Summary and Future Trends - Computational Intelligence in Biomedical Engineering - page 328

to work well and they remain useful primarily for two class problems (binary
classifier). If the problem is multiclass (three or more), NNs fare better in
terms of ease of use, but research is ongoing to extend the SVM formulation
to multiclass problems. A further issue is training speed, while training time
for NNs scale linearly with training set size, this is not the case for SVMs,
which have an almost quadratic slowing due to the quadratic programming
formulation. Nevertheless, research into decomposition techniques for more
ecient training of SVMs is proceeding.
When the measurements in a feature set are uncertain, rule-based methods
incorporating fuzzy logic have been applied, but this area of interest is rela-
tively new to the biomedical field and deserves further investigation. Hybrid
or expert systems for medical diagnostics have been previously proposed, but
there is an avenue for further improvement, especially in designing the best
rules or heuristics. These systems have proven useful as educational tools for
medical students and also as aids to physicians; again the research empha-
sis should be the accuracy of such systems. It could be beneficial to design
specific systems, rather than generalized medical systems, to take maximum
advantage of the selected CI technique and specific feature set.
As seen in the previous chapters, most core CI technologies have been
applied to diagnostic medicine to classify as healthy rather than pathology.
Efforts toward this goal have been successful, though some pathologies, for
example, neuromuscular disorders, require further investigation. The next step
is to apply these technologies to prognosis prediction and rehabilitation mon-
itoring where in prognosis prediction we envisage an automated CI system to
predict the probability of recovery. For example, if a patient had been diag-
nosed with cardiac arrhythmia, an important issue would be the chance of sur-
vival and how the disease would progress if untreated. Clinicians generally base
their prognosis on experience and there is left, therefore, considerable scope for
the application of CI techniques to improve the accuracy. In rehabilitation-
rate monitoring, we envisage the application of CI techniques to estimate
recovery rate due to a treatment, in an exercise intervention, for example,
to regain balance following a fall. An automated system capable of achieving
this would be useful in determining the effectiveness of drug treatment or
therapy.
Figure 8.1 outlines three areas in which we foresee research in CI applica-
tions flourishing. There are many challenges to be addressed but substantial
research is being undertaken in diagnosis and classification of pathologies.
In prognosis prediction, we require CI techniques related to time-series pre-
diction to estimate how well or how quickly a patient will recover. Research
in this direction will require physicians to record their patient's recovery to
generate examples of the required time-series data. Based on these data, it
should be possible to use CI techniques to predict a patient's future condition.
Fundamental questions that require attention are the accuracy of prediction,
the amount of data required, methods of performing the prediction, and index
measures or confidence values, which intuitively indicate the severity of dis-
ease. Accurate prediction would open up new ways of dispensing medication,