Dissertação
Machine learning techniques for detecting hypoglycemic events using electrocardiograms
Registro en:
CARMO, Natasha Rusty Silva. Machine learning techniques for detecting hypoglycemic events using electrocardiograms. 2021. 86 f. Dissertação (Mestrado em Ciência da Computação) - Universidade Federal de Sergipe, São Cristóvão, 2021.
Autor
Carmo, Natasha Rusty Silva
Institución
Resumen
Background Machine learning methods have long been employed to automatically analyze
electrocardiogram signals. In the past ten years, most studies have used a limited number of open
databases to test their results, most of which were collected in clinical settings. The growth in the
number of fitness trackers and other wearable devices that collect large amounts of data every
day offer a new potential to use data analysis to derive information that can improve the quality
of life for many people. Recently, an open database was released with data (electrocardiogram,
respiratory rate, motion data, food intake annotations and blood glucose) from patients with type
1 diabetes. It gives the opportunity to explore the potential of this data to predict hypoglycemic
events through a noninvasive method.
Methods The study uses pre-processing techniques to clean the data and extract features from
physiological signals in the dataset and verify how they correlate with blood glucose. Time
and frequency domain features are derived from the signal for the analysis. Automatic machine
learning is employed to determine the best classification model. The results are compared against
a 1D Convolutional Neural Network approach that automatically extracts features from individual
heart beats. The final models are evaluated in regards to performance metrics (accuracy, precision
and sensitivity) with respect to their ability to predict hypoglycemic events.
Results A 10-fold cross-validation provided the following percentage values for accuracy,
precision and sensitivity, respectively: 86.89 ± 2.8, 87.03 ± 2.7 and 86.90 ± 2.8 for the Random
Forest model and 93.00 ± 2.3, 93.08 ± 2.2 and 93.00 ± 2.3 for 1D CNN. The statistical evaluation
of the mean accuracy for both models from an unpaired T test returned a p-value lower than
0.0001, meaning that the distributions are significantly different and 1D CNN model outperforms
the decision tree model.
Discussion and Conclusion The small number of positive samples for hypoglycemia and high
data imbalance pose a challenge to classification. It is necessary to have reasonable number
of samples from both classes to achieve classification metrics that are suitable for medical
applications. When this condition is satisfied, data acquired from a wearable device under normal
living conditions has shown to be suitable for the task of classifying hypoglycemic events. São Cristóvão