Week 9 to 11 updates

 

Blog Update 2

Weeks 9 to 11 – 28/04/2025 to 16/05/2025

Feature Selection and Early Model Testing

Feature Diagnostics and Ranking

Using MATLAB’s Feature Diagnostic Designer, we calculated, ranked, and extracted the most relevant features for glucose estimation. These features were then applied across various quick-to-train machine learning models to measure performance improvements.

  • The best-performing model achieved an R² value of 0.14, marking a 14% improvement over the previously implemented neural network model.

  • This result was obtained using voltage sensor data, reinforcing the value of sensor-based inputs in glucose prediction.




ANN Prototyping and Evaluation

An Artificial Neural Network (ANN) was designed and tested using lifestyle data. The model achieved an accuracy of 55% on both validation and test sets.

  • While modest, these results demonstrate potential for model development, especially in scenarios where only lifestyle indicators are available.

As diabetes rates rise globally, the need for non-invasive, real-time glucose monitoring grows more urgent. This project contributes to that goal by leveraging artificial intelligence (AI) to estimate glucose levels without needles or strips, using a combination of sensor and lifestyle data.

In this second progress report, we explore the performance of different AI-based models to determine the optimal approach for portable, intelligent glucose monitoring.



Project Goal: Smarter, Simpler Glucose Monitoring

The primary aim of this stage was to evaluate and compare the following approaches:

  • Artificial Neural Networks (ANN)

  • Long Short-Term Memory (LSTM) networks

  • Curvilinear Component Analysis (CCA)-based regression models

These models were tested under three input configurations:

  1. Voltage-only sensor data

  2. Lifestyle data (e.g., diet, physical activity, sleep)

  3. Combined sensor + lifestyle data

The objective was to find the most accurate and computationally efficient model suitable for deployment in a real-time glucometer system.

Methodology Overview

  • Preprocessing: Missing values were handled, outliers were removed, and Z-score normalization was applied.

  • Dataset Split: 70% for training, 15% for validation, and 15% for testing.

  • Model Development: All models were implemented and evaluated using MATLAB.

AI Models in Focus

Artificial Neural Networks (ANN)

  • Structure: Two hidden layers, scaled based on input complexity.

  • Strengths: Effective at modeling non-linear relationships.

Results:

  • Voltage Data: R² = 0.53 – shows moderate correlation.

  • Lifestyle Data: R² = 0.07 – limited predictive capability.

  • Combined Data: R² = 0.50 – slight improvement, though lifestyle data may introduce noise.


Long Short-Term Memory (LSTM) Networks

  • Structure: Utilizes memory cells and gating mechanisms to model sequential data.

  • Strengths: Ideal for capturing time-dependent patterns in glucose levels.

Results:

  • Voltage Data: RMSE = 0.16, R² ≈ 0.70 – significantly better than ANN.

  • Lifestyle Data: R² ≈ 0.36 – improved over ANN but still limited.

  • Combined Data: R² ≈ 0.83 – best overall performance, showcasing LSTM’s ability to integrate diverse data types.



CCA-Based Regression Models

  • Approach: Applied Curvilinear Component Analysis to reduce 20 features to 9, followed by linear regression.

Results:

  • Linear Regression: R² = 0.99 (validation), R² = 1 (test), RMSE <2% of glucose range.

  • Significance: Extremely high accuracy and low computational cost, making CCA-based models ideal for embedded systems.

Key Takeaways

  • LSTM networks excel at modeling complex, temporal patterns in glucose levels—especially with combined sensor and lifestyle data.

  • ANNs underperform when lifestyle data is used in isolation or even in conjunction with sensor data.

  • CCA-based regression models offer remarkable accuracy with minimal computational load, making them strong candidates for real-time, portable deployment.



Upcoming stages will focus on:

  • Hyperparameter tuning

  • Exploration of ensemble modeling techniques

  • Clinical validation using Clarke’s Error Grid Analysis

  • Hardware integration for real-time testing of the top-performing model

We’re making strong strides toward an AI-driven, accessible, and non-invasive solution for blood glucose monitoring—and we’re just getting started.

Comments

Post a Comment

Popular posts from this blog

EE461 Project Breakdown

  EE461 Project Blog: Glucose Estimation Model Development Welcome to the official blog for our EE461 project! This space will serve as a central hub for logging our progress, sharing insights, backing up key developments, and tracking milestones throughout the project. Posts will be updated fortnightly, ensuring consistent documentation of our journey as we work towards building a robust glucose estimation model. Project Overview The primary goal of this project is to develop a reliable machine learning model capable of estimating glucose levels based on available physiological or sensor data. This task involves building the model itself and ensuring the quality and relevance of the input data, engineering meaningful features, and validating model performance through sound evaluation metrics. Project Objectives To achieve our goals, the project will be guided by the following key objectives: Data Preprocessing Cleaning and preparing the raw dataset is a crucial first ste...
Read more

Week 12 and 13 updates

  Blog Update 3 Week 12 and 13 20/05/25 - 03/06/25 Inference Pipeline Development and Final Presentation Preparation This week marked the transition from model experimentation to a fully functional inference pipeline, laying the groundwork for real-time glucose prediction. Alongside technical implementation, we also finalized the presentation materials and prepared a demo package for showcasing the end-to-end system. Inference Pipeline Implementation To simulate real-world prediction scenarios, a test sample was extracted and pushed to the end of the dataset. This allowed us to mimic unseen data inference while preserving a ground truth value for validation. The complete dataset, including this test sample, was then passed through the full preprocessing workflow, which included normalization and dimensionality reduction. Curvilinear Component Analysis (CCA) was used to transform the feature space, reducing it to 45 dimensions. This technique, already proven effective in prio...
Read more