Page 273 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

266 CHAPTER 13 Multiview Learning in Biomedical Applications

FIGURE 13.1
Example of multiview data related to clinical tests. (A) Magnetic resonance imaging;
(B) Electrocardiogram; (C) Blood test; (D) Blood pressure; (E) X-ray; (F) EEG.

describing the same phenomena. Bioinformatics and neurobiology are two prominent
examples of fields of application that deal with multiview data: for the same set of
samples, different kinds of measurements provide distinct facets of the same domain,
encoding multiple biologically relevant patterns. A model integrating different
perspectives or views of the data can provide a richer representation of the underlying
biological system than a model based on a single view alone.
The ability to combine different modalities assumes great relevance when trying
to unravel complex phenomena such as the physiological and pathological mecha-
nism governing the human neural system or the subtle molecular differences between
subtypes of diseases that are known to be the result of the interaction of several
causes.
To achieve this goal, several tools provided by different disciplines are needed,
including mathematical, statistical, and computational methods for the extraction of
information from heterogeneous data sources.

2. MULTIVIEW LEARNING
These days, researchers working with data integration favor mainly two kinds of ap-
proaches: the first methodology implies the combination of complementary informa-
tion coming from different sources describing complex phenotypes on the same set
of samples (multiview learning); the second strategy, known as metaanalysis, tries to
infer new knowledge integrating the information about the phenotypes of interest
with prior knowledge regarding already known phenotypes by means of comparative
methods. Fig. 13.2 reports a taxonomy of integration methodologies classified