⚡ electrical engineering

Non-invasive Blood Glucose Estimation from Wearable Physiological Signals

This paper proposes a deep-learning-based dynamic incremental learning framework that combines mutual entropy-optimized replay and uncertainty-quantified proxy gradient bridging to achieve accurate, label-efficient non-invasive blood glucose estimation from wearable PPG signals, supported by a new longitudinal benchmark dataset of 183 participants collected over 285 days.

Original authors: Zexing Zhang, Jichao Li, Yilong Wang, Kewei Yang, Tianyang Lei

Published 2026-07-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Zexing Zhang, Jichao Li, Yilong Wang, Kewei Yang, Tianyang Lei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A "Smart Watch" That Learns Without Needles

Imagine you want to check your blood sugar (glucose) every day. Currently, the best way involves pricking your finger with a needle or wearing a sensor under your skin that needs replacing. Both are painful, annoying, and expensive.

This paper proposes a new way: using a simple, non-invasive wristband (like a smartwatch) that shines a light through your skin to read your blood sugar. The problem is that the human body is messy. Your blood sugar changes based on what you ate, how you slept, or if you are stressed. A computer program trained on "yesterday's you" often fails to understand "today's you."

The authors built a system that acts like a super-adaptive student. Instead of memorizing a textbook and then forgetting it when the test changes, this system learns continuously, adapts to your unique body changes, and even learns from data it doesn't fully understand yet.

The Three Big Problems They Solved

1. The "Changing Student" Problem (Dynamic Incremental Learning)

The Issue: Imagine a student who studies hard for a math test in January. By June, the math curriculum has changed, and the student's brain has changed too. If you force the student to relearn everything from scratch every time a new topic appears, they get tired and forget the old stuff (this is called "catastrophic forgetting").
The Solution: The authors created a Dynamic Incremental Learning (DIL) system. Think of this as a student who keeps a "memory notebook." When new information arrives, the student doesn't throw away the old notebook. Instead, they add new pages while carefully protecting the old notes so they aren't erased. This allows the system to adapt to your changing body over months without needing a massive computer to retrain from scratch every day.

2. The "Confusing Crowd" Problem (Dynamic Clustering)

The Issue: Your body isn't just one thing; it's a mix of many different states. Sometimes you are calm, sometimes you just ate a big meal, sometimes you are exercising. If you treat all these moments as the same "type" of data, the computer gets confused.
The Solution: The team built a Smart Organizer (MERDC). Imagine a librarian who doesn't just sort books by title, but by the vibe of the story. This system looks at your heart rate signals and automatically groups them into "clumps" or "tasks" (e.g., "Post-Breakfast State," "Resting State," "Stress State"). It figures out these groups on its own without being told what they are. This helps the system know exactly which "rulebook" to use for the current moment.

3. The "Missing Answers" Problem (Proxy Gradient Bridging)

The Issue: To teach a computer, you usually need the "answer key" (the actual blood sugar number). But getting that number requires a finger prick, which is painful and rare. You have thousands of hours of heart rate data (the question) but almost no blood sugar numbers (the answer).
The Solution: They invented a Bridge Agent (PGBA). Imagine a teacher who has a few students with answer keys and thousands of students without them. Instead of guessing the answers for the students without keys, the teacher uses the few students who do have keys to learn a "pattern of thinking." Then, the teacher uses that pattern to guide the other students.
In technical terms, the system uses the few painful finger-prick measurements to learn how to "guess" the right direction for the thousands of painless measurements, allowing it to learn efficiently without needing a needle for every single data point.

The "Gym" They Built (The Dataset)

To prove this works, the researchers didn't just use a small, fake dataset. They built a massive Longitudinal Benchmark.

The Participants: 183 people (a mix of ages and health conditions, including diabetics).
The Duration: They tracked these people for 285 days (almost a year).
The Data: Every time they visited, they recorded:
1. PPG: The light signal from the wrist (the "question").
2. Blood Glucose: The finger-prick number (the "answer").
3. Blood Pressure: Extra health data.
The Result: This is one of the largest and longest-running datasets of its kind, designed specifically to test if a system can handle the messy, long-term reality of human biology.

The Results: How Well Did It Work?

The system was tested on people it had never seen before (a fair test).

Accuracy: The average error was 0.64 mmol/L. To put that in perspective, if the real sugar level is 5.0, the guess is usually between 4.36 and 5.64.
Safety: In the medical world, we use a "Clarke Error Grid" to see if a wrong guess is dangerous.
- Zone A: The guess is perfect or close enough to be safe.
- Zone B: The guess is a bit off but still safe enough to make a decision.
- Result: 97.7% of the guesses fell into the safe zones (A and B). This means the system is highly unlikely to give a dangerous reading.

What the Paper Does Not Claim

It is important to stick to what the paper actually says:

It is not a medical device yet. The authors explicitly state this is a "proof-of-concept" and a "benchmark." It is a research tool, not a product you can buy at a pharmacy today.
It is not a regulatory approval. The paper clarifies that while the results look good, they have not gone through the strict government testing required to sell a medical device.
It is not perfect for everyone. The study was done in one location in China. The authors admit the system might need more testing on different ethnic groups or with different types of watches before it works for everyone globally.

The Bottom Line

This paper presents a new way to teach computers to read blood sugar from a wristband. By using a system that learns continuously, organizes body states automatically, and learns from painless data, they achieved high accuracy on a massive, year-long dataset. It's a major step toward a future where you can check your blood sugar without a needle, but it is still in the research phase.

Technical Summary: Dynamic Incremental Learning for Non-invasive Blood Glucose Estimation

1. Problem Statement

Non-invasive blood glucose (BG) estimation using wearable photoplethysmography (PPG) signals faces three critical challenges that hinder real-world deployment:

Non-Stationarity and Distribution Drift: Intra-individual physiological regulation (driven by diet, activity, and pharmacology) and inter-individual/environmental variations cause continuous, multi-scale distribution shifts in PPG signals. Conventional deep learning models, trained on static datasets, suffer from catastrophic forgetting when exposed to these evolving streams and lack the stability-plasticity balance required for continual adaptation.
Label Scarcity and Cost: Reference capillary blood glucose labels are sparse, invasive, and costly to acquire, whereas unlabeled PPG data streams are abundant. Existing methods often fail to leverage this unlabeled data effectively without introducing bias or relying on pseudo-labels.
Fixed Distribution Assumptions: Most state-of-the-art (SOTA) methods assume fixed data distributions or predefined task categories (e.g., Class-Incremental Learning), which do not align with the continuous, regression-oriented, and latent task structures found in longitudinal physiological monitoring.

2. Methodology

The authors propose a Dynamic Incremental Learning (DIL) framework designed specifically for label-efficient adaptation to unlabeled PPG streams under distribution shift. The framework integrates three core components:

A. Mutual Entropy-Optimized Replay-Based Dynamic Clustering (MERDC)

To address the stability-plasticity dilemma and the lack of predefined task boundaries, the authors introduce a task discovery mechanism:

ME2AC Operator: A novel unsupervised clustering operator (Mutual Entropy-optimized Enhanced Adaptive Clustering) that partitions streaming PPG data into latent physiological tasks. Unlike standard DBSCAN, ME2AC uses:
- Local Signal Entropy: To detect core points based on the entropy of standardized signal values within adaptive neighborhoods.
- Histogram-Based Mutual Information: To merge preliminary clusters, ensuring that merged groups share significant morphological similarity.
Replay Strategy: Instead of storing data by glucose classes, the framework allocates a bounded episodic memory budget across the discovered latent tasks. This ensures that the model retains diverse feature representations from previous physiological regimes without discretizing the continuous regression target.

B. Proxy Gradient Bridging Agent (PGBA)

To leverage abundant unlabeled PPG data without assigning pseudo-labels, the authors propose PGBA:

Mechanism: PGBA is a learnable agent that maps the model's current output (prediction and uncertainty) to a proxy gradient in the output space.
Training: The agent is trained on a small set of labeled data using a "bridge loss," which minimizes the supervised loss after a hypothetical proxy update step.
Application: For unlabeled windows, the agent predicts a proxy gradient that serves as an update direction. This allows the model to learn from unlabeled streams by optimizing the direction of the update rather than the target value itself, bounded by a scaling factor to prevent large, biased updates.

C. Dynamic Incremental Learning (DIL) Backbone

Architecture: The backbone utilizes an InceptionTime network for feature extraction, followed by two linear heads: one for the mean blood glucose prediction ( $\hat{y}$ ) and one for the input-dependent variance ( $\hat{\sigma}^2$ ) to quantify uncertainty.
Optimization: The model minimizes regression loss on the current task while enforcing gradient projection constraints. Updates are projected onto the feasible space where the loss on past task memories (replay buffers) does not increase, thereby mitigating catastrophic forgetting.

3. Benchmark Dataset

The authors established a new longitudinal benchmark dataset to support this research:

Scale: 183 participants (mean age 57.02 ± 21.20 years) collected over 285 days.
Data: Synchronized PPG signals (1 kHz), reference capillary blood glucose (Accu-Chek Performa), and cuff blood pressure.
Scope: Includes 5,224 validated 4-second PPG windows, covering a glucose range of 4.2–16.3 mmol/L. The dataset captures both inter- and intra-participant variability, including diabetic and non-diabetic subjects with various comorbidities.
Protocol: A rigorous subject-independent 5-fold cross-validation (5CV) protocol was used, ensuring no data leakage between training and testing folds during task discovery, memory construction, or model training.

4. Key Results

Under subject-independent 5-fold cross-validation, the proposed method achieved:

Accuracy: Mean Absolute Error (MAE) of 0.64 ± 0.01 mmol/L and Root Mean Square Error (RMSE) of 1.29 ± 0.10 mmol/L.
Clinical Relevance: 97.69 ± 1.63% of estimates fell within Clarke Error Grid Zones A+B (clinically acceptable regions).
Comparison: The method outperformed reproduced SOTA baselines (including traditional ML, deep learning, and self-supervised approaches) by reducing MAE by 23%–86%.
Robustness: Aggregation-level analyses (averaging predictions over 40-second acquisitions and across subjects) confirmed that the favorable error distribution persists beyond window-level statistics.
Ablation: Removing any of the three core components (DIL, PGBA, or MERDC) resulted in significant performance degradation, confirming their complementary roles.

5. Significance and Claims

The paper positions its contributions as a proof-of-concept for adaptive, label-efficient non-invasive glucose estimation in wearable sensing. Key claims include:

Paradigm Shift: It introduces Dynamic Incremental Learning to this field, moving away from static training distributions toward continual adaptation to physiological drift.
Label Efficiency: The PGBA mechanism demonstrates that unlabeled PPG streams can be effectively utilized to improve generalization without the high cost of invasive labeling.
Task Discovery: The MERDC module provides a principled way to discover latent physiological tasks in regression problems without predefined categories, enabling effective replay-based continual learning.
Benchmarking: The release of a large-scale, longitudinal benchmark dataset addresses the scarcity of high-quality, diverse data in this domain, facilitating reproducible research.

Limitations and Modesty:
The authors explicitly state that these results represent a research-stage proof-of-concept. They acknowledge limitations including the single-center cohort (potentially limiting demographic generalizability), the use of capillary rather than venous plasma glucose as the reference, and the retrospective nature of the ISO 15197:2013 analysis. They emphasize that formal regulatory compliance and broader multi-center prospective validation are required before clinical deployment. The framework is designed to support personalized longitudinal monitoring but requires further integration with signal-quality gating and complementary modalities (e.g., ECG, accelerometry) for robust real-world performance.