Search

This paper proposes a novel deep learning architecture involving combinations of Convolutional Neural Networks (CNN) layers and Recurrent neural networks (RNN) layers that can be used to perform segmentation and classification of 5 cardiac rhythms based on ECG recordings. The algorithm is developed in a sequence to sequence setting where the input is a sequence of five second ECG signal sliding windows and the output is a sequence of cardiac rhythm labels. The novel architecture processes as input both the spectrograms of the ECG signal as well as the heartbeats' signal waveform. Additionally, we are able to train the model in the presence of label noise. The model's performance and generalizability is verified on an external database different from the one we used to train. Experimental result shows this approach can achieve an average F1 scores of 0.89 (averaged across 5 classes). The proposed model also achieves comparable classification performance to existing state-of-the-art approach with considerably less number of training parameters.

Hemorrhage is a frequent complication in surgery patients; its identification and management have received increasing attention as a target for quality improvement in patient care in the Intensive Care Unit (ICU). The purposes of this work were 1) to find an early detection model for hemorrhage by exploring the range of data mining methods that are currently available, and 2) to compare prediction models utilizing continuously measured physiological data from bedside monitors to those using commonly obtained laboratory tests. We studied 3766 patients admitted to the University of Virginia Health System Surgical Trauma Burn ICU. Hemorrhage was defined as three or more units of red blood cells transfused within 24 h without red blood cell transfusion in the preceding 24 h. 222 patients (5.9%) experienced a hemorrhage, and multivariate models based on vital signs and their trends showed good results (AUC = 76.1%). The hematocrit, not surprisingly, had excellent performance (AUC = 87.7%). Models that included both continuous monitoring and laboratory tests had the best performance (AUC = 92.2%). The results point to a combined strategy of continuous monitoring and intermittent lab tests as a reasonable clinical approach to the early detection of hemorrhage in the surgical ICU.

Sample entropy and approximate entropy are measures that have been successfully utilized to study the deterministic dynamics of heart rate (HR). A complementary stochastic point of view and a heuristic argument using the Central Limit Theorem suggests that the Gaussianity of HR is a complementary measure of the physiological complexity of the underlying signal transduction processes. Renyi entropy (or q-entropy) is a widely used measure of Gaussianity in many applications. Particularly important members of this family are differential (or Shannon) entropy (q = 1) and quadratic entropy (q = 2). We introduce the concepts of differential and conditional Renyi entropy rate and, in conjunction with Burg's theorem, develop a measure of the Gaussianity of a linear random process. Robust algorithms for estimating these quantities are presented along with estimates of their standard errors.

While heart rate variability has been measured in many clinical settings and has offered insights into how HR is controlled, rarely has it offered unique information that has led to changes in patient management. We review our experience in developing continuous HR characteristics monitoring to aid in the early diagnosis of sepsis in premature infants in the neonatal intensive care unit. A predictive algorithm, developed at one center and validated at another, has led to diagnosis and treatment of this subacute and potentially catastrophic illness prior to appearance of symptoms of severe illness.

Continuous electronic fetal heart rate (FHR) monitoring during labor is motivated by the clinical experience that fetal distress causes loss of FHR variation and the occurrence of decelerations late during uterine contraction. This practice is of uncertain clinical benefit, perhaps because the interpretation is qualitative. We have developed new quantitative measures and analyzed cardiotocograph records from 148 consecutive patients, 44 of whom had at least one "nonreassuring" epoch. In multivariate regression models, measures of deceleration and variability were significantly associated with the obstetrician's diagnosis (receiver operating characteristic area 0.84, p < 0.05). This approach may be useful clinically.