All research

Physiological signal processing · 2024

QU-GM: IoT-based glucose monitoring from wrist PPG, blood pressure and demographic data

A wearable and cloud-connected system that uses machine learning to estimate blood glucose and classify glucose severity without a finger-prick input.

Qatar National Research Fund NPRP13S-0205-200270 · Qatar Diabetic Society

Two-part diagram. The conventional method uses a lancet prick, a blood sample and a glucometer. The proposed method sends wrist PPG from a sensor and TinyPICO board over Bluetooth to an Android application, which exchanges data with machine learning models on Amazon Web Services.
Figure 1 from Shuzan et al. (2024). (a) Conventional finger-prick measurement and (b) the proposed QU-GM system.View full-size figureSource articleCC BY 4.0

Approach

Can wrist PPG, blood pressure and demographic information support real-time glucose estimation without a blood sample?

Finger-prick meters require repeated blood samples, and continuous glucose monitors use a sensor placed under the skin. QU-GM explores a less invasive route by combining a custom PPG wristband with blood pressure, demographic information and machine learning. The complete system links the wearable, an Android application and cloud-hosted models for glucose estimation and severity classification.

  1. Record wrist PPG and user data
  2. Estimate glucose and severity
  3. Return results to the application

Building the wristband and connected system

Hardware, mobile application and cloud

The wristband measures about 30 × 42 mm and combines a 609 nm pulse sensor, a TinyPICO ESP32 microcontroller and a 250 mAh battery that can last up to 24 hours. It samples PPG at 125 Hz and streams the signal to an Android application over Bluetooth.

The application combines PPG with blood pressure and demographic details, then sends the data to models hosted on Amazon Web Services. Estimated glucose and severity return to the phone every 10 seconds, and the design includes alerts for abnormal values.

Wristband size
30 × 42 mm
PPG sampling rate
125 Hz
Result update interval
10 s
Circuit diagrams of the device in work and charging modes, an exploded model of the wristband enclosure, and photographs of the finished device worn on the wrist.
Figure 2 from Shuzan et al. (2024). The prototype in (a) work and (b) charging modes, (c) its enclosure model and (d) the device worn on the wrist.View full-size figureSource articleCC BY 4.0
Three Android application screens: selecting the paired PPG sensor, entering demographic information while PPG is displayed, and viewing the glucose estimate with blood pressure fields.
Figure 7 from Shuzan et al. (2024). The Android application for connecting the device, entering user details and viewing the glucose estimate.View full-size figureSource articleCC BY 4.0

Collecting clinical data and engineering features

Qatar Diabetic Society · 139 participants

Three minutes of resting wrist PPG were recorded from each of 139 participants, with finger-prick glucose from an Accu-Chek meter and blood pressure and heart rate from an Omron monitor. The study had ethical approval from Hamad Medical Corporation, and all participants gave written informed consent.

Signals were filtered, cleaned with variational mode decomposition and divided into 4.096-second segments. The pipeline extracted 139 time-domain, statistical, Mel-frequency cepstral, demographic and clinical features. Removing highly correlated features left 78 for model selection.

Participant summary from Table 1 of Shuzan et al. (2024). Ranges are shown with the median in brackets.
Value
Participants139
With diabetes49.65%
Female41.84%
Age13–87 years (43)
Body mass index16.61–45.99 kg/m² (26.67)
Blood glucose66–600 mg/dL (118)
Flow diagram. Raw data passes through preprocessing and feature ranking into training, validation and test sets to select the best model. New data from the device follows the same feature steps to produce glucose and severity predictions.
Figure 3 from Shuzan et al. (2024). Model development and validation for glucose prediction and severity classification.View full-size figureSource articleCC BY 4.0

Estimating blood glucose on unseen participants

Regression with separate test participants

Participants were divided into separate training, validation and test groups in a 60:20:20 ratio. Three feature selection methods were compared with Gaussian process regression, decision trees and bagged ensemble trees. Bagged ensemble trees with 44 features selected by minimum redundancy maximum relevance gave the best result.

Demographic and clinical variables ranked among the most important features, alongside PPG-derived measures. On the Clarke Error Grid, 61.88% of predictions fell in zone A and 38.12% in zone B, with none in zones C, D or E.

Correlation with measured glucose
0.90
Mean absolute error
22.90 mg/dL
Predictions in Clarke zones A and B
100%
Selected regression results from Table 3 of Shuzan et al. (2024).
FeaturesCorrelationMean absolute errorRMSE
Bagged trees with MRMR440.9022.90 mg/dL30.74 mg/dL
Bagged trees with ReliefF520.8824.34 mg/dL32.15 mg/dL
Bagged trees with CFS530.8723.70 mg/dL32.56 mg/dL
Gaussian process with ReliefF480.8130.11 mg/dL40.77 mg/dL
Decision tree with MRMR560.7635.73 mg/dL49.44 mg/dL
Scatter plot of estimated against measured glucose from about 90 to 350 mg/dL, with the fitted trend line close to the ideal diagonal.
Figure 9 from Shuzan et al. (2024). Estimated against measured blood glucose.View full-size figureSource articleCC BY 4.0
Clarke Error Grid plotting predicted against reference glucose. All points lie within zones A and B.
Figure 11 from Shuzan et al. (2024). Clarke Error Grid analysis of the glucose estimates.View full-size figureSource articleCC BY 4.0

Classifying glucose severity

Three-class classification

A second model placed each sample into normal, warning or dangerous groups using measured glucose ranges. K-nearest neighbours, support vector machines and decision trees were compared with stratified five-fold cross-validation on 12,399 samples. K-nearest neighbours with 17 selected features performed best, using time-domain, statistical, Mel-frequency and demographic information.

Accuracy
98.12%
F1 score
98.12%
ROC AUC
0.998
Severity groups from Table 2 of Shuzan et al. (2024).
Blood glucose range
Normal66–109.8 mg/dL
Warning109.8–140.4 mg/dL
Dangerous140.4 mg/dL and above
Horizontal bar chart of the 17 features selected for severity classification. Weight, diabetes status and signal standard deviation have the highest relative importance.
Figure 12 from Shuzan et al. (2024). Features selected for severity classification, ranked by importance.View full-size figureSource articleCC BY 4.0

My contribution

I contributed to the machine learning and physiological signal analysis for the glucose monitoring framework, including technical evaluation and manuscript review.

Study scope

The models were developed on 139 participants recorded at rest, so larger and more diverse cohorts are needed. Blood pressure is currently measured separately and entered in the application. Severity classification used stratified cross-validation across samples, and longer-term testing during daily use is still required before clinical use.