating only the group assigned to the current stage. Selected earlier targets are reused to supervise later stages, encouraging the corresponding features to remain accessible in longer-input representations. Direct decoding from the final encoder shows that earlier features remain recoverable from longer-input representations, while later-stage features show higher mean decoding performance at their introduction durations. Controlled comparisons further show that prior-stage learning provides a better basis for learning newly introduced features at both transitions. Across 18 tasks from eight datasets, the final frozen encoder achieves the best observed score on 12 tasks compared with five existing PPG foundation models.

Figures & tables

Figure 1: RAE-PPG overview: duration-grounded targets, retain-and-extend pretraining, and frozen-encoder downstream evaluation.

Feature family	Signal-processing role	Realization in RAE-PPG
Rate	Pulse cadence	Local pulse rate at 10 s; full-window mean rate at 240 s $\dagger$
Morphology / vascular timing	Within-beat contour, fiducial timing, and reflected-wave structure	Local morphology at 10 s; reflected-wave timing at 30 s
Beat-interval variability	Successive interval changes and dispersion	PRV-RMSSD at 30 s; PRV-pNN50 and interval dispersion at 240 s
Waveform statistics	Baseline-sensitive autocorrelation and low-frequency modulation	Uncentered ACF index and low-frequency waveform variability at 240 s
Signal quality $\dagger$	Reliability of periodic pulse structure	Spectral purity at 10 s

Table 1: Physiology-relevant signal-feature families used by RAE-PPG.

Figure 2: RAE-PPG architecture and pretraining. Later stages reuse frozen groups and learn new targets while retaining selected earlier targets.

Downstream task	PaPaGei-P	PaPaGei-S	Pulse-PPG	SIGMA-PPG	AnyPPG	RAE-PPG
A. Regression: MAE $\downarrow$
DaLiA: HR 10	13.9397 [12.25–15.86]	16.2506 [13.25–19.37]	10.4389 [8.34–12.71]	14.8701 [12.28–18.98]	8.1192 [5.54–11.30]	8.6522 [6.31–11.18]
BUT PPG: HR 10	10.9192 [8.05–14.40]	10.8344 [8.06–14.22]	10.5904 [8.00–13.76]	10.6122 [7.91–13.87]	9.6609 [7.24–12.62]	10.5144 [8.21–13.37]
PTT-PPG: SBP 10	9.9489 [5.53–15.25]	10.4675 [5.81–15.99]	9.8620 [5.36–15.32]	10.3279 [5.64–15.98]	10.1314 [5.63–15.52]	9.5291 [5.24–14.87]
PTT-PPG: DBP 10	5.6742 [3.94–7.48]	5.9042 [4.02–7.90]	5.7877 [4.10–7.55]	6.0150 [4.19–7.91]	5.8514 [4.03–7.76]	5.4489 [3.78–7.21]
DREAMT: Age 10	10.8791 [8.66–13.41]	11.9138 [9.55–14.51]	11.0811 [8.63–13.91]	12.1837 [9.58–14.75]	10.9798 [8.37–13.81]	11.8329 [9.13–14.75]

Table 2: Frozen-encoder downstream comparison (95% CIs in brackets).

A. Retention / extension	Gain or $\Delta$ gain [95% CI]
retention $10\to 30$ s ( $n=9$ )	0.830 [0.815, 0.846]
extension $10\to 30$ s ( $n=2$ )	$0.611\rightarrow 0.898$ ; $+0.287$ [0.267, 0.307]
retention $30\to 240$ s ( $n=2$ )	0.666 [0.645, 0.687]
extension $30\to 240$ s ( $n=5$ )	$0.579\rightarrow 0.755$ ; $+0.176$ [0.163, 0.189]

Table 3: Signal-feature retention, extension, and downstream prediction.

	Nested/protected		Fully shared
Evaluation	A: ON	D: OFF	B: ON	C: OFF
Mean rank (18 tasks) $\downarrow$	2.03	2.58	2.56	2.83
Earlier features: aligned readout, gain $\uparrow$
$10\to 30$ s ( $n=9$ )	0.830	0.710	0.663	0.635
$30\to 240$ s ( $n=2$ )	0.666	0.534	0.747	0.530
New features: native-duration mean pooling, gain $\uparrow$

Table 4: Ablations of parameter organization and retained-target supervision.

Appendix figures & tables34 assets

Supplementary material from the paper’s appendix.

Appendix

Source	Partition	Groups: before $\to$ after	Records	Blocks	Source hours
MIMIC-III	TRAIN	5,614 $\to$ 5,544	174,858	3,101,986	421,887
MIMIC-III	VAL	729 $\to$ 724	21,560	465,577	66,397
MIMIC-III	TEST	737 $\to$ 736	22,447	413,427	56,047
VitalDB	TRAIN	4,930 $\to$ 4,609	4,609	7,216	13,841
VitalDB	VAL	604 $\to$ 566	566	898	1,698
VitalDB	TEST	622 $\to$ 582	582	867	1,774

Appendix

Table 5: Pretraining corpus after block-level screening. Counts before and after screening are shown for participants (MIMIC-III) or cases (VitalDB). Records are source segment files for MIMIC-III and case recordings for VitalDB. Hours are interval unions within each source record, before crop selection; they are not deduplicated across separate records.

Module	Operations and output shape	Parameters	Writable stage
Local feature extractor	Linear and local two-convolution branch; $B\times N\times 128$	50,560	10 s
Token projection	Content + position + duration embedding, LN; $B\times 64\times 384$	64,128	10 s; duration embeddings at their corresponding stages
Transformer	Six attention/FFN blocks; $B\times 64\times 384$	10,646,784	Current attention/FFN group; shared norms/biases at 10 s
Final LN	$B\times 64\times 384$	768	10 s
Feature heads	Eight affine maps $384\to d_{h}$ ; $B\times Q_{h}\times d_{h}$	7,315	Selected rows at their introduction stage
Exported encoder	Signal tokens; downstream mean is $B\times 384$	10,762,240	Frozen for evaluation

Appendix

Table 6: Architecture and parameter ownership.

B

is batch size,

N

the input length in samples, and

Q_{h}

and

d_{h}

the query count and output width of head

h

. Counts include biases and normalization parameters.

Feature (unit)	Support	Definition	Source
Introduced at 10 s
Peak-interval rate (bpm)	Peak pair	$60/I_{i}$ ; $0.3\leq I_{i}\leq 2$ s	Adapted from Goda et al. (2024)
Slope transit time (s)	Pulse	$[x(s_{i})-x(o_{i})]/\max_{[o_{i},s_{i}]}\dot{x}$	Natarajan et al. (2022)
APG $b/a$	Pulse	Signed amplitude ratio $b_{i}/a_{i}$	Takazawa et al. (1998)
APG $d/a$	Pulse	Signed amplitude ratio $d_{i}/a_{i}$	Takazawa et al. (1998)
Onset-to-APG- $b$ (s)	Pulse	$q_{i}^{b}-o_{i}$	Adapted from Natarajan et al. (2022)

Appendix

Table 7: The sixteen pretraining targets. Pulse denotes

[o_{i},o_{i+1})

; ratios are dimensionless unless a unit is shown.

P(B)

sums spectral power in band

B

. Source citations identify the underlying signal-feature definitions; the extraction settings and support requirements are specified in this section. Adapted from indicates a changed definition or extraction procedure. PRV denotes variability of PPG-derived pulse intervals.

Dataset	Final groups	ECG reference median [Q1–Q3]	Bias (PPG $-$ ECG)	Median / P90 absolute difference
MIMIC-III	50	$23.53\;[12.13,37.40]$ ms	$+3.23$ ms	$2.39$ / $11.91$ ms
VitalDB	19	$25.99\;[18.73,44.14]$ ms	$+2.84$ ms	$1.21$ / $14.74$ ms

Appendix

Table 8: Matched-event PPG–ECG comparison of pulse-interval SD in 240 s windows. Bias is the mean PPG-minus-ECG difference; brackets show ECG-reference quartiles. P90 denotes the 90th percentile of absolute differences.

Feature	Rank-check window (s)	Lower 95% bound (analysis / replication)	Held-out availability (%)
Introduced at 10 s
Peak-interval rate	10	0.932 / 0.919	97.8
Slope transit time	10	0.935 / 0.929	97.5
APG $b/a$	10	0.878 / 0.873	96.4
APG $d/a$	10	0.848 / 0.849	88.4
Onset-to-APG- $b$	10	0.794 / 0.782	96.4

Appendix

Table 9: Adjacent-window rank consistency and held-out availability of the final targets. Rank-check windows summarize pulse and pulse-pair values by their median and reflected-delay values by their mean; other targets are computed over the indicated window. Each bound is the minimum of the overall and per-dataset lower 95% bounds of the Spearman–Brown score, shown for the analysis and replication panels at the first tested window length. Held-out availability is the percentage of 87,362 held-out trajectories with an available final-target summary. Dashes indicate auxiliary targets excluded from the rank-consistency analysis.

Setting	10 s	30 s	240 s
Initialization	Random	10 s checkpoint	30 s checkpoint
Optimizer updates	857,664	340,309	125,551
Micro-batch $\times$ accumulation	$64\times 1$	$64\times 1$	$8\times 4$
Training roots	11,785,831	7,257,646	1,338,827
Processed roots	54.9 M	21.8 M	4.02 M
Nominal epochs	4.66	3.0	3.0

Appendix

Table 10: Stage-specific training settings and exposure.

Analysis	Fitting and evaluation	Draws / inference
Primary transfer; terminal ablations	Four inner TRAIN folds; refit TRAIN, evaluate VAL+TEST	1,000; score CIs and descriptive ranks
Cross-duration transfer	Same TRAIN/VAL+TEST roles and window centers	1,000; Holm-8 Retention, Holm-11 Extend
RAE-PPG cross-duration OOF	Five outer / four inner folds	Same inference as fixed-partition cross-duration evaluation
LOSO	One held-out participant; four inner folds	1,000; unadjusted contrasts
Feature decoding; feature ablations	Five outer / four inner folds; group-weighted fits; Huber selection, $\alpha\in\{10^{-1},\ldots,10^{6}\}$	5,000, dataset-stratified; retention/extension: Holm-4, BH-18; ablations: unadjusted
Feature-to-task Bridge	Nested TRAIN feature-probe cross-fitting ( $\alpha=1$ ) and task-probe selection; VAL+TEST evaluation	1,000; 32 reassignment controls; relation-wise lower bounds

Appendix

Table 11: Evaluation protocols. Folds and bootstrap resampling group participants or cases, and paired comparisons use shared draws. Task-probe search grids and selection metrics are specified in Appendix B.3 . Intervals condition on the fitted encoders and probes.

Dataset	Task(s)	Input (s)	Metric	TRAIN	Final evaluation
PPG-DaLiA	Heart rate	10	MAE	316 (9)	216 (6)
	Activity	30	macro-F1	311 (9)	207 (6)
BUT PPG	Heart rate	10	MAE	2,492 (30)	1,396 (20)
DeepBeat	Atrial fibrillation	10	AUROC	7,990 (119)	21,617 (30)
	Signal quality	10	macro-F1	7,990 (119)	21,617 (30)
PTT-PPG	SBP; DBP	10	MAE	84 (14)	48 (8)

Appendix

Table 12: Primary downstream tasks and evaluation cohorts. Counts are eligible examples (participants); final evaluation uses VAL+TEST.

Model	$f_{s}$ (Hz)	Preprocessing used	Input handling	Representation	Dim.
PaPaGei-P	125	z-score; 0.5–12 Hz BP; resample	10 s tiles	512-D embedding; tile mean + SD	512/1,024
PaPaGei-S	125	z-score; 0.5–12 Hz BP; resample	10 s tiles	512-D embedding; tile mean + SD	512/1,024
Pulse-PPG	50	Window z-score; no filter; resample	Full variable-length input	Global max	512
SIGMA-PPG	50	0.5–8 Hz BP; resample; min–max $[-1,1]$	2 s patches; native positions	Mean of non-class patch tokens; final normalization	360
AnyPPG	125	0.5–8 Hz BP; resample; window z-score	Full variable-length input	Internal global mean	512
RAE-PPG	64	12 Hz LP; resample; masked root z-score	10/30/240 s; 64 tokens	Signal-token mean	384

Appendix

Table 14: Model-specific interfaces for downstream evaluation. Sampling rates and dimensions refer to model inputs and probe representations. BP and LP denote band-pass and low-pass filtering. The DeepBeat-specific RAE-PPG preprocessing is described in Appendix B.2 .

Model	Checkpoint file	Source commit	SHA-256 prefix
PaPaGei-P	papagei_p.pt	0c537da	3a6850961af5
PaPaGei-S	papagei_s.pt	0c537da	79d68671e51b
Pulse-PPG	checkpoint_best.pkl	716eaf9	485ade5033b3
SIGMA-PPG	sigma_student_strict_ b185072e.pt	b1cc8d9	3fe728902855
AnyPPG	anyppg_ckpt.pth	661b877	99b9bb0a3c2b

Appendix

Table 15: Comparison checkpoint files, source revisions and SHA-256 prefixes.

Task / duration	RAE-PPG	Best external model	$\Delta U$ [95% CI]
DaLiA heart rate 10 s; MAE $\downarrow$	8.6138	AnyPPG 7.3293	$-1.2844$ [-2.0133, -0.4897]
DaLiA activity 30 s; macro-F1 $\uparrow$	0.3937	AnyPPG 0.3397	$+0.0540$ [-0.0021, +0.1194]
WESAD stress 30 s; AUROC $\uparrow$	0.9484	AnyPPG 0.9407	$+0.0077$ [-0.0080, +0.0242]
WESAD arousal 30 s; AUROC $\uparrow$	0.8016	AnyPPG 0.7781	$+0.0235$ [-0.1299, +0.1607]
WESAD valence 30 s; AUROC $\uparrow$	0.7095	PaPaGei-S 0.6549	$+0.0546$ [-0.0998, +0.2477]
PTT SBP 10 s; MAE $\downarrow$	9.4724	PaPaGei-P 9.4932	$+0.0207$ [-0.6675, +0.6648]

Appendix

Table 16: Primary-duration LOSO evaluation.

\Delta U

is RAE-PPG minus the observed-best external model, with utility defined as

-\mathrm{MAE}

, AUROC, or macro-F1. Brackets give paired, unadjusted 95% participant-bootstrap intervals. Bold marks the better point score.

	Observed scores		Bootstrap mean
Model	Firsts /18	Rank $\downarrow$	Firsts /18	Rank $\downarrow$
RAE-PPG	12	1.444	11.217	1.656
AnyPPG	5	2.389	4.647	2.480
SIGMA-PPG	0	3.722	0.697	3.795
Pulse-PPG	0	3.944	0.528	3.958
PaPaGei-P	1	4.167	0.572	3.997

Appendix

Table 17: Rank stability across the 18 downstream tasks. Bootstrap summaries use the 1,000 shared participant/case resamples within each task.

			Five additional runs		Rank range	Runs
Downstream task	Metric	Primary	Mean $\pm$ SD	Range	(six runs)	ahead /6
DaLiA: HR 10	MAE $\downarrow$	8.652	8.460 $\pm$ 0.262	8.152–8.864	2	0
BUT PPG: HR 10	MAE $\downarrow$	10.514	10.518 $\pm$ 0.112	10.441–10.716	2–4	0
PTT-PPG: SBP 10	MAE $\downarrow$	9.529	9.629 $\pm$ 0.064	9.561–9.709	1	6
PTT-PPG: DBP 10	MAE $\downarrow$	5.449	5.367 $\pm$ 0.055	5.283–5.433	1	6
DREAMT: Age 10	MAE $\downarrow$	11.833	11.909 $\pm$ 0.061	11.813–11.966	4–5	0

Appendix

Table 18: Sensitivity to the pretraining seed. Primary denotes the run used for the main results; five additional runs vary only the random seed.

Tokens $K$	Best-score tasks /18	Mean rank $\downarrow$
64	12	1.44
128	12	1.44
256	12	1.44

Appendix

Table 19: Token-count sensitivity on the 18 primary tasks. Each setting uses one complete pretraining run and the same five fixed comparison checkpoints.

Task	10 s	30 s	240 s	Retention	Extend
DaLiA heart rate MAE $\downarrow$ , ref. 10 s	8.6522 [6.3069–11.1771]	7.0623 [5.2881–8.7911]	8.8074 [7.8910–9.8989]	$\times$	–
DaLiA activity F1 $\uparrow$ , ref. 30 s	0.3097 [0.2425–0.3547]	0.4078 [0.3115–0.5070]	0.4581 [0.4109–0.4975]	$\times$	$\times$
WESAD stress AUROC $\uparrow$ , ref. 30 s	0.8436 [0.7286–0.9640]	0.8821 [0.7522–0.9750]	0.9410 [0.8083–1.0000]	$\times$	$\times$
WESAD arousal AUROC $\uparrow$ , ref. 30 s	0.7527 [0.6300–0.8685]	0.8571 [0.7583–0.9593]	0.7875 [0.6763–0.8964]	$\times$	$\checkmark$
WESAD valence AUROC $\uparrow$ , ref. 30 s	0.6907 [0.6469–0.7590]	0.6973 [0.6023–0.8406]	0.6080 [0.4887–0.7766]	$\times$	$\times$
MESA sleep staging F1 $\uparrow$ , ref. 30 s	0.3607 [0.3556–0.3656]	0.3821 [0.3758–0.3889]	0.4735 [0.4640–0.4833]	$\checkmark$	$\checkmark$

Appendix

Table 20: Fixed-partition cross-duration evaluation of RAE-PPG. Probes are selected and fitted on TRAIN and evaluated on VAL+TEST. Brackets give 95% participant/case-cluster bootstrap intervals.

Task	10 s	30 s	240 s	Retention	Extend
MESA SpO 2 desaturation burden MAE $\downarrow$ , ref. 240 s	0.2258 [0.2228–0.2291]	0.2166 [0.2137–0.2195]	0.1768 [0.1739–0.1797]	–	$\checkmark$
SDB AHI-code classification AUROC $\uparrow$ , ref. 240 s	0.4286 [0.2802–0.6050]	0.5000 [0.3483–0.6486]	0.6779 [0.5250–0.8184]	–	$\times$
DREAMT age MAE $\downarrow$ , ref. 10 s	11.8329 [9.1310–14.7515]	11.1262 [8.6348–13.9619]	10.3018 [7.5973–13.2041]	$\times$	–
DREAMT sleep staging F1 $\uparrow$ , ref. 30 s	0.2643 [0.2327–0.2999]	0.2687 [0.2361–0.3071]	0.3014 [0.2705–0.3322]	$\checkmark$	$\times$
DREAMT A/H detection AUROC $\uparrow$ , ref. 240 s	0.4958 [0.4282–0.5627]	0.4500 [0.3684–0.5383]	0.5225 [0.4355–0.6143]	–	$\checkmark$
Confirmed tasks / applicable tasks				2/8	6/11

Appendix

Table 20: Fixed-partition cross-duration evaluation of RAE-PPG (continued).

Task	10 s	30 s	240 s	Retention	Extend
DaLiA heart rate MAE $\downarrow$ , ref. 10 s	8.4987 [7.1158–10.1447]	7.3716 [6.2626–8.6637]	8.9185 [7.8384–10.3123]	$\times$	–
DaLiA activity F1 $\uparrow$ , ref. 30 s	0.3148 [0.2824–0.3469]	0.3795 [0.3291–0.4273]	0.4651 [0.4070–0.5196]	$\checkmark$	$\checkmark$
WESAD stress AUROC $\uparrow$ , ref. 30 s	0.7792 [0.6938–0.8729]	0.9446 [0.9071–0.9762]	0.8742 [0.7992–0.9582]	$\times$	$\checkmark$
WESAD arousal AUROC $\uparrow$ , ref. 30 s	0.7430 [0.6458–0.8500]	0.7556 [0.6825–0.8368]	0.7398 [0.6140–0.8644]	$\times$	$\times$

Appendix

Table 21: Grouped out-of-fold cross-duration evaluation of RAE-PPG, using five outer and four inner participant-grouped folds. Brackets give 95% participant/case-cluster bootstrap intervals; Retention and Extend follow the criteria in Appendix B.8 .

Task	10 s	30 s	240 s	Retention	Extend
WESAD valence AUROC $\uparrow$ , ref. 30 s	0.6757 [0.6011–0.7900]	0.6792 [0.5729–0.8272]	0.6414 [0.5312–0.7686]	$\times$	$\times$
MESA sleep staging F1 $\uparrow$ , ref. 30 s	0.3594 [0.3563–0.3627]	0.3789 [0.3754–0.3830]	0.4688 [0.4633–0.4749]	$\checkmark$	$\checkmark$
MESA A/H burden MAE $\downarrow$ , ref. 240 s	0.0934 [0.0920–0.0949]	0.0920 [0.0906–0.0934]	0.0840 [0.0827–0.0854]	–	$\checkmark$
MESA arousal-event burden MAE $\downarrow$ , ref. 240 s	0.0519 [0.0512–0.0526]	0.0504 [0.0498–0.0511]	0.0436 [0.0430–0.0443]	–	$\checkmark$
MESA SpO 2 desaturation burden MAE $\downarrow$ , ref. 240 s	0.2255 [0.2235–0.2275]	0.2164 [0.2144–0.2184]	0.1779 [0.1760–0.1798]	–	$\checkmark$
SDB AHI-code classification AUROC $\uparrow$ , ref. 240 s	0.5023 [0.4121–0.5992]	0.5769 [0.4737–0.6762]	0.6730 [0.5847–0.7647]	–	$\times$

Appendix

Table 21: Grouped out-of-fold cross-duration evaluation of RAE-PPG (continued).

A. Features introduced at 10 s; centered 10 s readout
Signal feature / readout	$N$	10 s	30 s	240 s
Slope transit time	85,150	0.930 [0.915, 0.942]	0.935 [0.922, 0.945]	0.782 [0.758, 0.805]
APG $b/a$	84,187	0.767 [0.720, 0.812]	0.775 [0.727, 0.819]	0.670 [0.624, 0.714]
APG $d/a$	77,251	0.643 [0.586, 0.697]	0.627 [0.572, 0.680]	0.450 [0.394, 0.500]
Onset-to-APG- $b$	84,187	0.866 [0.850, 0.881]	0.859 [0.842, 0.875]	0.727 [0.707, 0.746]
Normalized crest time	85,153	0.864 [0.841, 0.885]	0.855 [0.833, 0.875]	0.628 [0.597, 0.657]

Appendix

Table 22: Decoding of the sixteen signal features from the final frozen encoder. Column headings give input duration, and entries report gain [95% CI]. Readouts are specified by panel or row.

Task(s)	Source measurement and setting	Reported finding
Rate
DaLiA HR ∗ ; BUT PPG HR ∗	Smartphone PPG HR compared with simultaneous 10 s ECG recordings in a clinic subset ( Avram et al., 2019 ) .	HR was compared with ECG in that acquisition setting; the DaLiA and BUT PPG settings differ.
PTT SBP; PTT DBP	Resting HR and BP measured by the same automated BP monitor after seated rest ( Liu et al., 2010 ) .	Adjusted population-level associations of HR with SBP and DBP; not a PPG-based prediction experiment.
DeepBeat AF ∗	PPG from AF and sinus-rhythm groups; 120 s segments from 10 min recordings ( Basza et al., 2023 ) .	Univariate HR difference; the groups also differed in age and clinical characteristics.
DREAMT age	Smartphone PPG HR in a large cohort ( Avram et al., 2019 ) .	HR was associated with age; the source used smartphone rather than wrist PPG.
WESAD stress ∗	30 s fingertip-PPG intervals during Stroop and rest ( Tsai et al., 2025 ) .	Mean RR interval decreased during Stroop phases.

Appendix

Table 23: Literature sources for the fixed 33 Bridge relations. Tasks sharing the same source and finding are grouped within each feature family. Asterisks mark the 22 core anchors. Source measurements and findings, including mixed or nonsignificant results, are distinct from the Bridge outcomes reported in Table 24 A.

Task(s)	Source measurement and setting	Reported finding
Morphology / vascular timing (continued)
DREAMT sleep; MESA sleep	Fingertip-PPG timing and shape by sleep stage ( Yilmaz et al., 2023 ) .	Stage differences in normalized timing and shape; the present input durations are fixed evaluation settings.
Beat-interval variability
WESAD stress ∗	30 s fingertip-PPG RMSSD during Stroop and rest ( Tsai et al., 2025 ) .	Sex-specific RMSSD responses were described; pooled adjacent-phase differences were not marked significant.
WESAD arousal ∗	ECG normalized RMSSD over 60 s; the source specified at least 1 min for HRV analysis ( Brouwer et al., 2013 ) .	RMSSD was associated with arousal but not valence; the source signal was ECG.
DaLiA activity	Five-minute wrist-PPG variability at 06:00–07:00 compared with mean daily steps over the preceding 90 days ( Natarajan et al., 2020 ) .	Activity associations were quantified for HF and LF power; analogous findings for other HRV metrics were stated without data. Habitual steps differ from contemporaneous activity labels.

Appendix

Table 23: Literature sources for the fixed Bridge relations (continued).

Task	Seconds	Rate	Morph.	BIV	Wave.	SQ
DaLiA HR	10	$R^{*}\;+/+$	$U\;+/+$	–	–	$U\;-/-$
BUT PPG HR	10	$R^{*}\;-/-$	$U\;+/+$	–	–	$U\;-/-$
PTT SBP	10	$I^{\ddagger}\;-/-$	$R^{*}\;+/-$	–	–	$U\;-/-$
PTT DBP	10	$I^{\ddagger}\;-/-$	$I^{*}\;+/+$	–	–	$U\;-/-$
DeepBeat AF	10	$R^{*}\;-/-$	$R^{*\ddagger}\;+/+$	–	–	$U\;+/+$
DeepBeat quality	10	$U\;+/+$	$C\;+/+$	–	–	$C\;+/+$

Appendix

Table 24: Bridge results for all task–feature-family combinations (A) and source-sensitivity subsets (B).

Analysis set	Relations	PPG positive	Decoded positive	Both positive
Fixed analysis set	33	20	19	18
Alternative R/I subset	31	20	19	18
Excluding the two A/H–BIV relations	31	19	18	17
Both restrictions	29	19	18	17

Appendix

Table 24: Bridge results for all task–feature-family combinations and source-sensitivity subsets (continued).

	Nested/protected		Fully shared
Downstream task	A: ON	D: OFF	B: ON	C: OFF
PPG-DaLiA heart rate 10 s, MAE $\downarrow$	8.6522 [6.3069, 11.1771]	8.6522 [6.3069, 11.1771]	9.8918 [7.9631, 11.9085]	10.8341 [8.8499, 12.8399]
BUT PPG ECG-referenced heart rate 10 s, MAE $\downarrow$	10.5144 [8.2084, 13.3728]	10.5144 [8.2084, 13.3728]	10.8212 [8.0354, 14.1561]	10.7846 [8.0600, 14.0935]
DeepBeat atrial fibrillation 10 s, AUROC $\uparrow$	0.7774 [0.7164, 0.8087]	0.7774 [0.7164, 0.8087]	0.7040 [0.6700, 0.7336]	0.6895 [0.6464, 0.7170]
DeepBeat signal quality 10 s, F1 $\uparrow$	0.6718 [0.6360, 0.6903]	0.6718 [0.6360, 0.6903]	0.6601 [0.6158, 0.6916]	0.6647 [0.6297, 0.6882]
PTT-PPG systolic blood pressure 10 s, MAE $\downarrow$	9.5291 [5.2400, 14.8673]	9.5291 [5.2400, 14.8673]	9.8373 [5.5890, 15.0733]	10.0162 [5.5610, 15.2647]

Appendix

Table 25: Final downstream performance of the four ablation conditions. Each score is followed by its 95% participant/case-cluster bootstrap interval. Bold marks the best point score, including ties; F1 denotes macro-F1. Condition definitions are given in Appendix D.1 .

Task / input	Checkpoints	B: ON	C: OFF
DaLiA HR / 10 s	S $\to$ M	-0.6653 [-2.4852, +1.5717]	-1.4716 [-2.7188, +0.2765]
DaLiA HR / 10 s	S $\to$ L	-0.8051 [-2.0708, +1.4527]	-1.7474 [-3.0385, +0.4282]
BUT PPG HR / 10 s	S $\to$ M	+0.0412 [-0.3957, +0.4515]	-0.1614 [-0.4752, +0.0820]
BUT PPG HR / 10 s	S $\to$ L	-0.1816 [-0.5692, +0.1106]	-0.1450 [-0.4798, +0.1277]
DeepBeat AF / 10 s	S $\to$ M	-0.0568 [-0.0797, -0.0207]	-0.0356 [-0.0562, -0.0238]
DeepBeat AF / 10 s	S $\to$ L	-0.0654 [-0.0903, -0.0197]	-0.0799 [-0.0970, -0.0431]

Appendix

Table 26: Changes in downstream utility across pretraining checkpoints for fixed input windows. Entries are later-minus-earlier differences with paired, unadjusted 95% participant-cluster intervals. Utility is

-\mathrm{MAE}

, AUROC, or macro-F1, so positive differences indicate improvement.

Task / input	Checkpoints	B: ON	C: OFF
PTT SBP / 10 s	S $\to$ M	-0.0806 [-0.5095, +0.3675]	-0.2695 [-0.7449, +0.1594]
PTT SBP / 10 s	S $\to$ L	-0.2082 [-0.8874, +0.4442]	-0.3870 [-1.0548, +0.3268]
PTT DBP / 10 s	S $\to$ M	+0.4238 [-0.7027, +1.3793]	+0.2218 [-0.9236, +1.2553]
PTT DBP / 10 s	S $\to$ L	+0.2420 [-0.7670, +1.1346]	+0.1105 [-0.9515, +1.0945]
DREAMT age / 10 s	S $\to$ M	+0.7183 [-0.0397, +1.6900]	+0.5945 [-0.3183, +1.6819]
DREAMT age / 10 s	S $\to$ L	+1.2587 [+0.3049, +2.7789]	+1.1196 [+0.1184, +2.6454]

Appendix

Table 26: Changes in downstream utility across pretraining checkpoints for fixed input windows (continued).

Fixed input / feature set	Checkpoints	B: ON	C: OFF	Stage-change difference: B $-$ C
10 s / nine 10 s features	S $\to$ M	$-$ 0.1324 [ $-$ 0.1401, $-$ 0.1248]	$-$ 0.1637 [ $-$ 0.1732, $-$ 0.1542]	+0.0313 [+0.0252, +0.0372]
10 s / nine 10 s features	S $\to$ L	$-$ 0.1660 [ $-$ 0.1747, $-$ 0.1575]	$-$ 0.1866 [ $-$ 0.1965, $-$ 0.1769]	+0.0205 [+0.0140, +0.0271]
30 s / two 30 s features	M $\to$ L	$-$ 0.3128 [ $-$ 0.3327, $-$ 0.2931]	$-$ 0.3807 [ $-$ 0.4036, $-$ 0.3583]	+0.0679 [+0.0519, +0.0840]
10 s / nine 10 s features	M $\to$ L (secondary)	$-$ 0.0336 [ $-$ 0.0392, $-$ 0.0282]	$-$ 0.0228 [ $-$ 0.0297, $-$ 0.0161]	-0.0107 [-0.0197, -0.0019]

Appendix

Table 27: Changes in earlier-feature decoding across pretraining checkpoints at fixed input duration. B and C report later-minus-earlier feature-average gain changes; the final column reports the paired difference between those stage changes. Entries include unadjusted 95% percentile intervals.

Earlier features in longer inputs: aligned readout
		Nested/protected		Fully shared
Feature introduction / input	$n$	A: ON	D: OFF	B: ON	C: OFF
10 s / 30 s	9	0.8303 [0.8149, 0.8458]	0.7097 [0.6944, 0.7248]	0.6634 [0.6487, 0.6784]	0.6348 [0.6206, 0.6485]
10 s / 240 s	9	0.6441 [0.6284, 0.6591]	0.5866 [0.5710, 0.6017]	0.6647 [0.6497, 0.6790]	0.6104 [0.5973, 0.6234]
30 s / 240 s	2	0.6661 [0.6443, 0.6871]	0.5341 [0.5104, 0.5574]	0.7470 [0.7304, 0.7624]	0.5297 [0.5069, 0.5517]
New features at introduction duration: token mean

Appendix

Table 28: Feature decodability after pretraining. Panel A reports feature-average gain [95% CI] for each condition; Panel B reports paired gain differences for the same feature sets and readouts. Feature introduction / input gives the targets' introduction duration and the encoder input duration.

Feature introduction / input	$n$	A $-$ D	B $-$ C	A $-$ B
Earlier features in longer inputs: aligned readout
10 s / 30 s	9	+0.1205 [+0.1130, +0.1285]	+0.0287 [+0.0224, +0.0351]	+0.1668 [+0.1579, +0.1760]
10 s / 240 s	9	+0.0575 [+0.0504, +0.0645]	+0.0544 [+0.0470, +0.0620]	-0.0207 [-0.0277, -0.0137]
30 s / 240 s	2	+0.1320 [+0.1180, +0.1459]	+0.2173 [+0.2004, +0.2352]	-0.0809 [-0.0960, -0.0669]
New features at introduction duration: token mean
30 s / 30 s	2	+0.0060 [+0.0011, +0.0110]	+0.0723 [+0.0574, +0.0879]	+0.2927 [+0.2730, +0.3127]

Appendix

Table 28: Feature decodability after pretraining (continued).

Input	Newly introduced feature	FULL-PRIOR INITIAL gain	LEARNED gain	Error reduction (%)
30 s	Normalized reflected delay	0.7346	0.8449	41.57 [34.42, 47.88]
	PRV-RMSSD	0.8396	0.8771	23.41 [16.56, 29.74]
	Aggregate: two features	0.7871 [0.7670, 0.8061]	0.8610 [0.8472, 0.8731]	34.72 [29.23, 39.77]
240 s	PRV-pNN50	0.6294	0.7712	38.25 [32.35, 43.85]
	Uncentered ACF index	0.5973	0.7654	41.74 [35.43, 48.01]
	PPGV LF fraction	0.3998	0.5472	24.55 [20.22, 28.63]

Appendix

Table 29: Decoding newly introduced features after matched training continuations. Error reduction is relative to FULL-PRIOR INITIAL, as defined in Appendix D.4 . Brackets give unadjusted 95% participant/case-cluster bootstrap intervals; error-reduction intervals are paired.

Model	Preprocessing and objective	Admitted data	Optimization and checkpoint
PaPaGei-P/S	10 s, 125 Hz; 0.5–12 Hz Chebyshev-II filtering and 50 ms smoothing; segment $z$ -score; flat-line screening of the $z$ -scored segment and morphology-range admission. S: sVRI-bin contrastive loss with IPA/SQI regression. P: same-participant positives.	11,350,415 windows; 10,126 participants/cases.	Adam, $10^{-4}$ ; global batch 1,024; 15,000 updates. Five runs per model, differing only in random-number streams.
Pulse-PPG	240 s, 50 Hz; no filtering; per-participant $z$ -score. Learned motif-based distance (MotifDist), then relative contrastive training (RelCon).	1,337,278 windows; 2% of participants/cases held out for validation-loss checkpoint selection.	MotifDist: 20 epochs, batch 16, $10^{-3}$ . RelCon: 6 epochs, batch 64, $10^{-4}$ . a Lowest-validation-loss checkpoint at each stage.
SIGMA-PPG	240 s, 50 Hz; 0.5–8 Hz second-order Butterworth filtering; min–max scaling; quality scores over 2 s patches. Vector-quantized tokenizer, then teacher-guided masked modeling.	1,337,261 windows; released 80/20 file split within eligible TRAIN data.	Tokenizer: 100 epochs. Masked modeling: 80 epochs. Global batch 4,096; masked-modeling gradient accumulation 2.
AnyPPG	10 s PPG–ECG pairs. PPG: 0.5–8 Hz band-pass, 125 Hz. Lead-II ECG: 0.5 Hz high-pass, 50 Hz notch, 500 Hz. Flat-segment and ECG-quality screening; segment $z$ -score. Contrastive alignment with a frozen, pretrained single-lead ECGFounder teacher ( Li et al., 2025 ) .	8,205,102 pairs, MIMIC-III only; 2% of participants/cases held out for validation-loss checkpoint selection.	AdamW, $5\times 10^{-4}$ ; 2,560 pairs per process, two processes, accumulation 1. Run: 24 epochs, 37,848 optimizer updates. Lowest-validation-loss saved checkpoint: 33,116 updates. b

Appendix

Table 30: Model-specific retraining procedures. Data are drawn from the RAE-PPG pretraining TRAIN partition; counts refer to the admitted pools before the recipe-specific validation split where one is used.

A. RAE-PPG and the five retrained models
Downstream task	RAE-PPG	PaPaGei-P	PaPaGei-S	Pulse-PPG	SIGMA-PPG	AnyPPG
Regression: MAE $\downarrow$
DaLiA: HR 10	8.6522	8.7955 ▲	17.6340 ▼	14.4006 ▼	14.2939	9.0165 ▼
BUT PPG: HR 10	10.5144	10.1444 ▲	10.8978	10.9398 ▼	10.6272	10.1215 ▼
PTT-PPG: SBP 10	9.5291	9.6366	10.2725	10.3561 ▼	10.4055	10.2046
PTT-PPG: DBP 10	5.4489	5.5092	5.8239	6.0458 ▼	6.0637	5.9476

Appendix

Table 31: Downstream performance after comparison-model retraining. Panel A compares the unchanged RAE-PPG checkpoint with the five retrained checkpoints; bold and underline mark the best and second-best point scores. Panel B summarizes paired comparisons. Released-checkpoint scores are reported in Table 2 .

A. RAE-PPG and the five retrained models
Downstream task	RAE-PPG	PaPaGei-P	PaPaGei-S	Pulse-PPG	SIGMA-PPG	AnyPPG
Classification: AUROC $\uparrow$ ( $\dagger$ macro-F1 $\uparrow$ )
DeepBeat: AF 10	0.7774	0.7750 ▲	0.6541	0.6397	0.7275	0.8197 ▼
DeepBeat: Quality 10 $\dagger$	0.6718	0.6639 ▲	0.5986	0.6360 ▼	0.6139 ▼	0.6366
WESAD: Stress 30	0.9514	0.9499 ▲	0.8288	0.7983	0.8778	0.9035 ▼
WESAD: Arousal 30	0.8571	0.7875	0.6978	0.6886	0.7289	0.7637

Appendix

Table 31: Downstream performance after comparison-model retraining (continued).

A. Summary across the 18 tasks
Summary	RAE-PPG	PaPaGei-P	PaPaGei-S	Pulse-PPG	SIGMA-PPG	AnyPPG
Best-score tasks /18	12	2	0	0	2	2
Mean rank $\downarrow$	1.556	2.778	4.722	5.000	3.500	3.444

Appendix

Table 31: Downstream performance after comparison-model retraining (continued).

Self-Supervised Pre-Training

Self-supervised pre-training learns transferable representations from unlabeled data to reduce the labels needed for downstream tasks.

See all 250 papers

Oct 24, 2026

Unlocking the Specialization Capability of PPG Foundation Models for Wearable Signals through Structured Pruning and Adaptive Fine-Tuning

Photoplethysmography (PPG) plays a central role in wearable health monitoring and clinical decision support. However, existing approaches to universal PPG representation learning are predominantly trained on high-quality large-scale datasets obtained from oximeters in hospital settings and often struggle to adapt to sparse and low-quality real-world datasets, thereby limiting their generalization to clinically relevant tasks on prevalent PPG data such as collected from wearable devices. Surprisingly, even after full fine-tuning, PPG foundation models cannot consistently outperform general-purpose time-series models. Therefore, a key question is how to effectively adapt PPG foundation models to target prediction tasks. Through empirical studies on various PPG foundation models, we find that pretrained models often exhibit inherent sparsity and computational redundancy, suggesting that PPG foundation models have learned to activate task-relevant network substructures for different prediction tasks. To preserve this valuable prior knowledge, we propose a structured pruning-and-fine-tuning framework, SPAFT, which regularizes the adaptive fine-tuning process by constraining it to a more relevant and compact parameter space based on parameter importance and structural redundancy. Extensive experiments across four PPG foundation models, five benchmark datasets, and ten downstream tasks demonstrate that SPAFT achieves win rates of 82.8% and 87.5% on regression and classification tasks, respectively. This prune-then-finetune paradigm often enables PPG foundation models to achieve state-of-the-art performance, laying a promising foundation for the real-world clinical deployment of PPG foundation models and the development of personalized PPG foundation models.

Huaming Du, Zhixian Xie, Gaoyang Zhen +6

Southwestern University of Finance and Economics · Jizhi Technology Co., Ltd. · Fuzhou University +2

Time-Series Foundation Models Structured Pruning Wearable Health Monitoring

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

Oct 9, 2026

PPG-Bench: A Unified Benchmark for PPG Foundation Models

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.

Arvasu Kulkarni, Mahir Jain, Aditya Ray Mishra +2

Mannas AI

Photoplethysmography Time-Series Foundation Models

Ask PDF

Abstract

Self-Supervised Pre-Training

Explore similar work

Unlocking the Specialization Capability of PPG Foundation Models for Wearable Signals through Structured Pruning and Adaptive Fine-Tuning

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

PPG-Bench: A Unified Benchmark for PPG Foundation Models