DNA-PPG: A Foundation Model for Photoplethysmography via Dual Neighborhood Alignment

1 Introduction

The proliferation of telemedicine, wearable devices, and intensive care technologies has yielded massive clinical physiological time series. However, the deployment of deep learning in this domain is hindered by a severe "annotation bottleneck": high-quality clinical labeling is costly, expert-dependent, and prone to inter-annotator noise. Consequently, supervised paradigms fail to capitalize on the vast archives of accumulated unlabeled data. To bridge this gap, Self-Supervised Learning (SSL) has emerged as a dominant paradigm, leveraging the intrinsic structure of data to learn representations without supervision. This has catalyzed the development of physiological foundation models, which aim to overcome the inherent fragility of traditional task-specific approaches by establishing robust, transferable representations across heterogeneous devices and patient populations.

1 This work was done during an internship at Xiaomi Corporation. Code and pre-trained checkpoints are provided at https: //github.com/yyzyyz001/DNA-PPG.

To realize this vision, Contrastive Learning (CL) has become the standard for physiological foundation models due to its robust unsupervised learning capabilities. Representative frameworks, such as SimCLR [Chen et al., 2020], established the instance discrimination paradigm by treating intra-batch samples as hard negatives. This paradigm was subsequently adapted for PPG pre-training [Abbaspourazad et al., 2023], achieving patient-level feature disentanglement by strictly defining signals from different subjects as negatives. Building upon this, recent works like PaPaGei [Pillai et al., 2024] have attempted to integrate domain knowledge into the contrastive objective. Specifically, the PaPaGei-S variant employs a discretization (binning) strategy, where continuous hand-crafted morphological features are quantized into finite intervals to guide feature learning.

However, these paradigms face fundamental limitations. First, instance discrimination approaches are constrained by the rigid hard-negative assumption. By indiscriminately repelling pseudo-negative samples from different subjects, they overlook intrinsic morphological similarities-where distinct individuals often exhibit highly coherent time-frequency dynamics. This rigidity distorts the natural manifold structure of the feature space. Second, while approaches like PaPaGei-S incorporate domain priors, the discretization strategy disrupts the continuity of physiological states, inducing precision loss; furthermore, reliance on hand-crafted indices hinders transferability to other modalities or acquisition conditions.

To bridge these gaps, we propose DNA-PPG, a hybrid pre-training framework anchored in Dual Neighborhood Alignment. To mitigate the hard-negative bias, we design the Morphology-Aware Self-Supervised Branch (MorphSSL). Beyond standard subject consistency, it incorporates a Time-Frequency Soft Weighting (TF-Soft) mechanism that leverages spectral priors to measure dynamic affinity, adaptively relaxing the repulsion of morphologically similar false negatives to preserve time-frequency invariant neighborhoods. Simultaneously, to overcome discretization limitations, we introduce the Physiological Semantic Alignment Branch (Phys-Align). Discarding coarse binning, this branch models physiological indicators (e.g., Blood Pressure, Heart Rate, and Oxygen Saturation) directly as continuous manifold, enforcing the learning of smooth semantic transitions consistent with continuous physiological variations. Extensive ablations validate this synergy, demonstrating that their joint optimization significantly outperforms single-branch variants.

Our contributions are summarized as follows:

• We propose DNA-PPG, a novel physiological foundation model framework anchored in Dual Neighborhood Alignment. By addressing the manifold distortion from rigid hard-negatives and the precision loss from discretization, it constructs a continuous, structurepreserving representation space superior to existing instance discrimination or binning-based paradigms.
• We design two synergistic mechanisms to align feature neighborhoods: Morph-SSL leverages the TimeFrequency Soft Weighting (TF-Soft) mechanism to maintain morphology-invariant neighborhoods, while Phys-Align exploits continuous physiological indicators to form smooth semantic manifolds. Their synergistic optimization ensures representations possess both structural robustness and semantic precision.
• Extensive experiments, leveraging a pre-training dataset of 10.7 million segments from over 8,400 subjects, demonstrate that DNA-PPG outperforms SOTA base-

lines, achieving an 18% MAE reduction in regression and an 11.5% improvement in classification performance. These results validate its efficacy as a robust universal feature extractor for PPG signals, evidenced by robust generalization across diverse healthcare domains.

2 RelatedWork

2.1 PPG-based HealthMonitoring

Photoplethysmography (PPG) has emerged as a ubiquitous wearable sensing modality for non-invasive health monitoring [Nie et al., 2024]. Deep learning has significantly advanced this field, showing remarkable efficacy in tasks such as blood pressure estimation [Wang et al., 2020; Kim et al., 2022], arrhythmia detection [Kwon et al., 2019], and sleep apnea analysis [Papini et al., 2020]. However, despite their performance, these fully supervised pipelines heavily rely on costly expert annotations and often suffer from poor generalization across varying devices and protocols. These limitations underscore the necessity of label-efficient pre-training for learning robust and transferable PPG representations.

2.2 Self-Supervised Learning for Time Series

Self-supervised learning (SSL) for time series has primarily categorized into discriminative and generative paradigms. Discriminative approaches, such as SimCLR [Chen et al., 2020] and BYOL [Grill et al., 2020], learn representations by maximizing similarity between augmented views of the same instance, utilizing strategies like negative sampling or asymmetric architectures to derive distinctive features. Conversely, generative frameworks like TiMAE [Li et al., 2023] adopt masked modeling to capture temporal dependencies by reconstructing missing segments in the latent or signal space.

To enhance the robustness and generalization, recent studies have focused on refining training objectives. Addressing the false negative issue in rigid contrastive learning, SoftCLT [Lee et al., 2023] employs soft assignments across instance and temporal dimensions, while SupCon [Khosla et al., 2020] utilizes semantic labels to align intra-class samples. Furthermore, TF-C [Zhang et al., 2022] incorporates spectral priors to ensure time-frequency consistency. Inspired by the scaling laws of large language models, the field is transitioning towards foundation models: Chronos [Ansari et al., 2024] tokenizes time series for forecasting, while MOMENT [Goswami et al., 2024] leverages massive pre-training on the "Time Series Pile" to derive universal features.

However, applying these generic paradigms to physiological signals remains suboptimal. Most methods treat time series as homogeneous data, overlooking the fine-grained morphological details critical for clinical analysis. Moreover, general foundation models typically ignore the continuous manifold structure of physiological states, failing to align representations with continuous clinical indicators. This necessitates a domain-specific framework that explicitly harmonizes morphological integrity with physiological continuity.

2.3 FoundationModels for Physiological Signals

Recent work explores foundation models for diverse physiological modalities to learn universal representations. For instance, RelCon [Xu et al., 2024] leverages motif-based distances for accelerometry to learn fine-grained motion semantics, while Brant [Zhang et al., 2023] employs dual-encoder Transformers to model spatiotemporal dependencies in EEG. Similarly, WBM [Erturk et al., 2025] aggregates longitudinal metrics to encode high-level behavioral patterns.

In the specific domain of PPG, [Abbaspourazad et al., 2023] established a benchmark by scaling contrastive learning to over 141,000 participants. However, this approach relies on a rigid instance discrimination paradigm, where signals from different subjects are strictly treated as negative pairs. This strategy indiscriminately repels morphologically similar samples solely due to subject identity, thereby distorting the natural feature manifold. To incorporate domain guidance, PaPaGei-S [Pillai et al., 2024] aligns representations with hand-crafted metrics like the stress-induced vascular response index (sVRI). While incorporating priors, it relies on discretizing continuous features to construct positive pairs. Such quantization introduces hard decision boundaries that fragment the intrinsic continuity of physiological variations and induce precision loss. Moreover, the dependence on specific pre-defined indices limits the framework's adaptability. This highlights the urgent need for a unified approach that respects continuous physiological semantics while flexibly maintaining morphological neighborhoods.

3 Methodology

3.1 Overview

The proposed DNA-PPG framework serves as a foundation model designed to learn universal, transferable representations for photoplethysmography (PPG) data. The overall workflow is structured into two distinct phases: a hybrid pre-training phase and a downstream inference phase. Figure 2 specifically illustrates the architecture of the proposed hybrid pre-training phase. Formally, given a raw input sequence x, we apply stochastic augmentations to generate a transformed view ˜x. The shared backbone encoder fθ(·) maps this view into a high-dimensional representation vector h = fθ(˜x). During pre-training, h is subsequently projected by a non-linear projection head gϕ(·) into a latent embedding z = gϕ(h) for contrastive optimization.

During the pre-training phase, the framework employs a Dual Neighborhood Alignment strategy to construct a semantically rich feature manifold. This is achieved by optimizing two complementary branches simultaneously. The Morphology-Aware Self-Supervised Branch (MorphSSL) is designed to capture intrinsic signal dynamics and structural patterns. By employing a Time-Frequency Soft Weighting mechanism, this branch identifies samples with coherent time-frequency dynamics across different individuals, thereby mitigating the manifold distortion caused by rigid hard-negative sampling. Parallel to this, the Physiological Semantic Alignment Branch (Phys-Align) explicitly incorporates physiological priors into the representation learning process. By treating physiological indicators as auxiliary supervision signals, this branch forces the model to capture the intrinsic continuity of physiological states, ensuring that signals reflecting similar health conditions are naturally clustered in the embedding space. These two branches are jointly optimized via a progressive learning strategy, enabling the shared encoder to synergize morphological robustness with physiological semantic precision.

After pre-training, the projection head gϕ(·) is discarded. The encoder fθ(·) is frozen and serves as a universal feature extractor. Lightweight, task-specific heads-such as linear regressors or classifiers-are attached to the representation h, enabling efficient adaptation to diverse healthcare applications with minimal overhead.

3.2 Morphology-Aware Self-Supervised Branch

In the Morph-SSL branch, we adopt a foundational contrastive assumption where two distinct PPG segments cropped from the same subject are treated as a natural positive pair (i, i′ ), while all other samples within the mini-batch of size B serve as candidate negatives. Given an anchor i and target j, the standard InfoNCE [Oord et al., 2018] probability pI(i, j) is defined as:

where z is the normalized latent embedding and τ = 0.1 is the temperature hyperparameter.

Standard contrastive paradigms rely on a rigid, hardnegative assumption, indiscriminately repelling samples from different subjects. This ignores intrinsic morphological affinities and distorts the feature manifold. While direct similarity metrics (e.g., Euclidean distance or cosine similarity) could theoretically address this, they prove inadequate for raw PPG waveforms due to their susceptibility to phase misalignments and amplitude scaling, which induce high variance even among semantically similar signals. To address these challenges, we introduce the Time-Frequency Soft Weighting mechanism. This module integrates a dedicated auxiliary encoder based on the TF-C architecture [Zhang et al., 2022], pre-trained on our unannotated PPG dataset. By performing joint time-frequency modeling, the encoder extracts robust representation vectors vi that capture essential signal dynamics. Crucially, this auxiliary module remains frozen during the main training phase. Consequently, the additional computational overhead is limited to a single forward pass per sample, ensuring a negligible burden on the overall training efficiency. Specifically, we quantify the pairwise affinity in this auxiliary latent space using cosine similarity s(vi, vj):

To reduce noise and prevent signal dilution from dominant negatives, we define the Morphology-Aware weight wij as:

where I(·) is the indicator function. Here, ξ is a similarity threshold, set to 0.5 to filter out weakly correlated pairs. The scaling factor β = 0.05 is empirically chosen to attenuate the influence of soft negatives, thereby preserving the dominance of the primary positive pair (i, i′ ).

By assigning soft weights to sequences with similar timefrequency morphologies, we preserve local neighborhood structures within the InfoNCE framework. This is formalized through a calibrated target distribution T M i , where the positive pair (i, i′ ) is assigned a fixed weight of 1, and negative pairs utilize wij . These weights are normalized to yield a valid probability distribution:

The final Morphology-Aware loss is defined as the crossentropy between the soft target distribution T M i and the predicted probability distribution pI(i, ·), effectively mitigating the adverse effects of repelling morphologically similar samples:

3.3 Physiological Semantic Alignment Branch

To structure the representation manifold consistently with physiological priors, we introduce the Physiological Semantic Alignment Branch (Phys-Align). Importantly, within the DNA-PPG framework, physiological indicators are not utilized as input features for the backbone encoder fθ(·). Instead, they serve strictly as auxiliary supervision signals to calibrate the embedding space, ensuring that signals reflecting similar health conditions are naturally clustered.

We treat the physiological state of the i-th sample as a continuous semantic vector yi ∈ RK , comprising key hemodynamic and physiological indicators: blood pressure (BP), oxygen saturation (SpO2), and heart rate (HR), i.e., yi = [BP, SpO2 ,HR]⊤. Given physiological priors yi ∈ RK and validity masks mi ∈ {0, 1}K , we first apply robust standardization to handle varying scales and potential outliers. For the k-th dimension, the standardized value ˜ yi,k is computed as:

where µk and σk denote the median and the Median Absolute Deviation (MAD) of the valid samples, respectively, serving as robust estimators for location and scale.

We define a robust pairwise distance metric δij,k for missing data. Assuming missing entries follow N(0, 1) in the standardized space (samples with all physiological labels missing are filtered out during preprocessing; see Sec. 4.2), the expected squared distance per dimension is:

which are aggregated across all K dimensions as Dij = 1K ∑Kk=1 δij,k.

Unlike prior works such as PaPaGei-S [Pillai et al., 2024], which rely on rigid, interval-based discretization for hard contrastive sampling, our approach adopts a soft weighting mechanism. Such hard boundaries often lead to information loss and artificial manifold fragmentation. In contrast, we transform the Euclidean distance Dij into an affinity matrix A, enabling the model to capture fine-grained degrees of physiological similarity. The diagonal is masked to zero to avoid self-reinforcement:

where γ = 1.0 is the bandwidth parameter. The physiological target distribution is then obtained by row-normalization, T P ij = Aij/ ∑Bl=1 Ail, and the physiological alignment loss LP is the cross-entropy between this distribution and the contrastive probability pI(i, j) in Eq. (1):

3.4 Joint Optimization Strategy

To achieve holistic representation learning, we propose a joint objective that synergizes the intrinsic structural robustness captured by Morph-SSL with the precise clinical semantics encoded by Phys-Align. The total objective function L is defined as:

where α denotes the dynamic coefficient that modulates the injection of physiological priors into the representation space.

We implement a progressive learning strategy, inspired by curriculum learning [Bengio et al., 2009] principles. Specifically, α is initialized at 0 and linearly ramps up to 0.7 throughout the training process. This schedule allows the backbone encoder to first establish a stable manifold based on signal dynamics via Morph-SSL before gradually incorporating complex physiological semantic constraints. This strategy effectively prevents the high variance or noise inherent in physiological labels from disrupting feature alignment in the early stages.

4 Experiments

4.1 Datasets

We evaluate our framework using a large-scale pre-training dataset and six downstream benchmarks. Following the PaPaGei [Pillai et al., 2024] benchmark, we align evaluation settings for fair comparison while utilizing a refined preprocessing pipeline. Statistics are summarized in Tables 1 and 2.

Pre-training Datasets

We leverage unlabeled PPG signals from VitalDB [Lee et al., 2022] and MESA [Zhang et al., 2018; Chen et al., 2015] (see Table 1). To unify the format, raw signals were processed into non-overlapping 10-second segments. Unlike the original PaPaGei benchmark, we exclude MIMIC-III [Johnson et al., 2016] to ensure full reproducibility, given its recent access restrictions.

Downstream Datasets

We evaluate the generalization capability of our foundation model across six downstream datasets summarized in Table 2. Regression Tasks These tasks evaluate the model's precision in hemodynamic estimation using Mean Absolute Error (MAE). For PPG-BP [Liang et al., 2018], we estimate three indicators: Systolic Blood Pressure (SysBP), Diastolic Blood Pressure (DiasBP), and Heart Rate (HR). Similarly, VV [Toye, 2023] is used for SysBP and DiasBP estimation. For PPG-DaLiA [Reiss et al., 2019], which contains ambulatory data with motion artifacts, we perform HR estimation. Classification Tasks All classification tasks are formulated as binary problems to assess the model's ability to capture high-level physiological semantics, evaluated using the Area Under the Receiver Operating Characteristic (AUROC). For sleep analysis, we utilize SDB [Garde et al., 2014], where the original four-level Apnea-Hypopnea Index (AHI) is converted into a binary target by distinguishing between normal (0) and abnormal (> 0) cases. In the domain of affective computing, we address label subjectivity on WESAD [Schmidt et al., 2018] and ECSMP [Gao et al., 2021] by adopting a median-split protocol. Specifically, the continuous scores for Valence and Arousal in WESAD, as well as the Total Mood Disturbance (TMD) in ECSMP, are binarized based on their population medians to differentiate between high and low physiological states.

4.2 Implementation Details

Data Preprocessing Pipeline

We adopt a standardized pipeline for pre-training data to ensure signal quality, while employing task-specific strategies for downstream benchmarks. Pre-training Processing. For VitalDB, raw PPG signals are first resampled to 125 Hz and synchronized with physiological labels. We isolate continuous valid intervals, applying bandpass filtering (0.5-12 Hz) and smoothing to eliminate baseline wander and highfrequency noise, before segmenting into non-overlapping 10second windows. To ensure data integrity, we discard segments where the PPG signal is entirely absent or where both HR and SpO2 labels are missing. Additionally, flatline artifacts are filtered utilizing a dynamic threshold (0.01× the 5th-95th percentile amplitude range). MESA follows a similar procedure, processing raw signals in 20-minute slices. Finally, valid segments from both sources undergo Z-score normalization. Downstream Processing. Following pretraining, downstream signals undergo bandpass filtering, resampling (125 Hz), and Z-score normalization. Variablelength inputs are adjusted to a fixed 10 seconds via symmetric padding or center-cropping. For regression tasks involving discrete measurements (PPG-BP, VV), we aggregate feature representations by averaging multiple segments per subject. For continuous monitoring in PPG-DaLiA, a sliding window protocol (8s window, 2s stride) is applied. For longterm recordings (WESAD, SDB, ECSMP), we trim the initial and final transition periods to exclude noise, and subsequently sample representative 10-second clips from the remaining effective duration.

Backbone Architecture

The backbone encoder maps raw 1D PPG signals into a highdimensional feature space. To determine the most suitable architecture for physiological feature extraction, we evaluated three representative networks: a ViT-1D [Dosovitskiy, 2020] with 12 Transformer blocks, which processes the sequence via patching and aggregates global information utilizing the [CLS] token, and convolutional baselines EfficientNet-1D [Tan and Le, 2019] and ResNet-1D [He et al., 2016]. Empirically, we adopt ResNet-1D, configured for a 512-dimensional output, as our default backbone, as it yielded the superior performance in our contrastive learning framework (validated in Sec. 4.4).

Training Setup

Experiments were conducted on a single NVIDIA A800 GPU. Models were trained for 10 epochs (batch size 256) with fixed 10-second inputs (1,250 time steps). We employed the Adam optimizer [Kingma, 2014] (initial lr=1 × 10-3 ) and a cosine annealing scheduler [Loshchilov and Hutter, 2016] featuring 5% linear warmup and subsequent decay to zero. During training, data augmentations-including Gaussian noise, random scaling, and time warping-were applied to construct diverse views for contrastive learning.

4.3 Comparison

We compare DNA-PPG against seven baselines: statistical features (Hand-crafted), general contrastive learning methods (SimCLR [Chen et al., 2020], BYOL [Grill et al., 2020]), time-series specific frameworks (TF-C [Zhang et al., 2022]), and time-series foundation models (Chronos [Ansari et al., 2024], Moment [Goswami et al., 2024]). We also include PaPaGei-S [Pillai et al., 2024] as a representative PPGfocused benchmark, reporting the better result between the official checkpoint and a version retrained on our pre-training data. Notably, REGLE [Yun et al., 2024] is excluded because only its spirogram inference code is available; applying a spirogram-trained model to PPG creates a modality mismatch, rendering the comparison unreliable.

The quantitative results are summarized in Table 3. In the regression tasks (Panel A), DNA-PPG outperforms all baselines with the lowest average MAE of 9.50. Our model shows a significant advantage in hemodynamic estimation, reducing the SysBP error on PPG-BP by approximately 18% compared to PaPaGei-S. This improvement suggests that our Physiological Semantic Alignment Branch effectively captures fine-grained physiological semantics, which general contrastive methods often fail to model. Furthermore, on the noise-prone PPG-DaLiA dataset, DNA-PPG maintains superior performance (MAE 13.87) over foundation models like Chronos andMoment, indicating that theMorphology-Aware Self-Supervised Branch disentangles intrinsic physiological patterns from motion artifacts.

For classification tasks (Panel B), DNA-PPG achieves the highest average AUROC of 0.661. In affective computing benchmarks (WESAD, ECSMP), our framework demonstrates exceptional sensitivity to subtle signal variations, surpassing Moment in WESAD Valence detection by a large margin (0.737 vs.0.616). While PaPaGei-S exhibits a slight edge in sleep apnea detection (SDB), DNA-PPG delivers a more balanced and robust performance across the full spectrum of cardiovascular, sleep, and emotion tasks, verifying its capability as a versatile foundation model.

AUROC (↑)

4.4 Ablation Study

Contribution of Key Components

Table 4 quantifies the contribution of each module. Singlebranch variants (Settings 1 & 3) yield suboptimal generalization, whereas their combination (Setting 4) significantly improves performance, confirming the synergy between morphological learning and physiological alignment. Furthermore, incorporating Time-Frequency Soft Weighting (TFSoft) consistently boosts results across all configurations, evidenced by the improvements of Setting 2 over 1 and DNAPPG over Setting 4. The full framework achieves the best metrics (0.661 AUROC, 9.50 MAE), verifying that refining sample interactions via time-frequency consistency is crucial for robust representation learning.

Backbone Analysis

To determine the optimal backbone encoder for our framework, we first evaluated three candidate architectures: ResNet-1D, ViT-1D, and EfficientNet-1D. As shown in the left panel of Fig. 3, ResNet-1D yields the best performance, attaining the lowest regression error (9.50 MAE) and highest classification score (0.661 AUROC). Based on this, we further investigate model scaling by varying the channel width of ResNet-1D (1.32M, 4.99M, and 19.41M parameters). As plotted in the right panel of Fig. 3, performance does not improve monotonically with scale. While the 4.99M variant yields significant gains over the 1.32M baseline, further scaling to 19.41M results in diminishing returns: regression error decreases negligibly (9.50 → 9.46 MAE) while classification performance degrades (0.661 → 0.652 AUROC). Consequently, we adopt the 4.99M ResNet-1D as the default backbone to balance performance and computational efficiency.

Sensitivity to Dual-Branch Balancing

Finally, we analyzed the sensitivity to the hyperparameter α ∈ {0.3, 0.5, 0.7, 0.9}, which controls the trade-off between the morphological and physiological objectives. As plotted in Fig. 4, the performance varies significantly with α. While α = 0.5 achieves the highest classification score (0.668 AUROC), it incurs a substantial increase in regression error (10.62 MAE). Setting α = 0.7 provides the most balanced performance, achieving the lowest MAE (9.50) while maintaining a competitive AUROC (0.661). Therefore, we adopt α = 0.7 as the default setting.

5 Conclusion

In this paper, we presented DNA-PPG, a unified foundation model addressing manifold distortion and semantic precision loss in physiological representation learning. Through a Dual Neighborhood Alignment strategy, we bridge intrinsic signal morphology and clinical semantics. Specifically, TimeFrequency Soft Weighting captures inter-subject morphological affinities, while the Physiological Semantic Alignment branch exploits continuous targets for smooth health representations. Comprehensive experiments across cardiovascular, sleep, and affective domains verify that DNA-PPG significantly outperforms existing state-of-the-art frameworks. Future work will extend this paradigm to multi-modal signals and low-resource environments. Moreover, we plan to leverage the vast wearable-collected PPG signals lacking physiological indicators by using prediction-algorithm outputs as approximate gold standards, thereby further scaling pre-training with pseudo-supervision.