117828226

Chromatographic Fingerprints ofTwenty Salvia Species 523

loss of its information content, thus considerably reducing the computa-tional effort necessary to preprocess the signals. Such a signal preprocessing is, therefore, frequently used before signal alignment, for instance using the COW algorithm [25]. On the other hand, when chromatograms require de-noising, linear interpolation to adjust their lengths cannot be used, because this operation affects noise characteristics.

Corning back to processing our collection of the HPLC fingerprints, it is relevant to stress that ail have a satisfactory signal-to-noise ratio. Therefore, noise elimination proved unnecessary, yet the fingerprints were of different length. Bearing in mind a satisfactory signal-to-noise ratio, fingerprint lengths were adjusted using linear interpolation. After this transformation, the longest chromatogram (chromatogram 12; extract of S. triloba) was com-pressed by a factor of approximately 3.6 (from 10 859 to 3000 sampling points). As shown in Fig. 2b, important signal details relevant to the problem studied were still present after compression.

Brnę (min)    Bme (mm)

Fig. 2. HPLC chromatogram obtained from S. triloba extract (sample 12) before (a) and after (b) linear interpolation reducing the number of sampling points from 10 859 to 3000

The objective of another type of initial fingerprint preprocessing is to remove uninformative regions from the signals. For example, peaks that are not expected to explain variability among the samples should be removed. An uninformative region between 0 and 8.5 min, where the predominant signals were from carbon dioxide and oxygen, was therefore removed from all HS-GC-MS fingerprints.

Step 2: Enhancement of Signal-to-Noise Ratio for Chromatographic Fingerprints

In the second preprocessing step, we concentrated on enhancement of the signal-to-noise ratio of the herbal fingerprints. To remove the background from these fingerprints and to suppress noise, the penalized asymmetric