[vc_row][vc_column][vc_column_text]
CalCardiac worked to reduce false hospital alarms due to arrhythmias. Collaborating with the UCSF School of Nursing, the team used various machine learning techniques to create a browser-based utility to view and label ECG waveform images.
[/vc_column_text][vc_column_text]Capstone Team: Alex Ackroyd, Adam Andrews, Siddhant Issar, Segev Malool, Umesh Thillaivasan, Yuntao Wang.
Research in collaboration with Jacob Abba and Ran Xiao at UCSF; Advised by Xiao Hu, UCSF, and Gabriel Gomes, UC Berkeley.[/vc_column_text][vc_row_inner][vc_column_inner width=”2/3″][vc_column_text]
Challenge
CLASSIFYING HEARTBEATS
When the heart beats irregularly, it is known as an arrhythmia. A common heart arrhythmia, known as premature ventricular contraction (PVC), ac-counts for the highest number of non-actionable and false-positive in-hospital patient monitoring alarms. Current in-hospital patient monitoring systems do not have the capabilities to discern true-PVC alarms from false-alarms, and therefore medical professionals experience alarm fatigue, a desensitization to alarms leading to lower quality of care. This presents an opportunity to apply sophisticated machine learning methods to improve the accuracy of these alarms.
Actions & Results
MACHINE LEARNING
Our team explored several machine learning approaches to handle and classify electrocardiogram (ECG) signal data from two data sets: the famous MIT-BIH Arrhythmia labeled data set, and UC San Francisco’s massive unlabeled data set. The MIT-BIH Arrhythmia Database has 48 records of half-hour ECG signals. UCSF’s data set has over 10 million records each 10-seconds long.
LOGISTIC REGRESSION, RANDOM FOREST, AND NEURAL NETWORKS
We sliced the MIT-BIH data set into 0.8-seconds ECG signal strips, 0.4-seconds before and 0.4-seconds after each annotated beat. Then, using a data set of 7129 PVC alarms, and 7129 non-PVC records derived from the full MIT-BIH sliced data set, Logistic Regression, Random Forest, and a Neural Network was trained.
TSNE, K-MEANS CLUSTERING, AND TRANSFER LEARNING
We also explored techniques to distinguish types of heartbeats. T-distributed stochastic neighbor embedding (TSNE) at-tempts to preserve high dimensional neighborhoods using a probabilistic interpretation of transformed data (Left). Anomaly detection with K-Means clustering algorithms can be used for QRS complex detection in time series ECG signal data (Center). Transfer Learning using one of Google’s pre-trained models can lead to improved heart beat classification (Right).[/vc_column_text][/vc_column_inner][vc_column_inner width=”1/3″][vc_video link=”https://www.youtube.com/watch?v=la_5fC1mWNk&feature=youtu.be”][vc_column_text]*CalCardiac was the 2017 winner of the Fung Institute Alumni Award for the Most Innovative Exposition. This award honors a team that presents their project in the most creative way to a general audience.[/vc_column_text][/vc_column_inner][/vc_row_inner][vc_separator][vc_single_image image=”41087″ img_size=”full”][/vc_column][/vc_row]