Contrastive learning has shown great potential in the application of electrocardiogram (ECG) and photoplethysmogram (PPG) signals, providing several benefits in their analysis. One application is in unsupervised representation learning, where contrastive learning can extract meaningful features from large amounts of unlabeled ECG and PPG data. By contrasting similar and dissimilar signal segments, contrastive learning can capture underlying patterns and structure, enabling the model to learn representations that reflect the intrinsic characteristics of the signals. These learned representations can then be utilized for various downstream tasks such as arrhythmia detection, heart rate variability analysis, or blood pressure estimation.

Furthermore, contrastive learning can be employed for domain adaptation and transfer learning in ECG and PPG analysis. By pre-training a model on a large dataset from a source domain, such as a different hospital or population, and then fine-tuning it on a smaller labeled dataset from the target domain, the model can adapt and generalize well to the specific characteristics of the target domain. This is particularly useful in scenarios where labeled data in the target domain is limited or expensive to obtain. Contrastive learning allows the model to leverage the abundant unlabeled data from the source domain to learn robust and transferable representations, improving the performance and generalization capability on the target domain.

However, applying contrastive learning to ECG and PPG signals also presents challenges. One challenge is the selection of appropriate positive and negative pairs for the contrastive loss calculation. In the case of ECG and PPG signals, determining similar and dissimilar segments can be non-trivial due to variations caused by different physiological conditions, noise, artifacts, or measurement devices. Careful consideration is required to design effective strategies for constructing the positive and negative pairs that capture the relevant similarities and differences in the signals. Additionally, the computational complexity of contrastive learning can be demanding, especially when dealing with long ECG or PPG sequences. Efficient sampling methods and optimization techniques need to be employed to make the training process feasible and scalable.

In summary, contrastive learning offers valuable applications in ECG and PPG signal analysis, including unsupervised representation learning and domain adaptation. By leveraging unlabeled data and learning meaningful representations, contrastive learning can enhance the performance of various tasks in ECG and PPG analysis. However, careful consideration of pair construction and computational efficiency is essential to address the challenges specific to these signal types. A novel approach was introduced by John et al. [15] for determining the heart rate of individuals using facial video. Instead of relying on existing blood pressure data, this method focuses on extracting the photoplethysmography (PPG) signal present in facial expressions through self-supervised contrastive learning. To enhance the accuracy, a supervised heart rate estimator is trained to closely align with the actual heart rate measurements. Ye et al. [16] introduces a new technique called Discriminative Clustering Enhanced Adversarial Domain Adaptation (DC-ADA) for detecting mental stress across different individuals using ECG data. This method enhances the training process of existing bi-classifier adversarial UDA methods by incorporating a self-supervised loss on the target data. The proposed approach involves obtaining pseudo labels for the target data through a clustering-based strategy. These labels are then utilized as supervision to guide the classification of ambiguous target data into the appropriate stress levels.

Retraining models for new users in healthcare applications often necessi-tates the collection of a significant amount of labeled data, which can be both challenging and expensive. This is particularly true for applications like atrial fibrillation detection. To address the scarcity of labeled data, unsupervised and self-supervised techniques have emerged as promising solutions. Fonseca et al. [17] proposes the utilization of contrastive learning to enhance the performance of a CNN that classifies atrial fibrillation in scenarios with limited labeled data, small models, and noisy data. The effectiveness of this strategy was evaluated using the largest publicly available ECG dataset, and the study presents results in terms of the F1-score for various proportions of unlabeledlabeled data and different model sizes. The findings indicate that the proposed strategy surpasses the baseline approach, achieving up to a 30% improvement in the 10-fold mean F1-score compared to a 5.8% enhancement in AUC, which is the current state-of-the-art performance metric.

Remote Photoplethysmography (rPPG) systems offer a non-contact, costeffective, and widely available means of regularly monitoring heart rate (HR) [18]. They achieve this by analyzing the scattered light reflected from changes in blood volume in human skin tissues (known as PPG). However, these systems have not been widely adopted due to the absence of a standardized approach for estimating HR from skin videos in various practical situations. Conventional supervised methods necessitate a substantial number of accurately synchronized annotations between video and rPPG signals, which significantly hinder the development of comprehensive end-to-end rPPG models. Hasan et al. [19] introduces Self-rPPG, a technique that directly learns the optical and physiological mechanisms of rPPG from unlabeled videos without requiring synchronized rPPG signal annotations. The authors devised a self-supervised contrastive learning-based pretraining strategy to learn the representation of the inherent frequency and phase of diffusion signals, as well as the temporal coherence of video frames, using unlabeled sequences collected from multiple public datasets. They conducted experiments to determine the optimal contrastive learning schemes (loss functions and sampling strategy) and to evaluate the significance of the features learned by Self-rPPG. The results showed that the self-supervised representations can effectively encode the frequency and phase of diffusion signals, while also demonstrating resilience against temporal distortion. The effectiveness of different augmentations for contrastive self-supervised learning of electrocardiogram (ECG) signals and the optimal parameters are thoroughly examined in this study [20]. The baseline approach comprises two primary components: contrastive learning and the downstream task. Initially, an encoder is trained using various augmentations to extract ECG signal representations that are generalizable. Subsequently, the encoder is frozen, and a few linear layers are fine-tuned using differ- ent quantities of labeled data for arrhythmia detection in downstream tasks. [21] introduced ALPINE, a captivating approach named A noveL rPPG technique for Improving remote heart rate estimatioN using contrastive lEarning. ALPINE incorporates the power of contrastive learning to overcome limited labeled data, fostering data diversity for enhanced network generalization. Moreover, they introduce an innovative hybrid loss, encompassing contrastive loss, signal-to-noise ratio (SNR) loss, and data fidelity loss, further enhancing their methodology.

[21] introduces a remarkable unsupervised learning framework for signal regression, eliminating the reliance on labeled video data. By embracing the essence of periodicity and finite bandwidth, this approach unveils the blood volume pulse from unlabelled videos. The study concludes that promoting sparse power spectra within the realm of normal physiological frequencies, coupled with variance across batches of power spectra, proves ample for learning the visual attributes of periodic signals. [23] introduced a captivating Intra-Inter Subject Self-Supervised Learning (ISL) model tailored for multivariate cardiac signals. This model skillfully incorporates medical expertise to effectively glean insights from differences within and between subjects. For intra-subject selfsupervision, the ISL model harnesses a channel-wise attentional CNN-RNN encoder to extract heartbeat-level features from each individual. Additionally, a stationarity test module captures temporal dependencies among heartbeats. In the realm of inter-subject self-supervision, the model employs a series of data augmentations inspired by the unique clinical attributes of cardiac signals. By employing contrastive learning, it acquires distinct representations for different patient types. [24] presents a captivating framework called CROCS, which employs supervised contrastive learning. Here, cardiac signal representations linked to specific patient attributes, such as disease class, sex, and age, are drawn towards learnable embeddings known as clinical prototypes. These prototypes serve a dual purpose: facilitating the clustering and retrieval of unlabeled cardiac signals based on various patient attributes. [25] presents an exquisite, all-encompassing framework for semi-supervised learning across domains, incorporating the elegance of contrastive learning and the prowess of adversarial training strategies. [26] introduced an exquisite unsupervised contrastive learning framework. This framework ingeniously incorporates a fresh contrastive loss, harnessing the power of a data augmentation scheme that gracefully combines two data samples to create novel instances. The central objective of this framework lies in predicting the mixing component, which gracefully serves as soft targets within the loss function. [27] present an exquisite contrastive learning framework that harnesses metadata to curate positive and negative pairs during the training of unlabeled data. Specifically, the contrastive learning model ingeniously integrates metadata from clinical records, such as age, sex, weight, and sound location, to establish a profound connection between lung and heart sound recordings on a patient level, effectively enhancing the overall context.

Proposing a novel technique for remote photoplethysmography (rPPG), [28] introduced an innovative approach that eliminates the need for costly physiological training data. Their method leverages self-supervised training, combining contrastive learning with a subtle prior on the frequency and smoothness of the target signal. By integrating a learned saliency resampling module, they demonstrated the ability to reduce reliance on manually engineered features while offering insights into the model’s behavior and potential limitations.

[29]introduces an exquisite deep learning framework that effortlessly captures the essence of physiological waveforms. By employing convolutional neural networks (CNNs), it distinguishes genuine arrhythmia alarms from false ones. Through the ingenious utilization of Contrastive Learning, the framework minimizes a binary cross entropy classification loss while incorporating a remarkable similarity loss derived from pairwise waveform segment comparisons. Moreover, to enhance the models, they integrate learned embeddings from a rule-based approach, effectively harnessing prior domain knowledge specific to each alarm type.

[30] present an elegant solution called the Contrastive Accelerometer–Gyroscope Embedding (CAGE) model. This innovative approach utilizes a two-stream convolutional neural network (CNN) to process the accelerometer and gyroscope signals independently. By fusing modality-specific features at the feature-level, the model effectively tackles recognition tasks. Additionally, they introduce a self-supervised learning (SSL) task, where the accelerometer and gyroscope embeddings from the same activity instance are paired. In their work, [31] introduced the Dense Lead Contrast (DLC) approach, which utilizes contrastive learning on multilead ECGs. By employing a multibranch network (MBN), DLC fosters intralead and interlead invariance, resulting in a comprehensive global representation through a joint loss that combines intralead and interlead contrastive losses.