117828228

Chronuitographic Fingerprints ofTwenty Salvia Species 525

wavelet, and the decomposition level was set to two. To remove irrelevant details, the 'Visu' criterion and the soft thresholding połicy were used. Fig. 3d shows the HS-GC-MS fingerprint of the volatile fraction of S. azurea after denoising by use of the wavelet transform.

Step 3: Alignment of Chromatographic Signals

The HPLC fingerprints obtained from the extracts of S. azurea and S. sclarea, shown in Fig. 4a, illustrate the peak-shift problem in HPLC data. The same is tnie for the HS-GC-MS fingerprints, so both sets of fingerprints require alignment. These sets were aligned by use of the CO W method, imple-mented in the freely available MATLAB toolbox [35]. Initially, for each fingerprint type, the target signal was selected on the basis of the correlation coefficient criterion. As target signal, that with the highest mean correlation coefficient among ali the considered signals was selected [36] and further used to align the remaining signals. Then, for each group of fingerprints, warping variables were optimized. With the HPLC fingerprints, each signal was divided into 120 sections and the slack variable was set to 5. With the HS-GC-MS fingerprints, the optimum alignment was obtained when signals were divided into 90 sections and the slack variable was set to 5.

tme (min]

Fig. 4. Two superimposed HPLC fingerprints of samples 1 and 9 (extracts of S. azurea and S. sclarea.) before (a) and after (b) their alignment using the correlation optimized

warping

To demonstrate the effect of signal alignment, let us consider the pair of chromatographic fingerprints shown in Fig. 4a. Although the samples of the herbal extracts studied tended to differ substantially in their Chemical composition, from Fig. 4b it is evident that, using the CO W method, two fingerprints were aligned satisfactorily.