Parameter invariant statistics and their application to clinical decision support / Alexander Steven Roederer.

Format:

Book

Manuscript

Thesis/Dissertation

Author/Creator:

Roederer, Alexander Steven, author.

Contributor:

Lee, Insup, degree supervisor.

Hanson, C. William, III, degree supervisor.

Barash, Yoseph, degree committee member.

Taylor, Camillo J., degree committee member.

Ungar, Lyle, degree committee member.

Yang, Christopher, 1974- degree committee member.

University of Pennsylvania. Department of Computer and Information Science, degree granting institution.

Language:

English

Subjects (All):

Penn dissertations--Computer and Information Science.

Computer and Information Science--Penn dissertations.

Local Subjects:

Penn dissertations--Computer and Information Science.

Computer and Information Science--Penn dissertations.

Physical Description:

xiv, 136 leaves : illustrations ; 29 cm

Production:

[Philadelphia, Pennsylvania] : University of Pennsylvania, 2016.

Summary:

The proliferation of digital medical device technology in the modern hospital has led to an explosion in the amount of available patient data. This data deluge presents opportunities for improved patient care, but in the current clinical environment it also creates significant challenges. One approach to alleviating the data deluge burden has been to develop clinical decision support systems, which automatically analyze data and provide the results to clinicians.

With an eye to improving clinical decision support systems, in this thesis we describe parameter invariant classification, a technique for detecting changes in patient state. The technique requires little data, is resilient to interpatient variability and noise, and can achieve a good detection rate while guaranteeing a constant rate of false positives over each individual in the population. It does so by avoiding direct estimation of a patient's parameters; instead, it considers two time-invariant linear systems (each describing a possible patient state) and selects which is more likely to have generated the data at any given time. This selection process is designed to be invariant to nuisance parameters-parameters that do not help to distinguish between the candidate systems.

The predictability advantages of parameter invariance come at a cost: the average performance of a parameter invariant classifier is usually lower than that of other classifiers. One reason is any parameters left out of the nuisance parameter set with a low signal-to-noise ratio decrease the accuracy of the classifier. To remedy this, we describe an algorithm in which parameter invariant statistics are calculated over many possible feature subspaces. Then, feature selection is used to choose the combination of statistics which achieves the best performance, selecting for subspaces with high signal-to-noise ratios. We prove that a classifier trained over these features performs no worse than a parameter invariant classifier created from a single statistic.

Finally, we apply parameter invariant classification to a number of challenging problems in the medical domain: detection of hypovolemia in critically ill patients, automated meal detection in diabetic patients, and detection of pulmonary shunts in infants. We demonstrate promising performance in each scenario.

Notes:

Ph. D. University of Pennsylvania 2016.

Department: Computer and Information Science.

Supervisor: Insup Lee; C. William Hanson III.

Includes bibliographical references.

OCLC:

980708776