T1D Change Point Detection Benchmark

An open source benchmark for semi-supervised change point detection of type 1 diabetic meals from continuous glucose monitor time series data. Originally created to present to PyData Global 2024 in association with sktime and skchange.

Versioning

NOTE: The benchmark is still under heavy development and is subject to change. Consider its current state as alpha v.0.0.1

• 0.0.1 -> Patch/bug fixes
• 0.1.0 -> Minor data set updates where we add new patients, cgms, insulin pumps, or meal timing regimens.
• 1.0.0 -> Major data set updates where we add new modeling tasks, like new transfer learning settings.

Dataset Metadata

| property | value | |-------------|-------------------------------------------------------------------------------| | name | T1D Semi-Supervised Change Point Detection Benchmark | | url | https://github.com/Blood-Glucose-Control/t1d-change-point-detection-benchmark | | sameAs | https://github.com/Blood-Glucose-Control/t1d-change-point-detection-benchmark | | description | | | citation | | | license | |

Data Directory Structure

Overview

This repository contains three main data directories: raw, processed, and obfuscated, each serving different purposes in the data pipeline.

Directory Details

1. data/raw

Contains data generated directly from simglucose simulator.

Characteristics:

• Duration per patient: 90 days
• 30 patients (10 adults, 10 children and 10 adolescents)
• Source: Jinyu Xie. Simglucose v0.2.1 (2018)
• Reference: https://github.com/jxx123/simglucose

2. data/processed

Contains processed data derived from data/raw.

File Naming Convention

Pattern: {patientNum}_{cgmName}_{insulinPumpName}_{startDate}_{endDate}.csv

Example: ado001_Dexcom_Cozmo_2024-02-01_2024-04-30

| Component | Description | Example | |-----------------|------------------------------------------------------------------|---------------------------| | patientNum | Concatenation of first 3 and last 3 characters from patient name | ado001 (adolescent#001) | | cgmName | CGM device name | Dexcom | | insulinPumpName | Insulin pump device name | Cozmo | | startDate | First day of generated data | 2024-02-01 | | endDate | Last day of generated data | 2024-04-30 |

3. data/obfuscated

Contains data obfuscated from data/processed to simulate human behavior.

File Naming Convention

Pattern: {patientNum}_{cgmName}_{insulinPumpName}_{startDate}_{endDate}_{loggingBehaviour}_{loggingTiming}.csv

Example: ado001_Dexcom_Cozmo_2024-02-01_2024-04-30_all_normal.csv

| Component | Description | Example | |-----------------|------------------------------------------------------------------|---------------------------| | patientNum | Concatenation of first 3 and last 3 characters from patient name | ado001 (adolescent#001) | | cgmName | CGM device name | Dexcom | | insulinPumpName | Insulin pump device name | Cozmo | | startDate | First day of generated data | 2024-02-01 | | endDate | Last day of generated data | 2024-04-30 |

Logging Behavior Types

| Filename Indicator | Type | Description | Distribution | |--------------------|------------------------|--------------------------------------------------|--------------| | all | All meals | Logs every meal | 20% | | top2 | Multiple meals per day | Logs 1-2 largest meals (on average 1.8 logs/day) | 25% | | once | Once per day | Logs largest meal only | 20% | | weekly | A few times per week | Irregular logging (on average 3 logs/week) | 20% | | none | Never | No logging | 15% |

Note: Distribution percentages are subject to change

Logging Timing Patterns

| Filename Indicator | Pattern | Description | Distribution | |---------------------|---------------------|----------------------------------------|--------------| | late | Left skewed | Forgetful loggers (gamma distribution) | 38% | | early | Right skewed | Hasty loggers (gamma distribution) | 23% | | average | Normal Distribution | Centered around meal start time | 28% | | punctual | Unchanged | Logs exactly at meal start | 11% |

Note: Distribution percentages are subject to change

Note: Parameters for gamma/distribution are subject to change

Gamma/normal distribution for loggingTiming

Each graph contains 50 randomly generated curves

Average logger:

Late logger:

Early logger:

T1D Change Point Detection Benchmark

A benchmark dataset for evaluating change point detection algorithms on Type 1 Diabetes data.

Installation

You can install the package using pip:

pip install t1d-cpd-benchmark

Example usage

from t1d_cpd_benchmark.loader import T1dCpd

# Initialize the dataset
dataset = T1dCpd()

# Load all splits
all_data = dataset.loader()  # Returns {"train": DataFrame, "test": DataFrame, "val": DataFrame}

# Load specific split
train_data = dataset.loader(split="train")  # Returns train split Dataset
test_data = dataset.loader(split="test")    # Returns test split Dataset
val_data = dataset.loader(split="val")      # Returns validation split Dataset

Loading Specific Patient Data

TODO: Update the current numbers of patients There are currently 180 patients with id 0 to 179

# Load data for specific patients
patient_ids = [1, 3, 5]
patient_data = dataset.loader(patient_ids=patient_ids)  # Returns data for specified patients

# Load specific split for specific patients
train_patient_data = dataset.loader(split="train", patient_ids=patient_ids)

Return Format

The loader returns data in the following format:

{
    "id": int,
    "train": pd.DataFrame,
    "test": pd.DataFrame,
    "val": pd.DataFrame
}

License

This project is licensed under the MIT License - see the LICENSE file for details.