Ambient Diagnostics - Ambient Intelligence for Scientific Discovery - page 242

Energy, Entropy, Crack Index, and Fractal Dimension. We plan to add six more
in the near future. In addition, we will add TCM expert's verbal descriptions
into the model. We need to devise a scheme to map these qualitative descriptions
into fuzzy sets of quantitative values so that computers can process them.
As comparing one patient's tongue to another patient's tongue is dicult in
terms of shape, color and texture registration, we found that it is more accurate
to compare the tongue images at a personal basis, e.g. the images of a cancer
patient before and after chemo, etc. It could be used as a measurement of the
effectiveness of a treatment. This is similar to the self-monitoring process that has
been used in the head injury study with fMRI personal data. For each person,
we can establish a base line and then use ubiquitous computing technologies
to monitor the patient and compare the collected data to the base line at a
predefined duration.
Neural computing is a promising method for generating diagnostic results
and potentially can be hardened onto a chip that is less than one dollar. We use
two neural networks for testing the concept: a general regression neural network
(GRNN) and a probabilitistic neural network (PNN). As the two test scenarios
show, both have virtually the same performance. Also we learned that neural
networks work better when the target classes are fewer, e.g. in our cases, two
targets are better than three targets. Although PNN and GRNN can correctly
classify 2 Polyps test cases out of 3, it should not be implied that the methods
would work well with a larger sample size. We still need more work to verify the
selectivity and reliability of the models.
Visualization utilities are helpful in early stage data explorations. At this
stage, the aim of visualization is to provide tools to aid the analysis, rather
than to provide a precise proof of clustering decisions that can be provided by
statistical and data mining techniques. We have not yet coupled visualization
techniques with fuzzy algorithms. The 3D clustering figures were actually an
animation that allows viewing from different angles with different triplets of
variables. Clusters are observed in some cases, while in other cases, we could not
see clusters. It is also not possible to see this effect when viewed as a figure on a
flat piece of paper. The parallel coordinates approach provides multi-dimensional
visualization. Although the variable values may overlap, as clusters, they do have
distinct characteristics.
As more tongue data is available, we will be able to use more advanced
techniques for visualization and classification of tongues to obtain better analysis
and more accurate classification. We have received letters from a few colon cancer
patients who have had 'geographical tongues'. Those feedbacks are significant
resources for investigation. The combination of the tongue feature changes and
other medical indications will increase the effectiveness of early diagnosis. We
will continue to follow up the cases.
Affordable self-diagnostic kits are changing our lives. Decades ago, diabetes
patients had to check their blood glucose in a lab or burn the sample with a
candle. Today, they can buy a digital kit from a drugstore and test anywhere.
Modern electronic technologies have been the building blocks for eDiagnostic