This review paper systematically covers the transformative impact of deep learning on PPG data analysis and its diverse applications. The combination of PPG signals and deep learning enables higher accuracy and automated analysis across medical and non-medical domains compared to conventional techniques, yet clearly reveals remaining challenges such as data limitations, real-world validation, and interpretability. The paper provides an in-depth discussion of key tasks, models, data status, as well as barriers to real-world deployment and future directions in deep learning research on PPG data.

1. Introduction: The Meeting of PPG and Deep Learning

In recent years, photoplethysmography (PPG) sensors have been widely adopted in wearable devices and various medical equipment, thanks to their ability to non-invasively and continuously measure heart rate, arterial blood pressure, oxygen saturation, respiratory rate, and more. For example, PPG devices of various form factors -- Apple Watch, smart rings, scarves -- have been developed, with growing numbers of users.

As PPG data proliferates, manual interpretation by experts has reached its limits, creating a need for automated and accurate analysis methods. While conventional machine learning approaches require manual feature selection, advances in deep learning enable automatic extraction of meaningful patterns from large datasets and "end-to-end" analysis.

"Researchers simply input the data, and the model automatically and accurately handles everything needed."

This paper broadly organizes the latest trends in deep learning-based PPG data analysis from the perspectives of tasks, models, and data.

2. Background

2.1 PPG Principles and Devices

The core principle of PPG is that light absorption and reflection change according to blood pulsation through skin tissue. It typically consists of an LED (light source) and a photodiode (detector), with two modes:

• Transmittance: light passes through to the opposite side
• Reflectance: light is reflected back on the same side

Different wavelengths (light colors) offer varying resistance to motion artifacts and measurement efficiency, making wavelength selection matched to the application and device characteristics important.

Figure 1. Overview of PPG analysis by task, model, and data

2.2 Deep Learning Concepts

Deep learning uses neural network architectures with multiple hidden layers to automate feature extraction and prediction from large datasets. Unlike traditional methods (e.g., logistic regression, random forests, SVMs):

• CNN (Convolutional Neural Network): Excels at extracting local patterns/features from image and time-series data
• RNN (Recurrent Neural Network), LSTM/GRU: Suited for capturing temporal context
• CRNN: Combines the strengths of CNN and RNN
• Transformer: Attention-based, excellent at processing long-range context
• AE, GAN, U-Net and other generative models: Data generation and reconstruction

Different architectures serve different purposes.

3. Methods

The researchers searched Google Scholar, PubMed, Dimensions, and other databases from 2017 to July 2023 using ("deep learning" OR "DL") AND ("photoplethysmography" OR "PPG"), finding a total of 646 papers. After filtering for duplicates, non-English papers, non-PPG-related content, non-use of deep learning, and lack of quantitative evaluation, 193 papers were selected for analysis.

Figure 2. Literature search and selection flowchart

4. Key Findings

The 193 papers were analyzed across tasks, models, and data.

4.1 Task Classification

4.1.1 Medical Applications

• Blood pressure analysis: Deep learning + PPG for hypertension classification and prediction. For example, the "DeepCNAP" model predicts SBP, DBP, and MAP with high precision using PPG.

  "DeepCNAP, combining ResUNet and self-attention, extracted continuous arterial blood pressure waveforms with mean absolute errors of 3.40 +/- 4.36 mmHg (systolic) and 1.75 +/- 2.25 mmHg (diastolic)."

• Cardiovascular disease diagnosis: Arrhythmia classification including atrial fibrillation (AF), heart rate computation, vascular aging

  "AF detection is approached as 2- or 3-class classification, with deep learning enabling more stable heart rate extraction during exercise."

• Sleep health: Sleep stage classification, obstructive sleep apnea (OSA) detection using PPG + deep learning

  "A CNN-GRU-based CRNN achieved a best epoch accuracy of 80.1% and Cohen's kappa of 0.65 for 3-5 stage sleep classification."

• Mental health/emotional state: Deep learning classification of stress, anxiety, and other emotional states via PPG changes
• Respiratory analysis: Respiratory rate prediction, respiratory signal reconstruction

  "ResNet-based deep learning predicted respiratory rate with a mean error of 2.5 +/- 0.6 breaths/min."

• Blood glucose analysis/diabetes: Non-invasive blood glucose prediction and diabetes detection
• Others: Autonomic neuropathy, pain classification, sepsis, COVID-19 diagnosis, anesthesia depth prediction

4.1.2 Non-Medical Applications

• Signal processing: Noise removal, signal quality assessment, PPG segmentation and localization

  "CNN-based classification of PPG signal quality as 'good'/'bad' achieved up to 93.8% accuracy on 5-second segments."

• Biometric authentication: PPG-based individual identification using CNN, LSTM, and 2D-transformed images

  "CNN-based biometric authentication achieved 97-99% accuracy on VitalDB and other datasets."

• ECG reconstruction: PPG-to-ECG conversion, especially for non-contact cardiac monitoring
• Activity recognition: Separating cardiac, respiratory, and motion components within PPG signals; human activity classification

  "Successfully classified 5 activities (standing/walking/running/jumping/sitting) using PPG alone."

4.2 Model Trends

• Most frequently used: CNN (116 studies) > CRNN (44) > RNN (32)
• Generation/reconstruction: U-Net (15), GAN (6)
  • U-Net: Effective for information preservation and time-series generation
  • GAN: Challenges with instability and mode collapse

"CNNs demonstrate powerful pattern extraction in time-series data (not just images) and are well-suited for parallel processing."

4.3 Data Landscape

• Widespread use of open-source data: Notably MIMIC-III, UCI_BP, PPG-BP, PPG_DaLiA
  • Characteristics, populations, and links for each database are summarized (see Table 3)

5. In-Depth Discussion

5.1 Significance and Expandability of Deep Learning + PPG

Thanks to deep learning, higher-dimensional information that was previously inaccessible through traditional theory (e.g., heart rate measurement) can now be explored, expanding PPG into

• diverse neural-network-based novel applications such as biometric authentication, emotion recognition, and ECG reconstruction.

"Before deep learning, researchers relied on manual feature extraction or theoretical time-series characteristics. Now, latent information in high-dimensional spaces can be automatically discovered."

5.2 Diverse PPG Device Design and Real-World Application

• Scenario-specific custom device design is needed for smartwatches, rings, bands, etc.
• Integration optimization of deep learning with hardware, and clinical/user real-world validation remain key challenges.

  "Portability and stability of devices for specific applications must be considered, and further research on integration with deep learning is needed."

5.3 The Importance of Signal Preprocessing

• Segmentation, resampling, noise removal, low-quality signal rejection, and data augmentation significantly impact performance.

  "Segment-based imbalances arising from signal length differences across users can introduce bias in training." This deserves careful attention.

5.4 Pros and Cons of Multimodal Fusion

• Combining PPG with ECG, ACC, and other physiological signals can enhance information, but
• Practical limitations exist in alignment, computational burden, and feasibility.

5.5 Considerations in Model Design

1. Architecture selection: Choose among CNN, RNN, attention, deep/shallow structures based on data/task characteristics and complexity
2. Domain knowledge integration: Reflect expert judgment in design to enhance interpretability and practicality
3. Personalization: Adapt to individual user data
4. Interpretability: Build trust through transparency and explainability
5. Performance balance: Trade-off strategies between complexity-efficiency and interpretability-performance are critical

"What is needed is not just high accuracy, but a balanced design suited to the real-world application environment -- combining interpretability, effectiveness, and efficiency."

5.6 Leveraging Diverse Learning Strategies

• Combining supervised, semi-supervised, and unsupervised learning to better utilize label-scarce data and strengthen generalization.

  "Effective utilization of large-scale real-world unlabeled data is an important research topic."

5.7 The Importance of Validation

• Consistent internal/external validation and benchmarking are essential for reliability.

  "Theory alone has limits. Reproducibility validation across diverse scenarios and datasets is critical."

5.8 Barriers to Data Acquisition and Standardization

• Insufficient databases, individual (demographic) and device/measurement-site variability, and lack of standardization make developing universal, scalable models difficult.

  "These constraints make it difficult to extend research and conduct objective comparisons."

5.9 Data Security and Privacy

• Anonymization, access management, and encryption are required for legal/ethical compliance. PPG deep learning in particular learns individual-specific patterns, raising re-identification concerns.

5.10 Integration with Large Language Models (LLMs)

• Combining medical records and other text with PPG analysis can enable precision medicine.
• Using LLMs as chatbots can enhance natural user interaction.

  "Leveraging LLMs enables personalized health management and diagnostic support through the integration of PPG data and textual information."

6. Conclusion

Deep learning-based PPG data analysis has been confirmed to surpass conventional methods in accuracy, automation, and application breadth. However, data volume and quality, real-world evaluation, and model interpretability and scalability remain open challenges. Advancing this field requires more comprehensive and standardized data, multifaceted validation, and accumulation of real-world deployment cases.

Glossary

Closing Thoughts

The combination of deep learning and PPG is revolutionizing wearable healthcare and diverse biosignal-based applications. Continued progress on data quality improvement, model interpretability, and real-world deployment validation is expected, along with advances toward personalized smart healthcare.