Abstract: Deep learning methods refer to variants of neural network models where two or more layers of nodes are integrated with each other. Use of these models is in a renaissance given the substantial advances in the accuracy of neural network models for computer vision, natural language processing and speech recognition. Availability of large scale datasets, advances in high-performance computational devices such as Graphics Processing Units (GPU), and developments of new optimization techniques have permitted efficient learning of deep neural networks for a diversity of application domains.
Deep learning methods are being applied in the healthcare domain:
Automated feature learning: Autoencoders have been used for phenotyping purposes using time-series data. Skip-gram methods were used for learning the representation vectors for medical codes such as diagnosis codes and medication codes. Recurrent networks(RNN) have been used to learn representations from clinical notes. Convolutional networks(ConvNet) also showed great potential for learning effective representations of continuous vital signs.
Accurate predictive models: Training a high-capacity model for accurate prediction is one of the most sought-after direction from many researchers as well as practitioners. ConvNets have shown state-of-the-art performance for detecting retinopathy from retinal images and detecting skin cancer from skin photographs. RNN has also shown great potential for multi-label diagnoses prediction and early detection of heart failure onset.
Synthetic data generation: Generating synthetic electronic health records is another popular direction due to the sensitive nature of private records. Generative adversarial networks is actively being used to generate synthetic claims data and lab measures, or even to guarantee some level of privacy.
However, deep learning models still have many limitations in support healthcare applications:
Interpretation: The deep learning models are often complex black boxes, which are difficult to understand. However, clinical practitioners often prefer and trust simpler and interpretable models. We will discuss how to bridge the gap between model complexity and interpretation.
Causal learning: Great predictive models do not necessarily lead to causal discovery. For clinical applications such as treatment selection, we need to know what treatment change will lead to the best outcome, which often requires identifying causal relations.
Intended Audience: Medical informatic researchers and practitioners.
Learning Objective 1: understand the challenges and limitations of deep learning for healthcare applications
Jimeng Sun (Presenter)
M. Brandon Westover (Presenter)
Hong Yu (Presenter)
David Sontag (Presenter)
Marzyeh Ghassemi (Presenter)