Diagnostics of Complex and Rare Abnormalities Using Ensemble Decomposition Learning

Abstract

Diagnostics of complex and rare medical cases lacking clear symptoms of particular abnormalities is a very challenging problem. Case based reasoning (CBR) is known to provide helpful and generic guidelines for practitioners as well as for the design of medical expert systems dealing with such non-standard diagnostic problems. Single example learning (SEL) algorithms offer more formal machine learning framework for classification of rare and novel classes. However, both CBR and SEL approaches require significant number of wellstudied examples and a set of objective features capable to provide robust matching of novel examples with the previous ones. In practice, data sets for well-defined abnormalities suited for quantification with existing indicators are often limited. However, significant amount of valuable clinical information from cases labeled only as normal or abnormal without particular diagnosis remains underutilized. Recently, we have demonstrated that this information can be effectively employed to produce powerful normal-abnormal meta-classifiers using ensemble learning techniques applied to existing physiological indicators. This is achieved by an optimal weighted combination of complementary indicators which are experts in different regimes of the considered biological complex system. Therefore, partial information of wide variety of dynamical regimes becomes implicitly encoded in the obtained ensemble of classifiers, while only aggregated output is used. Extraction of this underutilized knowledge could be formalized in terms of ensemble decomposition learning (EDL) and used for representation of complex and rare cases in terms of intrinsic dynamical regimes. Such representation could prove to be more accurate and robust compared to traditional CBR and SEL approaches. Illustrative application of the EDL approach to cardiac diagnostics based on HRV analysis is presented and discussed.