A Review of Deep Learning Methods for Photoplethysmography Data

This paper reviews 460 studies from 2017 to 2025 on deep learning applications for photoplethysmography (PPG) data, highlighting its superior performance over traditional methods in diverse tasks while identifying key challenges such as data scarcity, real-world validation, and model interpretability.

The Gist

Imagine your body is a bustling city, and your heart is the central power plant pumping energy (blood) through a vast network of roads (arteries) to every neighborhood. Photoplethysmography (PPG) is like a tiny, non-invasive camera that sits on your skin (usually your finger or wrist) and takes a "traffic report" of this blood flow. It shines a light through your skin and measures how much gets absorbed or reflected by the moving blood.

For a long time, doctors and engineers looked at these traffic reports using manual maps and simple rules. But recently, they've started using Deep Learning—a type of artificial intelligence that acts like a super-smart, tireless detective who can spot patterns in the traffic data that humans would miss.

This paper is a massive review of 460 different studies published between 2017 and 2025. It's like a "State of the Union" address for how AI is currently using these heart-beat cameras to solve all sorts of health mysteries.

Here is the breakdown of what the paper found, explained simply:

1. What is the AI actually doing? (The Tasks)

The paper categorizes what these AI detectives are trying to solve. It's not just about counting heartbeats anymore.

• The Vital Signs (Blood Pressure & Glucose): Imagine trying to guess the water pressure in a pipe without a gauge. The AI looks at the shape of the blood wave to guess your blood pressure or even your blood sugar levels, potentially replacing the need for painful finger pricks or tight cuffs.
• The Sleep Detective: Just as a car engine makes different noises when idling vs. racing, your blood flow changes when you are in deep sleep vs. dreaming. The AI uses these changes to tell you if you have sleep apnea (stopping breathing) or to map out your sleep stages without you wearing a heavy, uncomfortable hospital mask.
• The Mood Reader: Stress and emotions change how your blood vessels tighten and relax. The AI can detect if you are stressed, anxious, or calm just by looking at your pulse.
• The Body Translator: One of the coolest tricks is ECG Reconstruction. The AI can look at the PPG signal (blood flow) and "translate" it into an ECG signal (electrical heart activity). It's like looking at the ripples in a pond and perfectly reconstructing the shape of the stone that caused them. This means your smartwatch could potentially diagnose heart attacks without needing the sticky electrodes of a traditional ECG.
• The ID Scanner: Because everyone's blood vessels are slightly different (like a fingerprint), the AI can use your pulse to verify who you are, acting as a biometric security key.

2. How does the AI think? (The Models)

The paper looks at the "brains" behind the AI. Think of these as different types of tools in a toolbox:

• CNNs (The Pattern Matchers): These are the most popular tools. They are great at spotting local shapes in the wave, like a specific bump or dip in the blood flow. They are like a magnifying glass looking for specific details.
• RNNs & LSTMs (The Storytellers): These tools are good at understanding sequences. They look at the history of the signal to understand how the heart rate is changing over time, not just at one single moment.
• Transformers (The Big Picture Thinkers): These are the newest, most powerful tools (like the ones behind ChatGPT). They can look at the entire signal at once and understand long-term connections, but they require a lot of data and computing power to work well.
• Generative Models (The Forgers): These are special AI that can create new, fake heart signals that look real. Scientists use them to "fill in the blanks" when a signal is noisy or to create more training data for other AIs.

3. The Fuel: Data

AI is only as good as the food it eats. The paper highlights that we need more and better data.

• The Problem: Most of the data we have comes from hospitals (ICUs) or controlled sleep labs. It's like trying to learn how to drive a car only by practicing in a parking lot. We need more data from real life—people running, dancing, and sleeping in their own homes.
• The Challenge: Everyone is different. A signal from a young athlete looks different from an elderly person, and a signal from someone with dark skin looks different from someone with light skin due to how light interacts with skin. The AI needs to learn to handle all these variations, or it might make mistakes.

4. The Hurdles (Why isn't this perfect yet?)

Even though the AI is amazing, the paper points out some growing pains:

• The "Black Box" Problem: We know the AI gives the right answer, but sometimes we don't know why. Doctors need to trust the AI, so they need to understand the reasoning, not just the result.
• The Noise Factor: In the real world, your watch moves, your skin sweats, and the light changes. This creates "noise" that confuses the AI. We need better ways to clean up the signal before the AI looks at it.
• Privacy: Since your pulse is unique, it's like a fingerprint. If someone steals your heart data, they could potentially identify you. We need strong rules to protect this sensitive information.

The Big Picture

This review is essentially saying: "We have built a incredibly powerful engine (Deep Learning) that can read our heart's story in ways we never imagined. It can predict diseases, monitor sleep, and even translate between different types of heart signals."

However, to make this engine safe and reliable for everyone, we need to feed it more diverse data from real life, teach it to explain its answers, and ensure it works equally well for people of all ages, skin tones, and backgrounds. The future of health monitoring isn't just about wearing a device; it's about having a smart, invisible assistant that understands your body's language better than you do.

Technical Summary

1. Problem Statement

Photoplethysmography (PPG) is a ubiquitous, non-invasive optical sensing technique used to capture hemodynamic information in both clinical settings and consumer wearables (e.g., smartwatches). While PPG data volume has exploded due to the proliferation of wearable devices, traditional analysis methods face significant bottlenecks:

• Limitations of Traditional ML: Conventional approaches rely heavily on manual feature engineering (extracting morphological, temporal, or frequency-domain features). This process is time-consuming, labor-intensive, dependent on expert knowledge, and often fails to capture the full complexity and non-linear patterns within PPG signals.
• Data Complexity: The growing scale and noise (motion artifacts, sensor variability) of PPG data exceed the capacity of manual inspection and conventional statistical methods.
• Gap in Literature: While previous reviews covered specific applications (e.g., atrial fibrillation detection, blood pressure estimation) or general PPG applications, there was a lack of a comprehensive overview specifically focusing on deep learning (DL) methodologies across the entire spectrum of PPG analysis tasks.

2. Methodology

The authors conducted a systematic literature review following the PRISMA guidelines:

• Search Strategy: A comprehensive search was performed across Google Scholar, PubMed, and Dimensions for papers published between January 1, 2017, and December 31, 2025.
• Query: ("deep learning" OR "DL") AND ("photoplethysmography" OR "PPG").
• Selection Criteria:
  • Inclusion: English language, focus on contact PPG data, application of deep learning methods, and quantitative evaluation.
  • Exclusion: Non-contact PPG, non-DL methods, lack of quantitative metrics, or non-English papers.
• Final Dataset: From an initial pool of 1,168 papers, 460 studies were selected for detailed analysis.
• Analysis Framework: The included studies were categorized and analyzed from three key perspectives:
  1. Tasks: Application domains and specific physiological parameters.
  2. Models: Deep learning architectures and generative vs. discriminative approaches.
  3. Data: Sources of datasets, collection devices, and population characteristics.

3. Key Contributions & Findings

A. Task Landscape (Applications)

The review identified 11 major task categories, demonstrating that DL has expanded PPG utility far beyond basic heart rate monitoring:

• Blood Pressure (BP) Analysis (N=154): The largest category. Tasks range from binary hypertension classification to continuous cuffless BP estimation and arterial blood pressure (ABP) waveform reconstruction using U-Nets and Diffusion models.
• Cardiovascular Monitoring (N=106): Includes arrhythmia detection (AF, PVC), myocardial infarction, heart failure, and risk stratification (e.g., 10-year CVD risk).
• Sleep Health (N=27): Sleep staging and sleep apnea detection, often utilizing transfer learning from ECG to PPG.
• Mental Health (N=31): Stress detection and emotion recognition (e.g., baseline vs. stress vs. amusement).
• Respiratory & Oxygenation (N=26): Respiratory rate (RR) estimation and SpO2 monitoring.
• Blood Glucose (N=31): Non-invasive glucose estimation, often fusing PPG with kinetic features or historical data.
• Signal Processing (N=41): Denoising, artifact removal, and signal quality assessment (SQA).
• Biometric Identification (N=27): User authentication based on unique PPG waveform morphology.
• ECG Reconstruction (N=16): Synthesizing ECG signals from PPG using GANs and Diffusion Transformers.
• Human Activity Recognition (N=7) and Others (sepsis, pain levels, anesthesia depth).

B. Model Architectures

The review analyzed the prevalence and suitability of various DL architectures:

• Convolutional Neural Networks (CNNs): The dominant architecture (283 papers). CNNs are favored for their ability to capture local morphological patterns (e.g., systolic peaks, dicrotic notches) efficiently.
• CRNNs (CNN + RNN): Widely used (96 papers) to combine CNNs' local feature extraction with RNNs' (LSTM/GRU) temporal modeling capabilities.
• RNNs (89 papers): Used for sequential dependencies but less common as standalone models due to training instability with long sequences and high computational costs.
• Transformers (38 papers): Emerging for modeling long-range dependencies. However, adoption is limited by the need for large-scale datasets and high computational complexity (quadratic scaling with sequence length).
• Generative Models:
  • Autoencoders (AE): Common for denoising and reconstruction.
  • GANs: Used for signal synthesis and augmentation, though adoption is limited by training instability.
  • Diffusion Models: An emerging trend (3 papers) showing promise for high-fidelity waveform generation (e.g., ECG from PPG).
  • U-Net: Frequently used as a backbone for signal reconstruction tasks.

C. Data Resources

The authors highlighted a critical bottleneck: limited availability of large-scale, high-quality, standardized datasets.

• Key Datasets: MIMIC-series (ICU data), UK Biobank (large-scale cohort), VitalDB (surgical), and PulseDB (cuffless BP benchmark).
• Challenges: Most datasets are clinical (ICU) rather than real-world daily life. There is significant heterogeneity in devices, sampling rates, and preprocessing pipelines, hindering model generalization.

4. Results & Discussion

• Performance: Deep learning methods generally outperform traditional machine learning approaches by automatically learning informative representations from raw signals, reducing reliance on handcrafted features.
• Emerging Trends:
  • Foundation Models: The rise of pre-trained models (e.g., AnyPPG, Pulse-PPG) trained on massive datasets to support diverse downstream tasks via fine-tuning.
  • Multimodal Integration: Combining PPG with ECG, accelerometer, and clinical text (via LLMs) for holistic health profiling.
  • Personalization: Addressing inter-individual variability through subject-specific adaptation and transfer learning.
• Critical Challenges:
  • Data Scarcity & Bias: Lack of diverse, large-scale datasets leads to poor generalization across skin tones, ages, and devices.
  • Real-world Validation: Many models perform well in controlled clinical settings but degrade in noisy, real-world environments (motion artifacts).
  • Interpretability: "Black box" nature of DL models hinders clinical trust; there is a need for physiologically explainable AI.
  • Privacy: PPG signals can act as biometric identifiers, raising data security and re-identification risks.

5. Significance

This review serves as a definitive reference for researchers and clinicians in the field of digital health. Its significance lies in:

1. Comprehensive Mapping: It provides the first holistic taxonomy of DL applications in PPG, moving beyond isolated tasks to a unified view of the field.
2. Guidance for Future Research: It identifies critical gaps, specifically the need for standardized preprocessing protocols, large-scale real-world datasets, and interpretable models.
3. Clinical Translation: By highlighting the transition from research to wearable deployment, it underscores the necessity of robust validation in daily-life scenarios to enable reliable, continuous, and personalized healthcare monitoring.
4. Technological Roadmap: It charts the evolution from simple CNNs to complex foundation models and generative AI, suggesting a future where PPG-based systems can provide comprehensive, multi-system health assessments.