We accessed Google Scholar, Web of Science, and PubMed to identify publications for the period 1900–2022, with the search string “Lavandula angustifolia + essential oil”, “linalool”, “pharmacological effects”, etc. Also, we included in the search string “Bulgaria”, “China”, “India”, and “France”–the countries that are recognized as main world producers of lavender essential oil, as well as “Greece”, “Ukraine”, “Poland”, and other neighboring countries. Following the PRISMA 2000 guidelines, the records were assessed for eligibility, and the inappropriate ones were excluded. We selected 32 publications that presented the full components’ lists as a result of GC or GC-MS analyses of lavender (L. angustifolia) essential oil obtained by hydrodistillation (either industrially or using Clevenger apparatus for 40 minutes to 4 hours) from various counties and varieties, under different cultivation. Based on published results of various samples (each publication presented analyses of 1 to 19 samples) was created a data set in Excel (Kozuharova et al. 2023; Naef and Morris 1992; Venskutonis et al. 1997; Milhau et al. 1997; Adam et al. 1998; Ghelardini et al. 1999; Renaud et al. 2001; Shellie et al. 2002; Chakopoulou et al. 2003; Schmidt 2003; Afsharypuor and Azarbayejany 2006; Ihsan 2006; Babu and Singh 2007; Cong et al. 2008; Smigielski et al. 2009; Hassiotis et al. 2010a, 2010b; Verma et al. 2010; Wesolowska et al. 2010; Raina and Negi 2012; Adaszyńska et al. 2013; Prusinowska and Śmigielski 2014; Konaktchiev 2015; Zagorcheva et al. 2016; Lafhal et al. 2016; Nikšić et al. 2017; Luo et al. 2019; Détár et al. 2020; Dong et al. 2020; Chen et al. 2020; Gangoo et al. 2021; Pokajewicz et al. 2021; Cong et al. 2022). Initially, it comprised 99 samples. In total, 309 names of components were used in these publications. We checked for synonyms in the compound names and eliminated the duplications (compounds were alphabetically sorted). The records of isomers and coeluates were combined and generalized since the main limitation of the published data was the lack of homogeny. Eleven samples were removed from the data set due to the lack of consistency (for example, linalool and linalyl acetate quantities were much lower than the minimum of the standard), namely samples from the USA (6), Romania (1), China (1), and unknown locations (3). Also, the components that were not detected in more than 10 of the samples were removed. The final dataset prepared for the analysis consists of 88 lavender essential oil samples and 16 components (Suppl. material 1). Additionally, we included in our data set the values of the standards of the International Organization for Standardization (ISO) and European Pharmacopoeia (10^th edn) (Table 1) (Lis-Balchin 2002; ISO 3515:2002 2018; European Pharmacopoeia 2020).

Chemometric methods were used to mine the data set of dimensions [90 × 16] (88 experimentally tested samples of lavender essential oils along with minimum and maximum values of the standards; see above) and subject it to multivariate statistical analysis. The lavender essential oil samples are from different geographical locations. Ukraine (UA, n=30), Bulgaria (BG, n=14), Poland (PL, n=10), France (FR, n=7), India (IN, n=6), Greece (GR, n=6), Italy (IT, n=3), Australia (AU, n=2), China (CN, n=3), Jordan (JO, n=2), Iran (IR, n=1), Hungary (HU, n=1), Lithuania (LT, n=1), Moldova (MD, n=1), United Kingdom (UK, n=1), and Bosnia and Herzegovina (BA, n=1) were analyzed for the content of 16 important chemical components.

The major goal of the chemometric study was to reveal hidden relationships between the objects of the study (partitioning with respect to the geographical location) and between the variables characterizing the objects (chemical descriptors). An additional task to the partitioning procedure was the determination of the specific descriptors for each group of similarity found in the partitioning procedure.

The multivariate statistical methods used for the data mining were hierarchical and non-hierarchical clustering (K-means clustering) and, additionally, factor analysis. All of the methods are well-known and described fully in the literature (Massart et al. 1987, 1998). In general, cluster analysis aims to find patterns of similarity (clusters) in a large dataset. It could happen by spontaneous clustering without preliminary conditions (unsupervised approach, hierarchical clustering) or by a supervised approach where the number of clusters is a priori determined by experts or by checking a preliminary hypothesis (K-means clustering as a supervised method). The clustering methods follow specific algorithms such as data standardization, choice of similarity measures, linkage of the similar objects into clusters, and check of the cluster significance. In hierarchical clustering, the output is a graphical plot called a dendrogram, since in K-means clustering, the results of partitioning are presented as tables describing the membership of each object or variable.

The first step of our analysis was to build the hierarchical dendrograms for linkage of the objects (Fig. 1) and variables (Fig. 2) into patterns of similarity. The exact membership of each object is presented for convenience in membership tables (Tables 2, 3) as the output of the K-means clustering (it is worth noting that the clustering results for both methods are very similar). On both dendrograms, it is quite obvious that the objects are partitioned into four groups of similarity (clusters). The chemical descriptors are separated into four clusters (Fig. 1). Also, four clusters of objects (locations of origin of the lavender essential oil samples and standards’ minimum and maximum values) are formed (Fig. 2).

Figure 1. 

Hierarchical dendrogram for clustering of 16 chemical variables–the main components of the lavender essential oil.

Figure 2. 

Hierarchical dendrogram for clustering of 90 objects (88 samples and two standards).

As already mentioned above, a priory hypothesis, along with expert opinion, required partitioning of the variables and the objects into 4 clusters for each category. The members of each cluster of objects and variables are presented in Tables 2, 3. The results from K-means clustering enable the interpretation of the four clusters and the discussion of the membership of each variable.

The four clusters of variables (Fig. 1, Table 2) give an indication of the factors responsible for the explanation of the dataset structure. It becomes possible to suggest that the important conditional components of the structure are the total explained content of the Lavandula angustifolia essential oil being correlated to 1,8-cineole + phelandrene, D-limonene, Z-β-ocimene, linalyl acetate, and 3-Octanone as components in lower concentrations; another two factors are related to camphor, 1,8-cineol, and ocimene derivates as aromatic components; and one factor for components in higher concentrations as a macrocomponent factor. Interestingly, the components linalool and linalyl acetate, responsible for the specific pharmacological effects of lavender essential oil and characterizing its quality, fall in different clusters, namely Cluster Number 1 and Cluster Number 4 (Table 2).

It is of substantial interest to find out which descriptors are specific to each one of the identified clusters. The plot of averages for each variable for each identified cluster is presented in Fig. 3. For better interpretation of the plot, below is the sequence of all variables (on the plot, 8 out of all 16 variables are presented). The results for specific descriptors of the four clusters are presented in Tables 3, 4.

Figure 3. 

Plot of means for each descriptor for each identified cluster.

The four clusters of objects (locations of origin of the lavender essential oil samples and minimum and maximum values of the standards, Fig. 1, Table 5) inform some specific geographical dependencies related to Lavandula angustifolia essential oil quality.

Cluster Number 1 contains 32 members (Table 5), 14 of which are lavender essential oil samples with Bulgarian origin (100% of the Bulgarian samples), and three samples of French origin (43% of the French samples), three samples of Ukrainian origin (11% of the Ukrainian samples), two samples of Australian origin (100% of the Australian samples), two samples of Indian origin (33% of the Indian samples), two samples of Greek origin (33% of the Greek samples), one sample from China (33% of the Chinese samples), and three other single samples as follows - from Bosnia and Hercegovina, (the only one from this location), from Italy (33% of the Italian samples), and from Poland (10% of the Polish samples). It is important to note that this cluster includes the variable “standard max” (the maximum values). This is an indication that the Bulgarian samples, as well as the other listed samples, are closely bonding to the maximum values of the standard requirements. This cluster could be a conditionally named “Bulgarian” pattern.

Cluster Number 2 includes 41 members and is the biggest one (Table 5). It includes predominantly samples from Ukraine (23 out of a total of 41 samples, or 56%). This is also 82% of all Ukrainian samples. Next are four samples of French origin (57% of the French samples), followed by four samples of Indian origin (67% of the Indian samples), two samples from Poland (20% of the Polish samples), two samples from Italy (67% of the Italian samples), two samples from China (67% of the Chinese samples), two samples from the United Kingdom (100% of the UK samples), and two single samples (each of them, the only one from this location) as follows from Hungary and Lithuania. Again, a standard sample (the minimum values) belongs to this cluster, and it allows the assumption to name this cluster the “Ukrainian” pattern conditionally.

Analogically, Cluster Number 3 consists of 10 members (Table 5) and represents a “Polish” pattern since the samples from Poland are seven (out of a total of 10 samples from Poland, or 70% of all samples in this cluster and 70% of all Polish samples). It is not a surprise that the remaining three samples are from Ukraine (10% of the Ukrainian samples), which is a geographical neighbor of Poland.

Finally, Cluster Number 4 is the smallest one with only seven members (Table 5), with dominantly Greek samples (4 out of total 7, and 66% of all Greek samples), and the cluster could be a conditionally named “Greek” pattern.

Factor analysis (often named principal components analysis, or PCA) is a typical projection method. It helps to clarify the dataset structure by reducing variables and replacing the initial variables with latent variables (factors). The new factors are characterized by factor loadings (indicating relationships between the initial variables) and factor scores (representing the new special coordinates of the objects). The output is usually a specific table with factor loadings, which helps to interpret the physical meaning of the latent factors. In Table 6, the factor loadings for four factors are presented.

The first latent factor explains nearly 22% of the total variance of the system. It indicates the role of lavandulol, lavandulyl acetate, and terpinen-4-ol on the data structure and could be a conditionally named “terpene effect” factor, which characterizes Polish samples (Table 6).

The second latent factor explains another 18% of the total variance and is related to the high loadings of camphor and 1,8-cineole being very important characteristics of the lavender oil. The Greek and some Mediterranean samples are related to this conditional “Camphor” factor (Table 6).

The next latent factor, with an explanation of almost 17% of the total variance, is related to the special derivates of ocimene and limonene. The third latent factor could conditionally be named an “aromatic” factor (Table 6).

The last latent factor 4 (about 14% of the total variance) indicates the significance of α-terpineol and linalyl acetate as major components to form the total content of the sample quantity; conditionally, this factor could be known as the “macro components“ factor (Table 6).