Blood Glucose Prediction Using Continuous Monitoring Data

This repository contains the solution for the Kaggle competition Blood Glucose Prediction. The goal is to predict blood glucose levels one hour into the future using a dataset collected from young adults in the UK with type 1 diabetes. Participants used continuous glucose monitors (CGMs), insulin pumps, and smartwatches to collect data aggregated into five-minute intervals.

Competetion:NoteBook

Dataset Description

The dataset consists of:

• Training data: Samples from the first three months of the study for 9 participants, including features and the target blood glucose value (bg+1:00).
• Testing data: Samples from the later study period for 15 participants, excluding the target value, with no overlap with the training data and in a randomized order.

Key Features

• Medical Data: Includes blood glucose readings, insulin doses, carbohydrate intake, heart rate, step count, calorie burn, and activity labels for six-hour historical windows.
• Challenges:
  • Missing values and noise.
  • Different device models used by participants.
  • Test data contains unseen participants.

Data Files

• activities.txt: List of activity labels.
• train.csv: Training data with features and target.
• test.csv: Test data with features only.
• sample_submission.csv: Sample submission file format.

Solution Overview

Key Components

1. Preprocessing:

   • Handles missing values using interpolation and imputation.
   • Adds cyclical features for time of day (sine and cosine transformations).
   • Standardizes the dataset for machine learning.

2. Model Training:

   • Trains a LightGBM regression model to predict blood glucose levels.
   • Supports hyperparameter tuning via grid search.
   • Logs training results and metrics using MLflow.

3. Outputs:

   • Preprocessed Data: For faster loading in subsequent runs.
     • X.npy: Preprocessed features for training.
     • y.npy: Preprocessed target labels.
     • df_test_final.npy: Preprocessed test data features.
   • Model: Saved LightGBM model in the models/ folder.
   • Submission Files:
     • Submission_LGB_No_Tuning.csv: Predictions from the baseline model.
     • Submission_LGB_Fine_Tuned.csv: Predictions from the fine-tuned model.

How to Run

1. Set Up Environment:

   • Install dependencies:

     pip install -r requirements.txt

   • Start an MLflow tracking server:

     mlflow server --backend-store-uri sqlite:///mlflow.db --default-artifact-root ./mlruns

2. Run the Pipeline:

   • Preprocess data, train models, and generate submissions:

     python main.py

3. Outputs:

   • Preprocessed data saved in the preprocessed_data/ folder.
   • Trained LightGBM model saved in the models/ folder.
   • Submission files saved in the project root directory.

Results

• Baseline Model:
  • Mean Absolute Error (MAE): Logged in MLflow.
  • Root Mean Squared Error (RMSE): Logged in MLflow.
• Fine-Tuned Model:
  • Achieves improved performance with hyperparameter optimization.

Future Work

• Explore advanced models like recurrent neural networks (RNNs) or transformers for time-series prediction.
• Implement feature selection and importance analysis for better interpretability.
• Experiment with additional imputation methods for missing data.