Deep Learning with Multimodal Data for Healthcare

Healthcare plays a significant role in communities in promoting and maintaining health, preventing and managing the disease, reducing health disability and premature death, and educating a healthy lifestyle. However, healthcare information is well known for its big data that is too vast and complex to manage manually. The healthcare data is heterogeneous, containing different modalities or types of information such as text, audio, images, and multi-type. Over the last few years, the Deep Learning (DL) approach has successfully solved many issues. The primary structure of DL lies in the Artificial Neural Network (ANN). It is also known as representation learning techniques as these approaches can effectively identify hidden patterns of the data without requiring any explicit feature extraction mechanism. In other words, DL architectures also support automatic feature extraction. It is different than machine learning techniques, where there is no need to extract features separately in DL. In this dissertation, we proposed three DL architectures to handle multiple modalities data in healthcare. We systematically develop prediction models for identifying health conditions in several groups, including Post-Traumatic Stress Disorder (PTSD), Parkinson's Disease (PD), and PD with Dementia (PD-Dementia). First, we designed the DL framework for identifying PTSD among cancer survivors via social media. After that, we apply the DL time series approach to forecast PD patients' future health status. Last, we build DL architecture to identify dementia in diagnosed PD patients. All these work are motivated by several medical theories and health informatics perspectives. We have handled multimodal healthcare data information throughout these years, including text, audio features, and multivariate data. We also carefully studied each disease's background, including the symptoms and test assessment run by healthcare. We explored the online social media potential and medical applications capability for disease diagnosis and a health monitoring system to employ the developed models in a real-world scenario. The DL for healthcare can become very helpful for supporting clinician's decisions and improving patient care. The leading institutions and medical bodies have recognized the benefits it brings, and the popularity of the solutions are well known. With support from a reliable computational system, it could help healthcare decide particular needs and environments and reduce the stresses that medical professionals may experience daily. Healthcare has high hopes for the role of DL in clinical decision support and predictive analytics for a wide variety of conditions