Clinical Dataset Analysis and Patient Outcome Prediction via Machine Learning

• Main Author: Buettner, Alex
• Format: Dissertation
• Language: English
• Published: ProQuest Dissertations & Theses 01.01.2018
• Subjects: Computer science; Health care management; Operations research
• ISBN: 9780438841598; 043884159X

We analyze and evaluate relevant machine learning methods for use in extract- ing and understanding clinical data sets in the context of optimization of clinical processes. Three data sets were considered to demonstrate the types and style of data found in the healthcare field: (a) the Pima Indians diabetes dataset (PIDD), a non-time-dependent diabetes onset study, (b) an alcoholism EEG dataset (AED), studying responses of alcoholic and control subjects when exposed to image stimulus, and (c) the diabetes readmission dataset (DRD), that focuses on factors that relate to diabetic patient readmission times. Each dataset is modeled using a variety of machine learning methods, including Bayesian, neural network, and decision tree methods, to better understand the advantages and disadvantages as applicable to rapid dependency extraction and understanding of the information contained therein. The goal of this work is to analyze the potential of machine learning for use in management of clinical processes and operations. Neural network models are used to assess all three data sets; two using dense neural networks, and one using convolutional neural networks. The dense neural network model used on the PIDD resulted in a maximum prediction accuracy of 81.77%. In contrast, the use of neural network (NN) models on the much larger DRD demonstrated some drawbacks that were not expected upon initial analysis of the data. We found that the NN model performs poorly on this dataset, with classification accuracy no higher than 61.17%, due to the complexity of the dataset and potential need for more data. The use of convolutional neural networks for analysis of time series data was demonstrated on the alcoholism EEG dataset, resulting i in subject classification accuracies between 91.41% and 98.82% depending on the training and testing sets used to analyze the model. Bayesian methods are used to analyze all three datasets, both in supervised and unsupervised manners. Supervised learning analysis on the PIDD showed improvement over published results, but are generally in agreement with the literature. Classifications accuracies of resulted in a maximum of 84.49% on a preprocessed dataset, and 79.75% on an unmodified dataset. Similarly, supervised learning was applied to the DRD, resulting in maximum classification accuracies on a three-class and two-class model of 58% and 62%, respectively. Unsupervised Bayesian methods are applied to the alcoholism EEG dataset in order to extract the true number of classes present in the model, in which all trials correctly identified two subject classes without the aid of labeling. Hidden Markov models are also applied to the alcoholism EEG dataset in an unsupervised fashion, allowing us to extract characteristic states in each EEG sample, for each subject class. The PIDD and DRD sets are also processed using decision tree models; gradient boosting classifiers (GBC) are applied to the PIDD, and extreme gradient boosting classifiers (XGBC) are applied to the DRD. The GBC model used to analyze the PIDD resulted in a maximum classification accuracy of 82.48% on preprocessed data, and 80.60% on an unmodified dataset. The DRD showed difficulty in model development, with maximum classification accuracy reaching approximately 55%, and with insensitivity to two of three data labels. The model seems unable to capture the components of the third class, showing that the distinguishability of the classes may be lacking.