Predicting a Binary Outcome with Dose-Response Data

• Activity Number: 522
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Yang Zhang*+
• Companies: University of Pittsburgh
• Keywords: classification ; SVM ; breast cancer ; dose response curve ; chemotherapy

Abstract

An in vitro test system to quantitatively assess the chemotherapy sensitivity of tumor cells has been developed and the assay results may provide clinically relevant information for treatment decision. We consider in this study predicting a binary clinical outcome in cancer patients by using the in vitro assays data. Traditionally, we first create a summary statistic, such as area under the dose-response curve of the in vitro data or the concentration of a drug at which a certain predetermined proportion (e.g. 50%) of growth is inhibited as compared to drug-free control. This summary statistic is then used to predict the clinical outcome using regression models. Such a two-stage approach often loses information from the original dose-response data and may be ineffective if this summary statistics is weakly associated with the clinical outcome. The relationship between the in vitro and in vivo systems is usually complex and the dose-response information at one or some particularly dosages may predict the outcome better. Here we propose a support vector machine-based method that uses the original dose-response data in an integrated and efficient manner.