? This SBIR aims to produce next generation classification and regression software based upon ensembles of decision trees: bagging, random forests, and boosting. The prediction accuracy of these methods has caused much excitement in the machine learning community, and both challenges and complements the data modeling culture prevalent among biostatisticians. Recent research extends the methodology to likelihood based methods used in biostatistics, leading to models for survival data and generalized forest models. Generalized forest models extend regression forests in the same way that generalized linear models extend linear models. ? ? This software would apply broadly, including to medical diagnosis, prognostic modeling, and detecting cancer; and for modeling patient characteristics like blood pressure, discrete responses in clinical trials, and count data. ? ? Phase I work will prototype software for survival data, and investigate the performance of ensemble methods on simulated and real data. For survival applications, we will assess out-of-bag estimates of performance, and investigate measures of variable importance and graphics that help clinicians understand the results. Experience writing prototypes and using them on data will lead to a preliminary software design that serves as the foundation of Phase II work. ? ? Phase II will expand upon this work to create commercial software. We will research and implement algorithms for a wider range of applications including generalized forest models, classification, and least squares regression. We will also implement robust loss criteria that enable good performance on noisy data, and make adaptations to handle large data sets. ? ? This proposed software will enable medical researchers to obtain high prediction accuracy, and complement traditional tools like discriminant analysis, linear and logistic regression models, and the Cox model. ? ?
