Project Summary/Abstract With a constant stream of patient data generated at the hospital bedside, clinicians are asked to interpret this data along with patient medical records and lab results in real time. The proposed project offers an approach to automated clinical decision support (CDS) in parsing through some of this abundant data, focusing on the time series summarization (TSS) of the electrocardiogram (ECG) and approximations to the related vectorcardiogram (VCG) using techniques at the interface of data science and applied mathematics. Given the fact that bedside monitor signals can be corrupted by noise, it is important to distinguish between noise/artifact, cardiac arrhythmia, and normal cardiac rhythms; while the literature approaching such issues is growing, there is still a need for addressing this problem for the pediatric population – especially for pediatric patients with electrical conduction abnormalities as seen in the Cardiac Intensive Care Unit (CICU). Through collaboration between investigators at the University of Central Oklahoma (UCO) and at Baylor College of Medicine and Texas Children’s Hospital (TCH), this project combines the application of deep learning algorithms and subset selection techniques such as the discrete empirical interpolation method (DEIM) to classify and summarize data recorded from the pediatric CICU at TCH. Specifically, the objective of this project is two-fold: (1) apply variational autoencoders (VAEs) to differentiate between noise, arrhythmias, and normal sinus rhythm, and (2) evaluate both existing and newly developed subset selection algorithms, with an added emphasis on DEIM-related methods in application to cardiac data. Undergraduate students at UCO will evaluate VAE architectures for noise detection, performing model selection and then applying the chosen model to patient data for further analysis. Additional VAE models will be trained and selected for recognizing ECG and VCG waveforms containing pathologies. VAE results will be compared to those generated using existing methods in the literature and will inform the subsequent summarization of patient data. While DEIM has demonstrated viability in class- identification tasks in prior work, DEIM and its related methods were originally developed for applications such as mathematical model reduction, not class identification. For this reason, students will perform a necessary comparison of DEIM-related methods applied to a variety of data types, giving particular attention to experiments involving ECG waveforms; while doing so, students will also develop a novel extension of such methods tailored to this specific medical context. In addition, the comparison of these techniques for class identification purposes will offer valuable insight regarding DEIM-related methods to both the larger biomedical informatics and data science communities. Once established, this TSS framework will provide a means of presenting to clinicians a representation of...