EAS Doctoral Dissertation Defense by Manoj Thakur
Abstract:
In a clinical trial, to assess the safety profile of the treatment under investigation, safety tables are presented that compare the proportion of patients (subjects) experiencing distinct adverse events(side effects) between the two or more than two groups.
A patient may experience multiple adverse events after taking the treatment under investigation. Therefore, the data for several adverse events (AEs) are collected on each subject, and the response scale is "Yes" or "No" to having a specific adverse event. Consequently, multiple binary response variables (adverse events) are assessed, each on the same subject, leading to multivariate correlated binary data.
Researchers are interested in comparing the overall safety profile of the experimental group (treatment group) with the control group. When the global null hypothesis of equality of the two vectors of population proportions for the treatment group and the control group is rejected, researchers are also interested in knowing which specific adverse events or set of adverse events caused the difference.
In current practice, the proportion of subjects with a specific adverse event or any adverse events between two or more than two groups is compared using descriptive summary statistics or a chi-squared test, or an exact test. There are at least two issues with the current practice of analyzing safety data. Firstly, the univariate analysis ignores the correlation between the adverse events. Consequently, the precision of the estimates and the power may be lower than that achieved by other approaches. Secondly, the univariate analysis does not compare the vectors of proportions between the two groups. However, it compares the proportion of patients who experience each distinct adverse event between the two groups separately and does not take into consideration each adverse event together. Consequently, it may fail to shed light on the overall safety outlook of a new treatment and lead to an erroneous decision.
In this dissertation, to evaluate the overall safety profile of the treatment under investigation, multivariate approaches are presented to analyze the safety data that incorporate the correlation among responses. However, these approaches are getting more computationally demanding and impractical as the number of adverse events increases.
The random forest (RF) and LambdaMART algorithms are proposed to handle a large number of adverse events to analyze multivariate correlated binary data, as the existing multivariate approaches are getting computationally intensive and unfeasible to obtain the parameter estimate due to a large number of adverse events.
The random forest and LambdaMART are also used to rank the adverse events. In addition to RF and LambdaMART, the adverse events are also ranked based on the Food and Drug Administration's (FDA) guidance and the weighted average of the proportion of subjects who experienced a specific adverse event between the two groups. The ranking of adverse events is used to assign weight to adverse events according to their ranking to perform the weighted score-type test.
The ranking of adverse events may help pharmaceutical sponsors to decide which adverse events need to be included in the package insert. In addition, it will also be helpful in precision medicine to prescribe safer drugs or vaccines to the right patient. The multivariate approaches will be useful for comparing the overall safety profile of two treatment groups and for interpreting the composite binary endpoints. In addition, it will facilitate to data monitoring committee to assess the overall safety profile of the treatment under investigation and make a correct recommendation to the sponsor on whether to continue, modify, or stop the study due to the safety profile of the treatment under investigation.
In this dissertation, we also proposed a modified truncated Hochberg procedure to control the global family-wise error rate (FWER) for hierarchical structure multiple endpoints.
Safety data collected from the Phase II vaccine study and the Phase III drug study are used to illustrate the proposed and existing methods.
Advisor: Dr. Donghui Yan, Department of Mathematics
Committee members:
- Dr. Hongkang Xu, Department of Accounting & Finance
- Dr. Ming Shao, Department of Computer & Information Science
- Dr. Yuchou Chang, Department of Computer & Information Science
Note: All EAS Students are ENCOURAGED to attend.
TXT 105 : Virtual - Please contact Dr. Donghui Yan (dyan@umassd.edu)