Supervised and unsupervised learning in biostatistics

点击翻译按钮获取中文摘要

As consequence of major technical advances in computing and in molecular biology, huge databases of biomedical interest are now available. In order to take advantage of the potential information they contain, new methods of data analysis are needed. One challenging aspect of the effort required, is that data are available at virtually all levels of observation: social, psychological, clinical, systemic and molecular. There is a growing need of synthesis and it is increasingly recognized that it is essential to link distinct levels of observation to improve understanding. The new methods of data analysis that seem to be appropriate for the task are considered intelligent, i.e. able to mimic as much as possible the thinking of human experts. They borrow substantially from the discipline known as machine learning, but with specific attention to problems of statistical inference and to practical needs, such as dealing in real time with huge data sets. Indeed, relatively new terms have been created in the last 15-20 years, such as data mining and statistical learning theory.In my work I attempt to fully exploit the parallelism between the cognitive strategies of an expert and some computer-based tasks of data analysis. Highlights of my current and future work include: 1) Tree-growing algorithms based on statistical models for increasingly complex data structures, e.g. data from multi-level systems. 2) Neural Network architectures based on statistical models. 3) Pattern discovery (clustering) for multilevel data and other complex data structures, both in an exploratory and in a model based mode. 4) Latent class models to explain complex dynamical patterns in observed variables via simpler dynamics of underlying conceptual entities.