Affiliations: [a] Department of Bioinformatics and Biostatistics, University of Louisville, Louisville, KY, USA | [b] Department of Epidemiology and Department Health, University of Louisville, Louisville, KY, USA | [c] Department of Anatomical Sciences and Neurobiology, University of Louisville, Louisville, KY, USA | [d] Biostatistics Shared Facility, James Graham Brown Cancer Center, University of Louisville, Louisville, KY, USA
Correspondence:
[*]
Corresponding author: Shesh N. Rai, Biostatistics Shared Facility, James Graham Brown Cancer Center, University of Louisville, 505 South Hancock Street, Room 211, Louisville, KY 40202, USA. Tel.: +1 502 852 4030; Fax: +1 502 852 7979; E-mail: [email protected]
Abstract: Phase III clinical studies are most often powered to detect an overall
difference in response to treatment between two treatment arms. However in
many cases, any response to treatment is restricted to a subset of patients.
Therefore, traditional randomized clinical trials with broad eligibility
criteria may result in missing effective treatments. If predictive assays
have been developed beforehand to accurately identify patients who are
likely to benefit, they should be used. However these are often not
available before a Phase III study. The Adaptive Signature Design is an
interesting approach that addresses this situation. In order to lay a
framework for potential parametric extensions to this design, parametric
classification methods are first reviewed, with focus on distances specific
to each classification method. Modification of these methods for treatment
subset prediction is then discussed.
Keywords: Classification, machine learning, dimension reduction, interaction, clinical study
