A workflow for the modeling and analysis of biomedical data

The use of data mining techniques for the classification of shape and structure can provide critical results when applied biomedical data. On a molecular level, an object's structure influences its function, so classification based on structure can lead to a notion of functional similarity. On a more macro scale, anatomical features can define the pathology of a disease, while changes in those features can illustrate its progression. Thus, structural analysis can play a vital role in clinical diagnosis. When examining the problem of structural or shape classification, one would like to develop a solution that satisfies a specific task, yet is general enough that it can be applied elsewhere.; Before any analysis process can begin, however, the data must be transformed into an appropriate input. There are a wide variety of modeling methods to chose from, but ideally, one would like the transformation process to satisfy the following criteria: (1) capture the clinically- or biologically-relevant features, (2) be computationally feasible, (3) reduce the dimensionality of the data and (4) provide easily interpretable results. It is usually impossible to completely satisfy all four criteria with a single modeling method, so sacrifices must be made. In many fields, there are certain domain-specific methods that are commonly used to model data. While they might provide acceptable performance for the application in question, they may also be specific to a single researcher, making the comparison of results impossible.; In this work, we propose a workflow that can be used to model and analyze biomedical data, both static and time-varying. This workflow consists of four general stages: (1) Modeling, (2) Biomedical Knowledge Discovery, (3) Incorporation of Domain Knowledge and (4) Visual Interpretation and Query-based Retrieval. For each stage we propose either new algorithms or suggest ways to apply existing techniques in a previously-unused manner. Where appropriate, we compare against traditional or accepted standards. Not every technique will be appropriate or effective on all types of data, but we hope that researchers can use our work as a guide in order to determine the most appropriate modeling and analysis path for their particular application.; We present our work as a series of case studies that implement various stages of the above workflow. Through these case studies, we also propose and address a number of specific research questions. We show that generalized modeling methods can be used to effectively represent data from several biomedical domains. We detail a multi-stage classification technique that seeks to improve performance by first partitioning data based on global, high-level details, then classifying each partition using local, finegrained features. We create an ensemble-learning strategy that boosts performance by aggregating the results of classifiers built from models of varying spatial resolutions. This allows a user to benefit from models that provide a global, coarse-grained representation of the object as well as those that contain more fine-grained details, without suffering from the loss of information or noise effects that might arise from using only a single selection. Finally, we propose a method to model and characterize the defects and deterioration of function that can be indicative of certain diseases.