Representation Learning on Large-Scale Neural and Healthcare Data: A Practitioner’s Perspective

Abstract

The rapid growth of data volume has promoted development of advanced computational methods, particularly machine learning (ML) and deep learning (DL) algorithms, to address challenges emerged in various fields. A key step to solve ML/DL tasks is to find a good data representation by mapping input data into a feature space where representations can entangle or disentangle the different explanatory factors of variation behind the data, and predictive modeling can be more accurate and reliable. In this thesis, we focus on the study of representation learning methods in neuroscience and healthcare, and propose algorithms based on recent ML/DL developments to address both critical and practical challenges in neural signal processing and healthcare predictive modeling. For neural signal processing, we first strive to bridge the gap of the fast-increasing scale of data acquisition in neural recording and the limited bandwidth of data links for transmission. We propose two unsupervised compression algorithms to reduce the bandwidth of neural signals without sacrificing their utilities in downstream tasks. This is mainly achieved by leveraging the morphological consistency of neural signals across geometrically adjacent recording sites to capture common data variations. Next, we propose a semi-automatic spike sorting algorithm to decompose multi-unit recordings into single-unit activities based on adversarial representation learning that can sort spikes from a small number of labeled examples, thereby mitigating the data-hungry limitation of DL-based classification models. For healthcare predictive modeling, we propose to represent the hierarchical and relational structures of medical entities (patients, doctors, and medical services) in patients electronic health records (EHR) using a collection of graph-based network embedding algorithms. The proposed framework can bring a number of advantages such as enhanced clinical outcome prediction accuracies and more interpretable modeling of patient medical profiles and treatment history, which suggest the potential of being used as a comprehensive and general-purpose solution for representation learning of EHR data.