HiPPT: Physiology-Grounded Wearable PPG Representation Learning via Multi-Scope Text Alignment

Jiaqi Wang; Wei Chen; Xiaoyu Li; Lexian Xia; ShuaiLong Tang; Chenran Zhang; Donglin Xie; Guangpu Zhu

cs.AIOct 10, 2026

HiPPT: Physiology-Grounded Wearable PPG Representation Learning via Multi-Scope Text Alignment

Authors: Jiaqi Wang, Wei Chen, Xiaoyu Li, Lexian Xia, ShuaiLong Tang, Chenran Zhang, Donglin Xie, Guangpu Zhu,

Organizations: Southeast University · OPPO · Shenzhen Technology University · Peking University

Abstract

Reusable wearable PPG representations should capture local pulse morphology and broader physiological regularities. Existing pretraining objectives, however, usually apply supervision at the recording or pair level, overlooking that different concepts become reliable over different scopes of waveform evidence. We propose HiPPT, a physiology-grounded framework that matches each supervision source to its appropriate evidence scope. We construct a PPG-Text corpus of over 4 million wrist-worn recordings from 202K users, pairing recordings with waveform-derived morphology and metadata-derived context descriptions. HiPPT combines record-level morphology alignment, which grounds individual recordings in pulse characteristics, with cross-recording context alignment, which aggregates recordings sharing the same metadata tuple to capture context-consistent regularities. Waveform-guided text modeling further recovers masked contextual concepts from PPG patch representations, providing fine-grained waveform-text supervision. Across six downstream tasks spanning blood pressure, vascular stiffness, physiological attributes, and cross-dataset heart-rate estimation, HiPPT achieves the best result on the primary linear-probing metrics. The aligned space also supports zero-shot profiling of age group, sex, and BMI category, individually and jointly, with performance improving as more recordings are aggregated. These results show that evidence-scope matching yields transferable and semantically accessible wearable PPG representations. Code and pretrained weights will be released.

Figures & tables

Figure 1: Overview of HiPPT. Record-Level Morphology Alignment grounds each PPG recording in its waveform-derived morphology description. Cross-Recording Context Alignment aggregates a collection of recordings sharing the same metadata tuple and aligns the aggregated representation with the corresponding context description. Waveform-Guided Text Modeling predicts masked contextual concepts conditioned on the remaining text and PPG patch representations, providing fine-grained waveform–text supervision.

Method	SBP		DBP		cfPWV
Method	MAE $\downarrow$	$R^{2}\uparrow$	MAE $\downarrow$	$R^{2}\uparrow$	Balanced ACC $\uparrow$	PR-AUC $\uparrow$
SimCLR ( Chen et al. 2020 )	11.89 $\pm$ 0.69	0.09 $\pm$ 0.04	6.71 $\pm$ 0.13	0.09 $\pm$ 0.04	0.84 $\pm$ 0.03	0.80 $\pm$ 0.04
MoCo-v3 ( Chen et al. 2021 )	11.98 $\pm$ 0.73	0.07 $\pm$ 0.04	6.67 $\pm$ 0.23	0.08 $\pm$ 0.04	0.85 $\pm$ 0.05	0.80 $\pm$ 0.06
MAE ( He et al. 2021 )	11.95 $\pm$ 0.47	0.07 $\pm$ 0.05	6.81 $\pm$ 0.14	0.05 $\pm$ 0.06	0.84 $\pm$ 0.03	0.80 $\pm$ 0.03
PaPaGei ( Pillai et al. 2024 )	12.42 $\pm$ 0.48	0.00 $\pm$ 0.05	6.94 $\pm$ 0.31	0.00 $\pm$ 0.04	0.83 $\pm$ 0.05	0.79 $\pm$ 0.05
Apple ( Abbaspourazad et al. 2023 )	12.01 $\pm$ 0.84	0.07 $\pm$ 0.06	6.67 $\pm$ 0.28	0.09 $\pm$ 0.05	0.85 $\pm$ 0.03	0.79 $\pm$ 0.03

Table 1: Linear-probing transfer to systolic blood pressure (SBP) and diastolic blood pressure (DBP) regression on HBPM-X, and high-stiffness arterial classification on cfPWV. Results are mean

\pm

std under 5 random seeds. Best results are in bold , and second-best results are underlined .

Method	Age		Sex		HR
Method	MAE $\downarrow$	$R^{2}\uparrow$	ACC $\uparrow$	ROC-AUC $\uparrow$	MAE $\downarrow$	$R^{2}\uparrow$
SimCLR ( Chen et al. 2020 )	7.68 $\pm$ 0.89	0.34 $\pm$ 0.11	0.68 $\pm$ 0.07	0.78 $\pm$ 0.08	8.87 $\pm$ 3.46	0.44 $\pm$ 0.30
MoCo-v3 ( Chen et al. 2021 )	7.73 $\pm$ 0.83	0.34 $\pm$ 0.11	0.66 $\pm$ 0.09	0.77 $\pm$ 0.09	9.24 $\pm$ 3.83	0.40 $\pm$ 0.34
MAE ( He et al. 2021 )	7.08 $\pm$ 0.82	0.41 $\pm$ 0.11	0.74 $\pm$ 0.05	0.80 $\pm$ 0.08	8.36 $\pm$ 3.11	0.44 $\pm$ 0.37
PaPaGei ( Pillai et al. 2024 )	8.83 $\pm$ 0.74	0.19 $\pm$ 0.11	0.66 $\pm$ 0.07	0.73 $\pm$ 0.08	8.29 $\pm$ 3.53	0.43 $\pm$ 0.42
Apple ( Abbaspourazad et al. 2023 )	7.53 $\pm$ 0.82	0.37 $\pm$ 0.10	0.68 $\pm$ 0.07	0.79 $\pm$ 0.07	8.18 $\pm$ 3.44	0.50 $\pm$ 0.39

Table 2: Linear-probing results for attribute probing on HBPM-X and external heart-rate transfer on PPG-DaLiA dataset. Best results are in bold , and second-best results are underlined .

Method	Age Group	Sex	BMI	Joint Acc.
Random	0.2591	0.4860	0.2384	0.0384
Supervised	0.7979	0.6610	0.4056	0.2136
Morphology Alignment	0.2749	0.4881	0.3514	0.0379
+ Context Alignment	0.6272	0.8189	0.5133	0.2560
+ WGTM (HiPPT)	0.7078	0.8313	0.5350	0.2955

Table 3: Zero-shot user-profiling accuracy for age group, sex, BMI, and their joint prediction. The supervised row is a task-specific reference trained with downstream labels. Best results are in bold , and second-best results are underlined .

Figure 2: Zero-shot user-profiling accuracy for different number of PPG recordings used to construct aggregated representation for a single subject.

Panel of Figure 3 — Figure 3: T-SNE visualization of two methods on the HBPM-X. Users with the same attribute tuple $a_{s}$ are marked in the same color.

Figure 4: Visualization of unimodal retrieval. The red star denotes the ground truth attribute tuple of the query PPG, and we use circle to plot the attribute distribution of recalled PPG signals(Age group distribution on the top and BMI group distribution on the right). We also visualize the result by Apple in the X-label.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: User-level attribute distribution of the HBPM dataset.

Figure 6: User-level attribute distribution of the VS dataset.

Figure 7: cfPWV cohort label statistics. (a) Class counts at the subject and record levels, revealing inverted imbalance. (b) Subject-level continuous cfPWV distribution with the clinical threshold at

10\,\mathrm{m/s}

Figure 8: Records-per-subject distribution in the cfPWV cohort. The histogram is truncated at 300 for readability; the boxplot shows the full range. Abnormal subjects contribute substantially more recordings on average.

Symbol	Definition
$T$	Pulse-cycle duration (s)
$\mathrm{IHR}$	Instantaneous heart rate (bpm)
$\mathrm{prop\_s}$	Fraction of the cycle spent on the systolic upslope
$\Delta T$	Interval between systolic and diastolic peaks (s)
$t_{\mathrm{sys}}$	Systolic duration (s)
$t_{\mathrm{dia}}$	Diastolic duration (s)

Appendix

Table 4: Handcrafted PPG features used to curate morphology descriptions.

Clause	Slots	Example surface form
Cycle	$T$ , IHR	cardiac cycle 0.96s, 62.76 bpm within normal range
Sys/Dia	$t_{\mathrm{sys}}$ , $t_{\mathrm{dia}}$	ventricular contraction 0.38s, relaxation 0.58s
Prop	$\mathrm{prop\_s}$	systolic upslope proportion 0.16
DeltaT	$\Delta T$	0.32s duration systolic-diastolic peak interval
AW	$\mathrm{AW}_{s}$ , $\mathrm{AW}_{d}$	amplitude to width ratio: systolic 1.20, diastolic 0.80

Appendix

Table 5: Clause inventory for morphology descriptions.

Variable	Condition	Category
Age	Age $<18$	adolescent
	$18\leq\mathrm{Age}<30$	young adult
	$30\leq\mathrm{Age}<50$	middle-aged
	Age $\geq 50$	elderly
Sex	male / female	male / female
BMI	BMI $<18.5$	underweight

Appendix

Table 6: Discretization rules for metadata categories.

Method	SBP MAE $\downarrow$	DBP MAE $\downarrow$	cfPWV PR-AUC $\uparrow$	Age MAE $\downarrow$	Sex ROC-AUC $\uparrow$
SimCLR ( Chen et al. 2020 )	12.40 $\pm$ 0.77	6.92 $\pm$ 0.29	0.70 $\pm$ 0.23	8.02 $\pm$ 0.79	0.76 $\pm$ 0.09
MoCo-v3 ( Chen et al. 2021 )	12.75 $\pm$ 1.10	7.19 $\pm$ 0.43	0.71 $\pm$ 0.21	9.28 $\pm$ 1.30	0.75 $\pm$ 0.09
MAE ( He et al. 2021 )	12.61 $\pm$ 0.47	7.21 $\pm$ 0.19	0.76 $\pm$ 0.26	7.15 $\pm$ 0.82	0.79 $\pm$ 0.05
PaPaGei ( Pillai et al. 2024 )	12.62 $\pm$ 0.53	7.34 $\pm$ 0.22	0.73 $\pm$ 0.24	7.66 $\pm$ 0.88	0.80 $\pm$ 0.07
Apple ( Abbaspourazad et al. 2023 )	12.24 $\pm$ 0.89	6.98 $\pm$ 0.30	0.73 $\pm$ 0.26	8.54 $\pm$ 1.42	0.77 $\pm$ 0.08
SupCon ( Khosla et al. 2020 )	12.27 $\pm$ 0.52	7.07 $\pm$ 0.22	0.64 $\pm$ 0.23	8.23 $\pm$ 1.00	0.82 $\pm$ 0.08

Appendix

Table 7: Linear-probing results with 10% of the labeled training data. Results follow the same evaluation protocol as the full-data experiments. Best results are in bold ; second-best results are underlined . Ties are marked equally.

Method	SBP MAE $\downarrow$	DBP MAE $\downarrow$	cfPWV PR-AUC $\uparrow$	Age MAE $\downarrow$	Sex ROC-AUC $\uparrow$
SimCLR ( Chen et al. 2020 )	12.14 $\pm$ 0.84	6.75 $\pm$ 0.30	0.68 $\pm$ 0.12	7.91 $\pm$ 0.80	0.77 $\pm$ 0.07
MoCo-v3 ( Chen et al. 2021 )	12.24 $\pm$ 0.83	6.79 $\pm$ 0.29	0.66 $\pm$ 0.08	8.75 $\pm$ 1.14	0.76 $\pm$ 0.08
MAE ( He et al. 2021 )	12.15 $\pm$ 0.32	6.94 $\pm$ 0.16	0.77 $\pm$ 0.06	7.14 $\pm$ 0.76	0.81 $\pm$ 0.08
PaPaGei ( Pillai et al. 2024 )	12.89 $\pm$ 0.48	7.17 $\pm$ 0.19	0.76 $\pm$ 0.10	7.70 $\pm$ 0.84	0.80 $\pm$ 0.07
Apple ( Abbaspourazad et al. 2023 )	11.97 $\pm$ 0.80	6.62 $\pm$ 0.20	0.74 $\pm$ 0.17	7.63 $\pm$ 0.78	0.78 $\pm$ 0.09
SupCon ( Khosla et al. 2020 )	12.12 $\pm$ 0.53	6.96 $\pm$ 0.15	0.73 $\pm$ 0.10	7.81 $\pm$ 0.76	0.82 $\pm$ 0.08

Appendix

Table 8: Linear-probing results with 30% of the labeled training data. Results follow the same evaluation protocol as the full-data experiments. Best results are in bold ; second-best results are underlined . Ties are marked equally.

Figure 9: Effect of pretraining data scale. HiPPT is pretrained using 10%, 25%, 50%, and 100% of the full corpus. (a) Zero-shot accuracy, recorded-sex probe AUROC, and cross-modal Recall@5. (b) SBP, DBP, and age MAE. Higher is better in (a), whereas lower is better in (b).

PPG encoder states	Masked-token Acc. $\uparrow$
Normal	0.709
Zero	0.172
Random	0.186
Drop (Normal $-$ Zero)	0.537

Appendix

Table 9: Lights-out diagnosis of WGTM under different PPG encoder states supplied to decoder cross-attention.

Method	SBP		DBP		cfPWV
Method	MAE $\downarrow$	$R^{2}\uparrow$	MAE $\downarrow$	$R^{2}\uparrow$	Balanced ACC $\uparrow$	PR-AUC $\uparrow$
HiPPT	11.67 $\pm$ 0.48	0.11 $\pm$ 0.03	6.57 $\pm$ 0.11	0.11 $\pm$ 0.05	0.85 $\pm$ 0.02	0.84 $\pm$ 0.03
HiPPT w/o Age	11.88 $\pm$ 0.50	0.08 $\pm$ 0.04	6.70 $\pm$ 0.16	0.08 $\pm$ 0.05	0.84 $\pm$ 0.03	0.80 $\pm$ 0.03
HiPPT w/o Sex	11.79 $\pm$ 0.43	0.10 $\pm$ 0.03	6.65 $\pm$ 0.14	0.09 $\pm$ 0.03	0.85 $\pm$ 0.03	0.81 $\pm$ 0.04
HiPPT w/o HR	11.77 $\pm$ 0.46	0.11 $\pm$ 0.04	6.73 $\pm$ 0.16	0.11 $\pm$ 0.03	0.84 $\pm$ 0.04	0.79 $\pm$ 0.03

Appendix

Table 10: Linear-probing transfer after removing one explicit attribute from pretraining descriptions. The age- and sex-related texts are removed from context description.

Method	Age		Sex		HR
Method	MAE $\downarrow$	$R^{2}\uparrow$	ACC $\uparrow$	ROC-AUC $\uparrow$	MAE $\downarrow$	$R^{2}\uparrow$
HiPPT	7.05 $\pm$ 0.79	0.42 $\pm$ 0.11	0.74 $\pm$ 0.06	0.83 $\pm$ 0.08	7.65 $\pm$ 3.09	0.49 $\pm$ 0.36
HiPPT w/o Age	7.22 $\pm$ 0.81	0.40 $\pm$ 0.11	0.71 $\pm$ 0.07	0.82 $\pm$ 0.07	8.55 $\pm$ 3.84	0.38 $\pm$ 0.54
HiPPT w/o Sex	7.08 $\pm$ 0.87	0.42 $\pm$ 0.11	0.74 $\pm$ 0.06	0.82 $\pm$ 0.09	7.72 $\pm$ 3.60	0.47 $\pm$ 0.51
HiPPT w/o HR	7.11 $\pm$ 0.82	0.42 $\pm$ 0.10	0.74 $\pm$ 0.06	0.83 $\pm$ 0.08	7.70 $\pm$ 3.59	0.48 $\pm$ 0.47

Appendix

Table 11: Attribute and external heart-rate probing after leave-one-attribute-out pretraining. The HR ablation removes the explicit instantaneous-heart-rate field from morphology descriptions.

Aggregation	$N$ =1	$N$ =2	$N$ =5	$N$ =10	$N$ =15
Attention	5.10 $\pm$ 0.42	10.21 $\pm$ 1.09	21.40 $\pm$ 2.37	29.55 $\pm$ 0.80	29.47 $\pm$ 1.89
Mean	5.57 $\pm$ 1.60	9.31 $\pm$ 1.96	17.46 $\pm$ 1.81	25.81 $\pm$ 2.21	27.49 $\pm$ 2.18
Max	4.88 $\pm$ 0.72	8.60 $\pm$ 1.36	14.40 $\pm$ 2.15	12.84 $\pm$ 1.53	13.00 $\pm$ 1.96

Appendix

Table 12: Joint zero-shot profiling accuracy (%) for different aggregation operators and numbers of PPG recordings per subject. Joint correctness requires age group, sex, and BMI category to be simultaneously correct, where acquisition time is marginalized. All results use top-

3

candidate retrieval. The best mean result in each column is shown in bold .

Figure 10: Zero-shot profiling accuracy versus retrieval depth

k

with 10 PPG recordings per subject. Shaded regions denote one standard deviation over five random trials.

Figure 11: Subject-level cfPWV classification error near versus far from the

10\,\mathrm{m/s}

clinical cutoff. Bars show mean

\pm

std over five splits (on average 54 near-threshold and 58 far-from-threshold test subjects per split).

Figure 12: Row-normalized BMI confusion matrix for zero-shot profiling, averaged over five random trials.

Representation Learning

Representation learning develops data transformations that expose useful structure for prediction, transfer, generation, and control.

See all 804 papers

CAP: Towards PPG Universal Representation Learning with Patient-level Supervision
Jun 13, 2026Chenyang He, Xinyi Shao, Shun Huang +4Representation Learning Cross-Modal Representation Learning
Nanjing University of Aeronautics and Astronautics Nanjing, Jiangsu, China · Peking University Beijing, Beijing, China · Independent Researcher +1
PPG-LM: A Photoplethysmography-Language Model with Multi-Level Clinical Alignment
Sep 27, 2026Xiaoda Wang, Minxiao Wang, Maxwell A Xu +10Multimodal Pretraining Photoplethysmography
Emory University · University of California, Los Angeles · Google +3
Multi-site PPG: An In-the-Wild Physiological Dataset from Emerging Multi-site Wearables
May 18, 2026Jiayi Shao, Jiaying Ye, Shengyao Liu +4Wearable Sensing Heart Rate Estimation
University of Washington · 2Purdue University

Jun 13, 2026J/K move · Enter open · S save

CAP: Towards PPG Universal Representation Learning with Patient-level Supervision

Chenyang He, Xinyi Shao, Shun Huang, Bosong Huang+3

Nanjing University of Aeronautics and Astronautics Nanjing, Jiangsu, China · Peking University Beijing, Beijing, China · Independent Researcher+1

Representation Learning Cross-Modal Representation Learning Multimodal Pretraining Photoplethysmography Multimodal Contrastive Learning Wearable Health Monitoring

Open paper AskPDF arXiv

Photoplethysmography (PPG) plays a central role in wearable health monitoring and clinical decision support. Yet existing approaches to universal PPG representation learning largely focus on signal-level objectives and often overlook patient-level health context, which limits generalization to complex clinical tasks and heterogeneous cohorts. To address this gap, we construct a large-scale paired PPG-EHR multimodal dataset by distilling fragmented medical histories and clinical records into cohesive, patient-level electronic health records (EHR). Building on this resource, we propose Clinical Anchored Pretraining for PPG (CAP). During pretraining, CAP performs cross-modal contrastive alignment that anchors PPG representations to patient-level clinical semantics, guiding the encoder beyond waveform fitting toward modeling consistency in a patient's overall physiological state. During downstream adaptation, the pretrained PPG encoder provides clinically grounded representations that strengthen inductive bias and improve robustness and transferability. Experiments demonstrate that CAP consistently outperforms strong baselines on four diverse downstream tasks. CAP achieves a particularly large gain on respiratory rate prediction (up to +87.6% relative improvement over the state-of-the-art baseline) and delivers an average relative +26.7% across all tasks. We further enhance the interpretability of our approach through comprehensive analyses, including ablations and multiple complementary visualizations of the learned representations. The code for our experiments is available at: https://github.com/gody123gody/CAP .

Abstract

Representation Learning

Explore similar work

CAP: Towards PPG Universal Representation Learning with Patient-level Supervision

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

Multi-site PPG: An In-the-Wild Physiological Dataset from Emerging Multi-site Wearables