An Ensemble Learning Approach for Exercise Detection in Type 1 Diabetes Patients

Ke Ma, Hongkai Chen, Shan LinarXiv 2023 · OpenReview

TL;DR

This paper introduces an ensemble learning approach to accurately detect exercise in Type 1 Diabetes (T1DM) patients, a critical missing piece for artificial pancreas (AP) systems. By combining a Siamese network for rapid initial detection with a more robust physiological model, they achieve significantly improved accuracy in identifying exercise events, crucial for preventing dangerous hypoglycemia.

What Problem Does This Paper Address?

Type 1 Diabetes (T1DM) is characterized by the body's inability to regulate blood glucose (BG) levels due to insufficient insulin production. Artificial Pancreas (AP) systems aim to automate insulin delivery to mimic a healthy pancreas. However, a major challenge for these systems is accurately detecting exercise. Exercise significantly impacts BG levels, often leading to a prolonged period of increased insulin sensitivity and glucose uptake (lasting 4-8 hours). Without proper detection and adjustment of insulin delivery, this can result in severe hypoglycemia, a life-threatening condition. Existing methods for exercise detection are often limited to single sensor data or employ inaccurate models, failing to provide the reliability needed for AP systems.

Methods

The core of this work is an ensemble learning framework that combines two distinct detection strategies:

  • Siamese Network-based Detector: This component utilizes multiple physiological signals (GSR, heart rate, step count, skin temperature) from wearable sensors. A Siamese network, specifically using Long Short-Term Memory (LSTM) networks, is trained to differentiate between exercise and non-exercise data sequences by measuring the similarity/distance between input pairs. This approach is designed for low latency and real-time detection.
  • Data-driven Physiological Model: This model incorporates a more complex understanding of how exercise and meal intake affect BG levels. It uses a set of differential equations (Equations 7-10) to describe the dynamics of plasma insulin, glucose, and heart rate. By analyzing BG readings from a Continuous Glucose Monitor (CGM), this model can identify deviations indicative of exercise. This method provides higher accuracy but may have a longer detection delay.

The ensemble approach uses the Siamese network as a first-stage detector. If it flags a potential exercise event, the more accurate model-based detector is then activated to confirm and refine the detection. This two-stage process aims to leverage the speed of the Siamese network and the accuracy of the physiological model.

Datasets

The study utilizes the OhioT1DM Dataset. This is a public dataset containing 8 weeks of data from 12 T1DM patients. It includes a rich set of physiological signals (glucose, heart rate, GSR, skin temperature, step counts, etc.) and self-reported information on meals, sleep, and exercise. The dataset was preprocessed by filtering out sleep data, smoothing ground truth exercise labels, and aligning data from different sensors.

Key Insights and Ideas

  • The critical need for exercise detection in AP systems: The paper clearly articulates the life-threatening risks of undetected exercise in T1DM patients and the limitations of current AP systems in this regard.
  • Synergy of speed and accuracy: The ensemble approach is a clever way to balance the trade-off between rapid detection (Siamese network) and high accuracy (physiological model).
  • Multi-modal sensing is key: Relying on a single sensor is insufficient. Combining data from wearables (GSR, HR, steps, temp) with CGM data provides a more comprehensive picture.
  • Exercise effects are prolonged: The paper highlights that exercise's impact on BG can last for hours, necessitating a detection system that can account for this extended influence.
  • Meal intake complicates exercise detection: The authors acknowledge that meals often occur around exercise, which can confound BG readings and impact detection accuracy.

Key Results

The proposed ensemble learning approach achieved:

  • True Positive Rate (TPR): 86.4%
  • True Negative Rate (TNR): 99.1%

These results significantly outperform the four baseline detectors evaluated, including the individual Siamese network and model-based detectors, as well as detectors based on energy expenditure and heart rate/acceleration data.

Limitations and Future Work

  • Limitations:
    • The paper notes that the Siamese network alone can have a high false detection rate due to its limited ability to differentiate exercise from other activities with similar movement patterns.
  • The model-based detector has a longer detection delay.
  • The OhioT1DM dataset has varying data sampling intervals and missing data, posing alignment challenges.
  • Some data fields are self-reported, introducing potential inaccuracies.
  • Future Work:
    • Incorporating meal detection: The authors plan to integrate a model-based meal detection system into their approach, recognizing the significant impact of meals on BG and exercise detection.
  • Exploring eating detection: Investigating eating detection using wearable acceleration sensors is another avenue for future research.
  • The ultimate goal is to improve the accuracy and robustness of the exercise detector to better serve T1DM patients in real-world scenarios.

Summary updated

Abstract

Type 1 diabetes is a serious disease in which individuals are unable to regulate their blood glucose levels, leading to various medical complications. Artificial pancreas (AP) systems have been developed as a solution for type 1 diabetic patients to mimic the behavior of the pancreas and regulate blood glucose levels. However, current AP systems lack detection capabilities for exercise-induced glucose intake, which can last up to 4 to 8 hours. This incapability can lead to hypoglycemia, which if left untreated, could have serious consequences, including death. Existing exercise detection methods are either limited to single sensor data or use inaccurate models for exercise detection, making them less effective in practice. In this work, we propose an ensemble learning framework that combines a data-driven physiological model and a Siamese network to leverage multiple physiological signal streams for exercise detection with high accuracy. To evaluate the effectiveness of our proposed approach, we utilized a public dataset with 12 diabetic patients collected from an 8-week clinical trial. Our approach achieves a true positive rate for exercise detection of 86.4% and a true negative rate of 99.1%, outperforming state-of-the-art solutions.
BibTeX
@misc{ma2023ensemble,
  author = {Ke Ma and Hongkai Chen and Shan Lin},
  title = {{An Ensemble Learning Approach for Exercise Detection in Type 1 Diabetes Patients}},
  year = {2023},
  eprint = {2305.10353},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.10353},
}
Figure