🤖 machine learning

An Exploratory Study of Blood Glucose Estimation from Photoplethysmography Signals using Machine Learning

This paper introduces a new paired dataset of smartwatch-derived photoplethysmography (PPG) signals and continuous glucose monitoring (CGM) values, presenting preliminary results that suggest the feasibility of using machine learning to non-invasively estimate blood glucose levels while highlighting the need for further research with larger datasets.

Original authors: Ruhani Bhatia, Vijval Ekbote

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ruhani Bhatia, Vijval Ekbote

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

Imagine your body is like a busy city, and blood sugar is the fuel that keeps the cars (your cells) running. Sometimes, the fuel level gets too low or too high, which is dangerous for people with diabetes. Right now, the best way to check this fuel level is with a Continuous Glucose Monitor (CGM), but that requires sticking a tiny needle under your skin every few days. It's accurate, but it hurts and can irritate the skin.

The researchers in this paper asked a simple question: Can we check the fuel level without the needle?

They decided to try using a smartwatch instead. Smartwatches already have a sensor that shines a light on your wrist to measure your pulse (called a PPG signal). The researchers thought, "Maybe the way your blood pulses changes when your sugar levels change, and if we use a super-smart computer brain (Machine Learning), we can figure out the sugar level just by watching the pulse."

Here is how they tried to make this work, explained simply:

1. Gathering the "Training Data"

To teach a computer to guess sugar levels, you need to show it examples where you know both the pulse and the actual sugar level at the same time.

The Team: They found 5 volunteers (3 men and 2 women).
The Gear: Each volunteer wore a smartwatch on their wrist and a CGM sensor (the needle device) on their arm for two weeks straight.
The Result: They created a massive "paired dataset." It's like having a diary where every single second of a heartbeat pulse is matched with the exact sugar level at that moment. This is rare because most previous studies only looked at data for a few minutes.

2. Cleaning Up the Messy Data

The data they collected was a bit messy, like trying to compare a high-speed video to a slow-motion photo.

The Problem: The smartwatch recorded data 64 times every second (very fast), but the CGM only gave a new sugar reading every 15 minutes (very slow).
The Fix: They used a mathematical trick called interpolation. Imagine drawing a smooth, curved line connecting the dots of the slow sugar readings. This allowed them to guess what the sugar level was at every single second, matching the speed of the smartwatch.

3. Teaching the Computer (The Experiments)

They fed this cleaned-up data into a computer model (a type of Artificial Intelligence called a Neural Network) to see if it could learn the pattern.

Experiment A: The "Same Person" Test
They trained the computer on one person's data and tested it on that same person's future data.
- The Result: It worked okay! The computer could guess the sugar level with some accuracy, but it wasn't perfect. It was like a student who studied hard for a specific teacher's test but might struggle with a different teacher.
Experiment B: The "Personalization" Test
They tried to train the computer on three people, then tweak it slightly to work for a fourth person.
- The Result: The computer got the average sugar level roughly right, but it was terrible at guessing the exact timing. It was like a weather forecaster who knows it will rain tomorrow but can't tell you if it will start at 9 AM or 2 PM. The predictions were out of sync with reality.
Experiment C: The "Time Travel" Test
They wondered, "Maybe the pulse changes before the sugar changes?" So, they tried to predict the sugar level 10, 15, or 20 minutes into the future.
- The Result: It was inconsistent. Sometimes it worked, sometimes it didn't. It depended heavily on the specific person, meaning it wasn't a reliable rule for everyone.

4. The Safety Check (The Most Important Part)

Before a medical device can be used on real people, it has to pass a strict safety test called the Clarke Error Grid. This grid checks if a wrong guess could lead to dangerous medical decisions.

The Goal: You want 100% of the guesses to be in the "Safe Zone" (Zone A).
The Reality: Only about 74% of their guesses landed in the safe zone.
The Danger: About 26% of the guesses landed in "Zone E." In this zone, the computer might say your sugar is low when it's actually high (or vice versa). If a real person used this, they might take the wrong medicine, which could be life-threatening.

The Bottom Line

The researchers successfully built a unique dataset and proved that there is a signal in the smartwatch data that relates to blood sugar. However, their current computer model is not ready for the real world.

They explicitly state that their results are not safe for clinical use yet. The technology is like a prototype car that can drive down a quiet street but isn't ready for highway traffic. They hope their dataset will help other scientists try to build a better, safer version in the future.

Technical Summary: Exploratory Study of Blood Glucose Estimation from PPG Signals

Problem Statement
Diabetes management relies heavily on Continuous Glucose Monitoring (CGM), yet traditional CGM devices are invasive, requiring skin piercing and posing risks of infection and irritation. While non-invasive alternatives are needed, existing machine learning (ML) approaches for estimating blood glucose from physiological signals often suffer from a lack of dense, long-duration paired data. Previous studies utilizing Photoplethysmography (PPG) signals have typically recorded data for very short durations (often under 3 minutes) or failed to pair PPG signals with continuous glucose measurements over extended periods, limiting the ability to train robust deep learning models.

Methodology
The authors propose a machine learning-based framework to estimate blood glucose levels using non-invasive PPG signals captured by smartwatches. The methodology is structured around the creation of a unique dataset and the application of specific signal processing and modeling techniques:

Dataset Creation: A paired dataset was constructed involving five volunteers (3 male, 2 female) monitored over a 14-day period.
- PPG Data: Collected via Empatica EmbracePlus smartwatches at a 64 Hz sampling rate.
- Glucose Data: Collected simultaneously via Freestyle Libre CGM devices (initially recorded at 15-minute intervals).
- Ethical Compliance: The study received prior permission from the Institute Review Board.
Data Preprocessing:
- Interpolation: To address the disparity between the high-frequency PPG data (64 Hz) and low-frequency glucose data (15-minute intervals), cubic spline interpolation was applied to the glucose values. This allowed for the generation of a smooth glucose curve, enabling the creation of a dense training set where every PPG timestep could be paired with a corresponding glucose value.
- Feature Extraction: Raw PPG signals were processed using a sliding window framework (60-second window, 50% overlap/30-second stride) to mitigate noise and ensure the inclusion of multiple cardiac cycles. Using the pyPPG library, 918 features were extracted per window, encompassing:
  - Time-domain features: Maximum/mean amplitude, skewness, kurtosis, energy.
  - Frequency-domain features: Dominant frequency, total power.
  - Heart Rate Variability (HRV) features: Mean inter-beat interval, standard deviation of normal-to-normal intervals, beats per minute.
Modeling: A Multi-Layered Perceptron (MLP) with two hidden layers (100 neurons each) and ReLU activation was employed. The model was optimized using the Adam optimizer. Experiments included training/testing on the same subject, personalizing models for specific subjects via fine-tuning, and predicting glucose values with time lags ( $\tau$ ) of 10, 15, and 20 minutes.

Key Results
The study presents preliminary experimental findings that highlight both potential and significant limitations:

Same-Subject Performance: When training and testing on the same subject (temporal 80-20 split), the model achieved a Mean Absolute Percentage Error (MAPE) ranging from 7.65% to 14.28% across different days, with an average MSE of 0.65.
Personalization: Fine-tuning a general model on a specific subject's data resulted in a Mean Absolute Error (MAE) of 1.0581 and an RMSE of 3.5597. However, the model exhibited a systematic underestimation (bias of -0.6850).
Synchronization Issues: Despite low RMSE values, the Lin's Concordance Correlation Coefficient (CCC) was extremely low (0.0142), indicating that the model's predictions were not effectively synchronized with the ground truth glucose trends.
Time Lag Prediction: Predicting future glucose values (10–20 minutes ahead) showed significant instability and performance variance between subjects.
Clinical Safety Assessment: Using the Clarke Error Grid, only 73.64% of predictions fell into Zone A (clinically accurate). Crucially, 25.63% of predictions fell into Zone E, representing dangerous errors that could lead to incorrect treatment. The results did not meet the ISO 15197:2013 standard for clinical approval.

Key Contributions

Dense Paired Dataset: The primary contribution is the creation and release of a dense, long-duration (14-day) paired dataset linking high-frequency PPG signals from smartwatches with continuous glucose measurements from CGM devices. This addresses a critical data scarcity in the field.
Exploratory Analysis: The paper provides a baseline evaluation of using MLPs for this task, demonstrating that while predictive signals exist, current models struggle with synchronization and clinical accuracy.
Methodological Framework: The study details a pipeline for handling multi-modal data with different sampling rates, including interpolation strategies and feature engineering for PPG signals.

Significance and Claims
The authors position this work as an exploratory study rather than a definitive solution. They explicitly state that while the results suggest the existence of predictive signals within PPG data, the current models are not yet suitable for clinical deployment. The paper emphasizes that the low CCC and high percentage of dangerous errors (Zone E) necessitate extensive further validation and clinical testing before any claims regarding accuracy or accessibility can be made. The significance of the work lies in providing a foundational resource (the dataset) and preliminary insights to inspire further research into non-invasive glucose monitoring, rather than offering a ready-to-use medical device.