DEEP LEARNING METHODS FOR MULTI-MODAL HEALTHCARE DATA
MetadataShow full item record
Abstract: Today, enormous transformations are happening in health care research and applications. In the past few years, there has been exponential growth in the amount of healthcare data generated from multiple sources. This growth in data has led to many new possibilities and opportunities for researchers to build different models and analytics for improving healthcare for patients. While there has been an increase in research and successful application of prediction and classification tasks, there are many other challenges in improving overall healthcare. Some of these challenges include optimizing physician performance, reducing healthcare costs, and discovering new treatments for diseases. - Often, doctors have to perform many time-consuming tasks, which leads to fatigue and misdiagnosis. Many of these tasks could be automated to save time and release doctors from menial tasks enabling them to spend more time improving the quality of care. - Health dataset contains multiple modalities such as structured sequence, unstructured text, images, ECG, and EEG signals. Successful application of machine learning requires methods to utilize these diverse data sources. - Finally, current healthcare is limited by the treatments available on the market. Often, many treatments do not make it beyond clinical trials, which leads to a lot of lost opportunities. It is possible to improve the outcome of clinical trials and ultimately improve the quality of treatment for the patients with machine learning models for different clinical trial-related tasks. In this dissertation, we address these challenges by - Predictive Models: Building deep learning models for sleep clinics to save time and effort needed by doctors for sleep staging, apnea, limb movement detection - Generative Models: Developing multimodal deep learning systems that can produce text reports and augment doctors in clinical practice. - Interpretable Representation Models: Applying multimodal models to help in clinical trial recruitment and counterfactual explanations for clinical trial outcome predictions to improve clinical trial success.