Medical Mining with Supervised and Semisupervised Methods

Classification models are widely used in a plethora of different applications to automatically assign objects into one of several pre-defined categories. In the context of Medical Mining, objects can be patients or study participants and the target outcome may be a disease or disorder under study. Here, the objects are often represented by multi-dimensional feature vectors and classification models are induced by learning associations between features and the medical outcome from a set of objects where the outcome is known. However, in real-world medical domains the objects can be complex and change over time, being described by various differently scaled features and background data containing additional information. To produce quality classification models here, relevant dimensions w.r.t. the class variable must be identified by utilizing methods that cater to the requirements of such objects while considering available background knowledge.
In our work we use labeled data, constraints on object similarity and historical records of patients / study participants to identify relevant explicit and implicit dimensions relevant to medical outcomes. We argue that current methods are not adequate in all regards for this task, inducing the need for new approaches:
 
Evolving objects are observed multiple times during their evolution. Traditional algorithms that identify relevant dimensions by using labeled training data cannot be applied. We therefore extend classical feature selection methods to handle evolving objects.
 
Sole consideration of labeled data to find relevant dimensions is not always practical. Such data may not exist or only in small quantities and considering additional background information regarding the objects under study may improve findings. We therefore develop methods that use constraints on the similarity of objects to substitute the need for labeled training data and find object-group specific relevant dimensions.
 
The evolution of objects described by their multiple observations can implicitly contain dimensions relevant to the classification task at hand. Omitting this dimensions can severely impede resultant classification model quality. We therefore develop strategies to derive dimensions from an object's evolution and develop a method to detect and codify relevant evolution patterns.