7 Steps to a Ready to Use Spectroscopic Diagnostic System

Create a new Knowledge Base and enter an arbitrary name for it. Done!
The system will save all the spectra as well as your settings in this Knowledge Base. This is done automatically, so you never lose your data.

Organize your spectra! Load the spectrum files manually, or use a directory tree to import complete directories of spectrum files - simply by Drag&Drop. If you have a catalog of your data in the form of a spreadsheet table at your disposal, you may import this table and have the listed spectra searched and loaded automatically.

Create the different classes that you wish to distinguish and assign the appropriate class to each spectrum. If you use the tools for importing spectra by directory tree or by table as mentioned above, the assignment of classes to spectra is done automatically by a simple mouse click.

You may decide to declare a number of spectra "unused". These will not be used during construction and training of the diagnostic system, so they can be used later to test the prediction performance of the system with "unknown" data.

The software PATTERN EXPERT airspect provides extensive tools for visualizing your spectra. View the spectra one by one or in groups. Examine the mean curves of all spectra belonging to the same class as well as the corresponding standard deviations. All that may be displayed in one common diagram, in separate diagrams, or in a perspective view. Or, you may use the "Gel View" to look at your spectra from above. No important detail will remain unnoticed!

Configure the various preprocessing steps which are necessary to make your spectra comparable. You may activate and configure a truncation of border areas, a smoothing operation, a correction of discontinuous jumps, a baseline correction, a rescaling of the Y-values to a common range, and a peak alignment. A preview window gives you a visual impression of your settings' effects. The settings will be applied to all spectra in the same manner.
Finally, you may select the spectral regions to be used for analyzing and classifying your spectra. Or you leave the default setting unchanged, which uses the entire spectral interval - the system will find and utilize relevant regions automatically during the training process!

This part is the heart of your system. Here, you have the intelligent analysis methods - the so-called "Recognition Modules" - constructed and trained. For this, learning Pattern Recognition algorithms are used, which are preconfigured and optimized for your current application.

At first, a set of spectral intervals, which might be useful for solving your classification problem, is determined fully automatically. This "Feature Computation" process is based on a script, which you may select freely from a list of prefabricated and well documented scripts. The next step, which can be configured via script selection as well, is the "Feature Selection". Here, it turns out, which of the spectral features determined before are actually useful and necessary for distinguishing the classes. Among the procedures selectable for this step are feature scoring methods as well as evolutionary algorithms. The final step - "Classifier Construction" - uses the selected spectral features to construct and train a well-adapted classifier. From now on, this classifier can be used to determine the proper class for spectra of unknown samples.

The software PATTERN EXPERT airspect offers extensive tools for analyzing the spectral intervals, which have been found to be useful and relevant: A feature list containing these intervals can be displayed and exported to other programs, or the intervals can be inspected visually in a graphical presentation.
Moreover, there are tools to discover "suspect" spectra, which might be of low quality, or which have been erroneously misassigned by the user.
The software displays quality measures for the created Recognition Modules, which represent estimates of the percentage of correct class assignments.

In this part of the software, you use the knowledge, which has been extracted from the known examples during the previous step, to assign classes to your unknown spectra.

If you have made use of the option to declare some of your spectra "unused" at the beginning, so that they have not been used in any of the processes of the previous step, you can have these spectra classified now. The software displays the system's class decision for each spectrum and additionally a confidence value for each decision. This class decision is automatically compared to the known actual class. In this way, you receive another measure for the system's reliability, which is based on a truly independent test set.

The main function of the "Recognition" part is the identification of the proper classes for new spectra of really unknown samples. Again, the system's class decision is supplemented by a meaningful confidence value. In addition, there are tools for comparing the unknown spectra to the known examples of your Knowledge Base and for searching the Knowledge Base for the most similar reference spectra. The "Recognition" part is the place for utilizing the knowledge extracted from the training spectra for practical applications.

This step is not mandatory for the construction of a working diagnostic system (or "Recognition Module"). It may be performed optionally in order to obtain a realistic estimate of the system's future classification correctness.

The External Cross Validation represents a simulation of "real-life" conditions. Cyclicly, the software creates a complete Recognition Module with some of your spectra excluded from all steps, uses the finished module to classify the excluded spectra, and then checks the correctness of the individual class decisions. This cycle is repeated several times until, finally, all your spectra have been treated as unknown test data. In this way, the system's behavior on completely unknown spectra is tested.
The complete process is performed fully automatically. All the results are presented graphically.