Biomedical Engineering Reference
In-Depth Information
potentially alter the routine treatment of data by combining several steps. A
single transform [78, 79] can act to determine dimensionality of the data, com-
press manyfold the data size for analysis and storage, eliminate background
effects, remove spikes [80], de-noise [81], and perform classification [82]. Ex-
tensive application of the method has been reported for spectral analyses
[83]. An attractive feature of the approach, in addition its universal ap-
plicability and utility, is the capability to examine data at different lev-
els of granularity [84]. The selection of a specific component allows for the
examination of fine or coarse structure in the data - both spectral and
spatial. The drawbacks discussed for PCA above also remain for wavelet
transforms.
8.2.7 Dimensionality Reduction
Spectra have significant regions of redundant information, spectral regions
that may not be useful in classification or regions that have no spectral infor-
mation content. Hence, a subset of recorded spectral regions is useful and the
process of finding this subset is termed as dimensionality reduction. The first
approach is to employ a completely objective and automated technique. For
example, regression techniques can be used to determine which indices are
most effective at explaining the relationship between recorded data and the
predicted result. Yet another means are factor-based approaches (e.g., PCA)
since most of the important information is captured in a few components. A
third approach may be to use expert algorithms (e.g., genetic algorithms) that
integrate well with the data analysis. A comparison of various approaches is
available against newer ideas [85, 86]. Last, a completely manual approach,
in which the spectroscopist examines spectra for important features, may be
used [87]. In general, the more automated the method, the more compre-
hensively it can examine data. Similarly, when the data reduction method
includes a human expert element, an enhanced understanding of the under-
lying biochemical knowledge can be obtained and prior knowledge can be
incorporated.
8.2.8 Summary of Pre-processing Methods
Pre-processing methods are key to understanding the true scientific content
of the data structure and are critical to enabling accurate, fast, and robust in-
formation extraction from the methods we will describe next. The purview of
such methods has expanded from simple spectral corrections to sophisticated
algorithms that assure and improve data quality, provide consistency for sub-
sequent processing, and optimal retrieval and storage of data. We emphasize
that these steps are crucial and necessarily integrative with the entire chain
of designing experiments to analyzing results. Hence, the practitioner must
exercise due care to use the methods carefully and understand implications of
these operations on results obtained.
Search WWH ::
Custom Search