Gangopadhyay, AryyaShi, Peichang2023-04-052023-04-052022-01-0112662http://hdl.handle.net/11603/27356Analysis of healthcare data could help reduce cost, improve patient outcomes and understand the best practices related to diseases.However,with rapid increase of massive amounts of health-related data, such as Electronic Health Records (EHRs),high dimensionality and large sample size have become challenges for traditional statis- tical approaches.Deep learning models have been proved to be powerful tools in computer vi- sion and machine learning in healthcare.However,despite their superior performance to traditional statistical methods,it remains challenging to understand their inner mechanism due to the black box effects. A variety of interpretability algorithms have been developed to help explain the deep learning models.However due to the trade off between model accuracy and complexity, the current interpretability algorithms have lower performance com- pared to original deep learning models, which cause some concern for high stakes in healthcare. Also,most of interpretability algorithms focus on correlation interpretation, highly correlated features may lead to biased causal inference, which may be more important in healthcare. In this dissertations paper,we proposed a new ensemble approach for deep learn- ing interpretation,Local surrogate Interpretable model-agnostic Visualizations and Explanations (LIVE), where we assumed all the predictions from deep learning model form a mixture of a finite number of Gaussian distributions with unknown parameters.We applied ensemble trees to obtain the mixing coefficients. The rule sets from the trees were used to build an interpretable model through randomized experimental design for interpretation. Our LIVE algorithm was validated using different types of datasets (image and structured datasets) with different deep learning model structures. Our experiments showed that LIVE algorithm could not only help improve the model accuracy, but also provide visual interpretation.application:pdfDeep learningHealthcareInterpretabilityText