Heart rate variability in infants for analysis of life-threatening events and sleep versus wake classification

Life-threatening events including bradycardia and apnea in infants are a major health concern for families and physicians. It is our hypothesis that biomedical signal analysis including statistical analysis and computational intelligence techniques applied to infant electrocardiogram data could be helpful in the understanding of life-threatening events.; This thesis is the study of infant electrocardiogram data from the NIH Collaborative Home Infant Monitoring Evaluation CHIME. This NIH dataset contains a wealth of data from over 1000 infants, over 360 gigabytes, at postconceptional ages ranging from 34 weeks to 63 weeks. The information contained in the dataset includes home monitor data for most infants and a polysomnograph (PSG) from the hospital, of about 70% of the population.; Our goal is to study the feasibility of heart rate variability measurements derived from the electrocardiogram of the CHIME home monitor to predict life-threatening events in infants. This thesis has two main components: (1) classification of sleep/wake using heart rate and (2) prediction of life-threatening events using the PSG dataset.; In order to strengthen the analysis of the CHIME data collected in the home, an automated sleep/wake classification algorithm is developed based only on electrocardiogram information. Typical sleep state classification algorithms use multiple physiologic measurements recorded during a polysomnograph not available in the home data. The algorithm developed will focus on designing a computational intelligence model that will provide a highly reliable decision or give no decision. For the prediction problem, other studies have shown better results when data is first separated by sleep state. Thus, it is important that sleep state classification is highly reliable. The sleep/wake classification algorithm can achieve 85%--87% correct classification while rejecting only 30% of the data. This is an improvement of about 7% over a traditional model without rejection.; The possibility of building a model to predict life-threatening events (LTE) is studied in two parts. Computational intelligence models are evaluated to determine if a LTE can be predicted some fixed time before the LTE occurs. Two different prediction windows are studied: one month and one week. The results assess whether events can be predicted and, if they can, how soon before they occur will the prediction be accurate. Also, a statistical approach is taken for data analysis to determine which features, if any, have statistical significance comparing event and nonevent infants. Different factors are tested for their influences in changing HRV measures. Once statistical significance is determined for HRV measures, features are extracted that show the most significant difference for separating event infants from the control infants. The resulting features are used with computational intelligence techniques for determining if event prediction is possible. Using the features returned from the statistic study, a neural network can predict 76% of the testing dataset correctly (sensitivity = 66.67% and specificity = 70.73%).; In summary, this dissertation focuses on applying engineering algorithms to infant electrocardiogram data in order to study two important hypothesis: (1) classification of sleep/wake using heart rate and (2) prediction of life-threatening events.