Have a UROP opening you would like to submit?
Please fill out the form.
Detecting Life-threatening Cardiac Arrhythmias Using Machine Learning Techniques
QI: MIT Quest for Intelligence
July 9, 2020
Background: False arrhythmia alarm rates in ICUs may be as high as 89%, leading to disruption of care and “alarm fatigue.” In the 2015 PhysioNet Challenge , discriminating between true and false arrhythmia alarms proved to be hardest for ventricular tachycardia (VT). Accordingly, we plan to systematically investigate machine-learning approaches to detect true arrhythmia alarms. Prior Work: Our prior work in Lehman et al.  conducted an investigation of the utility of a semi-supervised generative approach for false alarm detection. We employed nonlinear compact representations of ECG spectral dynamics, using the Supervised Denoising Autoencoder (SDAE), which simultaneously minimized the ECG reconstruction and false alarm detection errors. In the 2015 Challenge databases, our approach obtained a reduction from 79.6% to 11.3% in the VT false alarm rate in the test set: 151 false alarms were suppressed at the cost of missing 5 out of 45 (11.1%) true alarms. Despite promising results from our prior work in SDAE , reducing VT alarms remains a challenge, as our ultimate goal is to reduce all false VT alarms, without missing any true alarms. Proposed Work: The goal of this project is to investigate machine learning approaches for reducing false VT alarms. This Summer UROP involves the following components: (i) expand the VT alarm training data set by combining multiple sources of existing annotated datasets, including the Challenge 2015 dataset, (ii) develop new machine learning approaches to reduce VT alarms using the expanded dataset; incorporate recent advances in transfer and multitask learning methods to enable efficient learning; develop models using recurrent or convolutional neural networks to efficiently model the dynamics and recurrent morphological changes observed in multichannel waveform data, (iii) work with a team of engineers and physicians to develop an interactive alarm annotation interface to collect additional labeled VT alarms; and investigate the use of machine learning techniques (such as active learning) to efficiently annotate large-volumes of waveform data, and (iv) improve performance of the machine learning model developed in step ii using the newly collected labeled data from the previous step.  PhysioNet 2015 Challenge. https://www.physionet.org/content/challenge-2015/1.0.0/  "Representation Learning Approaches to Detect False Arrhythmia Alarms from ECG Dynamics," Eric P. Lehman, Rahul G. Krishnan, Xiaopeng Zhao, Roger G. Mark, Li-wei H. Lehman, Proceedings of the 3rd Machine Learning for Healthcare Conference , PMLR 85:571-586, 2018. http://web.mit.edu/lilehman/www/paper/representation_learning_LehmanEtAl_MLHC2018.pdf
The ideal candidate would have taken courses in machine learning or related courses. Knowledge and experience in one or more of the following areas would be desirable: deep learning, transfer learning. Familiarity with signal processing would be a plus. Apply by May 7th for funding, or by July 9th for credit.