TL;DR
This systematic review explores the growing body of research using photoplethysmography (PPG) and electrocardiography (ECG) signals for diabetes detection, glucose estimation, and complication management. It highlights the potential of these non-invasive methods, particularly with the rise of machine and deep learning, while emphasizing the critical need for standardization and clinical validation.
What Problem Does This Paper Address?
Diabetes mellitus (DM) is a chronic metabolic disease with significant health and economic burdens, often leading to severe cardiovascular complications. Traditional diagnostic methods are invasive and can be costly. This paper addresses the need for accessible, non-invasive, and continuous monitoring solutions for diabetes care. It specifically investigates how readily available biosignals like PPG (from devices like smartwatches) and ECG can be leveraged to detect diabetes, estimate blood glucose levels, and identify complications, bridging the gap between current diagnostic practices and future wearable health technologies.
Methods
This paper presents a systematic review of existing literature, adhering to PRISMA guidelines. The authors searched major databases (PubMed, Google Scholar, IEEE Xplore) for studies utilizing PPG and/or ECG signals for diabetes detection, glucose estimation, or complication management. They categorized the reviewed studies based on the analytical approach:
- Traditional Methods: Statistical analysis and mathematical models applied to extracted features.
- Machine Learning (ML): Algorithms like Support Vector Machines (SVM), Random Forests, and Decision Trees, which require feature extraction.
- Deep Learning (DL): Neural networks, particularly Convolutional Neural Networks (CNNs), capable of learning directly from raw signals, often bypassing manual feature engineering.
The review analyzes the advantages and disadvantages of each approach, focusing on data types, features used, and main outcomes.
Datasets
The paper does not introduce new datasets. Instead, it reviews studies that utilized various datasets, often described as "in-house" or "public." Specific public datasets mentioned include [85] and [118]. The authors note that a significant limitation across many studies is the lack of standardized or publicly available datasets, hindering reproducibility and direct comparison.
Key Insights and Ideas
- PPG and ECG are promising: These non-invasive signals offer a wealth of information about cardiovascular health, which is closely linked to diabetes. Their ease of use and integration into wearable devices make them ideal for continuous monitoring.
- Machine and Deep Learning are game-changers: While traditional methods provided initial insights, ML and DL have significantly advanced the accuracy and efficiency of diabetes-related signal analysis, especially by automating feature extraction.
- Feature extraction is still key (mostly): Despite DL's ability to learn from raw data, many successful approaches, particularly for glucose estimation and complication detection, still rely on well-engineered features, with Heart Rate Variability (HRV) being a prominent example.
- Diabetes detection is the most explored area: The majority of reviewed studies focused on detecting diabetes itself, followed by blood glucose estimation. Diabetes complication detection is a less explored but crucial area.
- Clinical validation is a major bottleneck: While many studies show promising results in controlled environments, a significant gap exists in translating these findings into clinically validated tools for real-world patient care.
Key Results
The review included 78 studies, categorized as follows:
- Diabetes Detection: 31 studies (10 traditional, 15 ML, 6 DL)
- Blood Glucose Estimation: 41 studies (5 traditional, 21 ML, 15 DL)
- Diabetes Complications Detection: 7 studies (5 traditional, 1 ML, 1 DL)
While specific quantitative results vary widely across studies due to differing methodologies and datasets, the review indicates that ML and DL approaches generally achieve higher accuracy, sensitivity, and specificity compared to traditional methods. For instance, some ML models for diabetes detection reported accuracies exceeding 90%, and DL models for blood glucose estimation showed promising results with low RMSE and high accuracy in Clarke Error Grid analysis.
Limitations and Future Work
The authors identify several critical limitations and areas for future research:
- Data Processing Standardization: A lack of standardized signal processing and feature extraction methods makes it difficult to compare results across studies.
- Reproducibility: Insufficient detail in reporting methodologies and dataset characteristics hinders the reproducibility of findings.
- Dataset Splitting: A common issue is the improper separation of training, validation, and testing sets, potentially leading to biased results. Future work should ensure rigorous subject-level splitting.
- Performance Analysis Standardization: The wide variety of performance metrics used makes direct comparison challenging. Guidelines for performance analysis are needed.
- Clinical Validation: Crucially, none of the reviewed methods have been clinically validated. This is a critical step for their adoption in clinical practice.
- Interpretability: Deep learning models, while powerful, often lack interpretability, which is essential for clinical trust and decision-making.
- Diabetes Complications: Further research is needed to explore PPG and ECG for detecting and managing diabetes complications.
Summary updated