Deep learning has been successfully applied to fields such as video and image analysis, text and speech processing, and autonomous driving. However, healthcare applications are few and mainly focus on areas such as medical image processing. As a result, there are few tools and best practices available.
Julie Zhu and Dima Rekes describe an effort to benchmark deep learning against traditional machine learning methods, such as logistic regression and XGBoost, and share a deep learning approach using Python and Keras for imputing a medical condition based on a multiyear history of prescriptions filled by an individual, combining time-related events and static variables such as age and gender. Julie and Dima discuss the differences in data preparation between deep learning and traditional machine learning, which enables the removal of the feature engineering steps, and detail the process of evaluating different network architectures, training the network, data exploration, grid search, transfer learning, and multi-GPU parallelization. Along the way, you’ll learn the relative merits of CNNs, RNNs (LSTMs and GRUs), and Siamese networks and compare specialist and generalist models.
Julie Zhu is director of data science at Optum.
Dima Rekesh is a senior distinguished engineer at Optum, a division of UnitedHealth Group, where he works on technology strategy with an emphasis on deep learning. Previously, Dima spent many years as a distinguished engineer at IBM, where he was involved in a wide variety of deep learning projects related to analytics in the cloud and at the edge.
©2018, O'Reilly Media, Inc. • (800) 889-8969 or (707) 827-7019 • Monday-Friday 7:30am-5pm PT • All trademarks and registered trademarks appearing on oreilly.com are the property of their respective owners. • firstname.lastname@example.org