UC Santa Barbara

Statistical methods, and in particular deep learning models, have achieved remarkable success in computer vision, speech recognition, and natural language processing due to the availability of powerful computational resources. Recently, neuroscience and healthcare have entered an exciting new age. Modern recording technologies, like Multi-Electrode Arrays (MEA) and electronic health records (EHR), offer unprecedented opportunities to explore neural systems and to improve health care. At the same time, they present extraordinary computational and statistical challenges. This Ph.D. dissertation presents knowledge we mined from neuroscience data and healthcare data.

In the neuroscience data mining part, we propose a deep learning framework for MEA classification of mouse and human derived induced Pluripotent Stem Cell recordings. We also introduce a scalable Bayesian framework for inference of functional neural networks from MEA data.

In the healthcare mining part, we first perform quantitative analysis on early multiple organ failure (MOF) prediction with comprehensive machine learning (ML) configurations, including data preprocessing (missing value treatment, label balancing, feature scaling), feature selection,classifier choice), and hyperparameter tuning. We introduce BERTSurv, a deep learning survival framework which applies Bidirectional Encoder Representations from Transformers (BERT) as a language representation model on unstructured clinical notes, for mortality prediction and survival analysis. We propose classifier-guided generative adversarial imputation networks (Classifier-GAIN) for MOF prediction, by incorporating both observed data and label information. Finally, we propose an end-to-end deep learning framework for chronic pain score assessment.

In the end, we summarize the strengths, weaknesses, and implications of our work, and discuss future research directions.