Abstract

Babies born prematurely are more likely to experience serious medical problems called Bradycardia, where their heart rate drops below 100 beats per minute(bpm). Bradycardia can decrease cerebral blood velocities by 10-50% from their normal level [6], potentially harming the infant's developing brain and other vital organs. Healthcare professionals working in Neonatal Intensive Care Units (NICUs) face challenges evaluating the likelihood of these events occurring in preterm babies based on their physiological signals. Forecasting and tracking bradycardia events before time will help save thousands of preterm infants from losing their lives. While time series analysis of heart rate data is one possible solution, creating a separate model for each infant is time-consuming and complex too. This thesis aims to test the hypothesis that the heart rate dynamics of preterm infants have unique system characteristics. We can study and track bradycardia events using only one infant, reducing the necessity of designing a specific model for each infant.

To test the hypothesis, a Long Short-Term Memory (LSTM) based deep learning network has been proposed for time series analysis of the heart rate of preterm infants using only heart rate as the input. Bradycardias are an important part of preterm infant cardiac dynamics due to which they are significant in this study. The LSTM model was trained using the heart rate data of a single infant and then tested on other infants having test segments of 10 minutes with at least 5 seconds of bradycardia event. Bradycardia events were categorized into three cases according to their nature of severity, severe, moderate, and mild. The developed model performed one-step, five steps, and ten-step time series prediction using three different sliding window scenarios, i.e., 30, 60, and 120 sec on the test data. The model was tested for severe, moderate, mild, and normal case scenarios for all three-time steps prediction, i.e., one-step, five-step, and ten-step, using sliding window sizes of 30, 60, and 120 seconds. The model performed well when tested on novel data, with a low average RMSE value.

After thoroughly analyzing the output data, it can be concluded that the proposed hypothesis about distinct system characteristics in each infant is accurate since the designed model can detect instances of bradycardia in each infant with a lower average root-mean-square error (RMSE). This could aid in efficiently managing and treating preterm infants in the neonatal intensive care unit (NICU).

Date of publication

Spring 4-16-2023

Document Type

Thesis

Language

english

Persistent identifier

http://hdl.handle.net/10950/4184

Committee members

Premananda Indic, Ph.D,Shawana Tabassum, Ph.D., Carla Lacerda, Ph.D.

Degree

Masters of Science in Electrical Engineering

Available for download on Thursday, April 24, 2025

COinS