Vision4PPG: PPG Analysis via Vision Models

• Vision4PPG is an emergent methodology that transforms one-dimensional PPG signals into 2D representations (e.g., spectrograms) to leverage vision foundation models.
• It employs parameter-efficient fine-tuning with LoRA and advanced spectral methods to extract features for accurate regression and classification of vital signs.
• Benchmarks demonstrate state-of-the-art performance and robust generalization across various physiological endpoints, supporting clinical and wearable applications.

Vision4PPG refers to an emergent methodology that leverages vision foundation models (VFMs)—originally developed for natural image and video analysis—to analyze photoplethysmography (PPG) signals for non-invasive physiological monitoring tasks, with a focus on vital signs such as blood pressure, heart rate, respiration, oxygen saturation, and related laboratory values. The core strategy is to represent the one-dimensional PPG time series as image-like two-dimensional inputs (e.g., spectrograms or recurrence plots), thereby enabling the application of large-scale, parameter-efficient vision models to extract feature representations and perform regression or classification for various physiological endpoints (Kataria et al., 11 Oct 2025).

1. Transformation of PPG Signals into 2D Representations

The central technical innovation in Vision4PPG is the “imagification” of the PPG waveform. The one-dimensional PPG signal $x[n]$ is transformed into two-dimensional representations suitable for vision Transformer backbones. A primary example is the Short-Time Fourier Transform (STFT), which decomposes the signal into its time–frequency representation:

$L(f, t) = \log{(|X(f, t)|^2 + \epsilon)}$

where $X(f, t)$ is the STFT of $x[n]$ and $L(f, t)$ is the resulting log-power spectrogram. The result is z-score normalized and replicated over three channels to yield a 3-channel "image." Alternate representations include:

• STFT phase channels (cos$\phi$, sin$\phi$)
• Recurrence plots (RP) computed from $x[n]$ and its first or second derivatives

This approach allows the use of VFMs such as DINOv3 and SIGLIP-2 by feeding in these 2D tensors as input, thus capitalizing on the models’ capability to capture and process complex spatial and spectral patterns present in the PPG data (Kataria et al., 11 Oct 2025).

2. Vision Foundation Model Architecture and Fine-Tuning

Vision4PPG repurposes standard vision Transformers, utilizing architectures such as DINOv3 and SIGLIP-2. Parameter-Efficient Fine-Tuning (PEFT), specifically Low-Rank Adaptation (LoRA), is employed during tuning to adapt the pre-trained VFM weights to PPG-specific tasks with minimal computational overhead. LoRA is incorporated selectively in the self-attention mechanism, modulating the $Q$, $K$, and $V$ matrices:

• LoRA hyperparameters: $r = 8$, $\alpha = 16$, $p_{drop} = 0.05$

Final regression heads process the pooled feature tokens, mapping to physiological output variables via:

$$\hat{y} = w_2^T \cdot \text{GELU}(W_1 \cdot \text{LN}(p) + b_1) + b_2$$

where $p$ is the pooled embedding, and $W_1$, $W_2$, $b_1$, $b_2$ are learned weights and biases (Kataria et al., 11 Oct 2025).

3. Performance on Physiological Monitoring Tasks

Vision4PPG achieves state-of-the-art results across a wide spectrum of physiological endpoints. Empirical benchmarks on seven BP datasets (including PPG-BP, Aurora Oscillometric/Auscultatory, CAS-BP, BCG, BUT-PPG) demonstrate that DINOv3 and SIGLIP-2, with LoRA tuning, consistently yield competitive or superior mean absolute error (MAE):

Dataset	SBP MAE (mmHg)	DBP MAE (mmHg)
Aurora-Oscillometric	18.57	5.32
CAS-BP	(SOTA)	(SOTA)

Additional tasks (heart rate, respiration, SpO$_2$, sodium, potassium, lactate) confirm the broad generalization properties, with DINOv3 leading in 3 of 8 tasks. For instance, heart rate estimation achieves errors on the order of 8.10–8.27 BPM, lactate errors at 1.22–1.24 mmol/L (Kataria et al., 11 Oct 2025).

4. Comparison with Time-Series Foundation Models (TSFMs)

Vision4PPG is systematically compared against TSFMs such as MOMENT (multi-domain, incl. ECG) and PPG-GPT (PPG-specific, trained on millions of ICU hours). Despite the sequential inductive bias of TSFMs, VFMs demonstrate:

• Higher or matched accuracy: VFMs win 9/14 BP estimation leaderboards and match or exceed TSFMs in auxiliary tasks
• Computational efficiency: Vision backbones + PEFT result in lower model size and fine-tuning cost than typical full TSFM training
• Flexibility: Robust to various input 2D representations (STFT, phase, RP) and generalize well to out-of-domain data

This strongly suggests that vision-trained models, when properly adapted, can outperform purpose-built sequence models in PPG analysis (Kataria et al., 11 Oct 2025).

5. Generalization and Task Versatility

Vision4PPG methodology is demonstrated to be robust not only to different input transforms but also to a range of vital sign and laboratory measurement tasks. The stability across heterogeneous datasets and target variables is highlighted by performance consistency and the ability to operate on out-of-distribution test sets. This generalization is attributed to:

• The pre-trained VFM’s exposure to diverse patterns in natural imagery
• The rich encoding obtained by multifaceted 2D transformations (including both spectral and recurrence-based views)
• The model’s abstract feature extraction and soft attention pooling, which together enable cross-task adaptability (Kataria et al., 11 Oct 2025)

6. Clinical Integration and Practical Implications

Vision4PPG holds significant promise for clinical deployment:

• A single, computationally light model enables estimation of BP, heart rate, respiration, SpO$_2$, and key laboratory markers non-invasively and in real time
• Minimal modification to existing vision infrastructure is required, as PPG signals can be “imagified” and processed via standard 2D input pipelines in VFMs
• PEFT strategies ensure low computational cost, which is particularly advantageous for deployment in wearables and point-of-care devices
• Out-of-domain generalization supports clinical robustness in diverse patient populations and across varying acquisition conditions

A plausible implication is that Vision4PPG may reduce the model burden for clinician-scientists and device engineers, streamlining physiological monitoring with a unified vision-based analytic framework (Kataria et al., 11 Oct 2025).

7. Future Research Directions

The findings in Vision4PPG open several investigative fronts:

• Combination of PPG with additional physiologically meaningful 2D transforms, such as wavelet scalograms
• Systematic paper of model calibration and input preprocessing for optimal downstream estimation accuracy
• Exploration of multi-modal fusion with other sensor streams (e.g., video, thermal imaging)
• Longitudinal and population-level studies to further validate generalization and address remaining edge cases

This suggests that Vision4PPG is a foundational step toward flexible, comprehensive, and clinician-ready physiological analysis systems powered by vision foundation models (Kataria et al., 11 Oct 2025).