The image dataset was used to test the ability of each of the projection methods to predict the correct image ordering, based on the size increment of the brain. Since the largest source of variability separating each image in the series is the increase in feature surface area, as a result of the head size, only the first eigenfunction for each method was used in the comparison. This calculation reduces the dimensionality from ℝ^16,384 to ℝ¹. To assess the accuracy of each method, a non-parametric measure of association (Spearman's rho coefficient) was used, by which the scores from the primary eigenfunction were ranked and correlated against the correct ordering. A straight-line fit is indicative of perfect image ordering.

Both the kernel PCA and the spectral method from Ng et al. require fitting for the smoothing epsilon term to optimize performance for the dataset. The results for this parameter optimization are provided in Figure 1. The maximum rho coefficient possible indicates the appropriate value for this epsilon term for each method. Based upon the observation that neither line reaches a maximum value of 1, it is apparent that neither method is capable of determining the correct ordering of all of the images.

The remaining methods: PCA-correlation, PCA-covariance, nonmetric MDS, and the spectral method from Lafon 34 do not require parameter fitting that is dependent upon performance (as is necessary for the previous two nonlinear methods), so the images can be directly rendered into a low dimensional representation.

Table 1 and Figure 2 illustrate the results for all of the projection methods. PCA-correlation has the lowest coefficient (rho = 0.902). The plot (Figure 2a) demonstrates that the global size change is minimal after image 20 and the subtler differences in shape are not picked up by this method. This method thus fails to predict the correct progression between the images after this point. Both PCA-covariance and nonmetric MDS (Figures 2b and 2c, respectively) show exactly the same ability to solve for the correct ordering of images (rho = 0.966). Since these two methods give the same results, it is implied that the ordering of dissimilarity values (as fit with classical MDS) and the rank of the ordering of dissimilarity values (as fit with nonmetric MDS) are identical, which implies that there is no benefit in using nonmetric MDS to recover the image ordering. These two methods fail to retain the ordering after image 22, although the deviations at this point are not as drastic as those observed for the PCA-correlation results. The two other nonlinear methods (Figures 2d and 2e), each of which requires parameter optimization, predict the image ordering more correctly than do the two linear methods, indicating that this dataset is not best summarized with linear methods. Where there are meaningful local relationships, or nonlinearities, that the linear methods fail to preserve in a low dimensional mapping, nonlinear methods will be a more appropriate analysis choice. However, neither kernel PCA nor the spectral method from Ng et al. accurately preserves the correct image ordering over the entire series (rho = 0.989 and rho = 0.980, respectively). In this study, only the spectral method from Lafon was able to correctly solve the implicit ordering of the complete set of images (Figure 2f). This spectral method shares the properties of nonlinearity with kernel PCA and the spectral method from Ng et al. (which is initially anchored on the transformation of the Euclidean distance to some form of Gaussian kernel), however, it does not require parameter fitting of the epsilon term in order to produce optimal performance. Instead, the minimum non-zero squared distance is calculated for the smoothing term (for this example, ε = min_i≠j ||x_i - x_j||² = 140,245), meaning that this method is dependent only upon the distribution of squared Euclidean distances. The image ordering as produced by both PCA-correlation and the spectral method from Lafon is provided in Figure 3.

The microarray dataset was used to evaluate the ability of the projection methods to accurately classify samples derived from three taxonomically separate species into their respective groups, without conducting any preliminary feature selection (a procedure usually conducted in order to better partition the groups). In each case, the results of a method were calculated such that dimensionality was reduced from ℝ^12,625 to ℝ². Thereafter three classifiers using leave-one-out cross-validation (LOO-CV) were calculated on this projected space, both to assess the classification accuracy for each method and to compare the differences in value of the optimal parameters for kernel PCA and the spectral method from Ng et al. A nonlinear classifier, k-nearest neighbors (KNN), was calculated, setting k = 2 and k = 3; two settings of k were used since each DR method renders the groupings differently, thereby favouring two nearest neighbors for some methods and three nearest neighbors for others. The average error rates were computed across 1,000 trials, to account for the variation arising from ties broken at random in the assignment of the nearest winning class which results in slightly different classification results for subsequent trials. In the event that the mean error rate is calculated to be greater than zero, even when one or more of the trials provide a smaller error rate, a range is denoted in the results (e.g. 0%–4.66%) to indicate that the occurrence of a lower error rate is possible. Another commonly utilized classifier, QDA, was calculated, as well as the linear classifier counterpart, LDA, to allow comparison of the dependence of the results of different classification methods on the tuning parameter fitting in the first DR step.

As was done with the image data, optimal parameters were determined for kernel PCA and the spectral method from Ng et al. These values were evaluated with each classification algorithm separately. In determining the most appropriate epsilon value for optimal classification accuracy, the optimal value for the spectral method from Ng et al. varies according to the classifier used (Figure 4a). There is a small window at a value of ε ~20,000 in which both QDA (blue line) and LDA (red line) reach respective minimum total classification error rates of 2.22% and 15.56%. However, the KNN classifiers for k = 2 (black line) and k = 3 (green line) require slightly higher epsilon values (21,000 indicated by a dashed vertical line) to reach their respective minimum total classification error rates (2.22%–13.93% and 8.88%–13.79%, respectively). It can be argued that increasing the trials of the KNN classifier might better adjust this minimum point in the two KNN curves, where it coincides with the QDA and LDA window for minimum error. However, without calculating error rates with three separate classifiers, but rather independently determining the value for a given classifier, this window would be unknown, in which case different optimal parameter values for each classifier would be suggested. For example, using LDA as a classifier to determine the minimum error rate, a value of ε = 20,000 can be chosen as the optimal parameter for the spectral method from Ng et al.; however, for a KNN (k = 3) classifier in the same example, this parameter value would not fall within the range of the minimum error rate. Instead, a value of ε = 20,000 would provide a local minimum error of 14.08%, as compared to the global minimum error of 13.79%.

The parameter optimization for kernel PCA shows similar trends to the optimal values from the spectral method from Ng et al. However, the differences in optimal parameter solutions between the KNN classifier and both QDA and LDA is much more pronounced with this DR method. For the QDA classifier, the epsilon value is optimized to provide a minimal error rate (0%) in the window of ε = 67,000 – 68,000, while for the LDA classifier, the epsilon value is optimized to provide a minimal error rate (2.22%) in the window of ε = 73,000 – 92,000. In addition, the epsilon values that provide the minimal KNN error rates for k = 2 (2.22%–6.79%) and k = 3 (8.56%) are at ε = 174,000 (indicated by a dashed vertical line in Figure 4) and ε = 57,000 (indicated by a solid vertical line in Figure 4), respectively. This result from kernel PCA is consistent with those obtained from the spectral methods of Ng et al., and demonstrates that each of these two nonlinear approaches have a dependence between the outcome of the classification algorithm and an appropriately optimized parameter. In addition, when comparing training set classification results to LOO-CV, the optimal parameter values are drastically different for both kernel PCA and the spectral method of Ng et al. More importantly, the parameter selection is completely dataset-dependent. Note that here the scale is vastly different from that seen in the image example, thus the optimal parameter is as well. For these examples we examined two very disparate data types, but the same conclusion of dataset-dependence would almost certainly occur if two microarray datasets were compared, since there would still be differences in distance distributions (the possibility of two microarray datasets having identical distributions of Euclidean distances is highly unlikely). In other work we have examined additional microarray datasets and examined functional sub categories as well as disease state, cited here for those who are interested 5.

All individual classification results are summarized in Table 2, and the two-dimensional projections for each method are shown (Figure 5). For the kernel PCA and spectral method from Ng et al., the projection plots were generated with the epsilon terms optimized for QDA (ε = 67,000 for kernel PCA and ε = 20,000 for the spectral method from Ng et al). From the total error rate results reported in Table 2, it is apparent that the nonlinear DR methods of kernel PCA and the spectral method from Lafon perform more accurately than do the two linear methods (and nonmetric MDS) across all three of the classification algorithms (though using KNN with k = 2, PCA-covariance and nonmetric MDS can achieve a minimal error rate of 2.22%, similar to kernel PCA). Of these three nonlinear DR methods tested (not including nonmetric MDS), the spectral method from Lafon 34 outperforms both kernel PCA and the spectral method from Ng et al. Not only does the Lafon spectral method project the different species into well-partitioned groups (Figure 5f) for a 0% error rate across all classification algorithms (excepting a single misclassified sample with the LDA classifier), but the KNN classifier does not exhibit any deviation in classification results across 1,000 iterations, unlike the results obtained with any of the other methods. These results, in addition to the property that tuning parameter optimization is only dependent on the distribution of squared Euclidean distances (for this example ε = min_i≠j ||x_i - x_j||² = 1.29 × 10⁹, indicates that the spectral method from Lafon has significant advantages in tuning parameter fitting as a nonlinear DR method by our two criteria.

The image data was obtained from the Computer Vision Laboratory at the University of Massachusetts at Amherst 26. A total of 54 slice-by-slice CATSCAN images were obtained for the human head, where each image has dimensions of 128 × 128 pixels. The average for each image was calculated and the median of these averages was determined to be 11.82. Each image was then scaled to a target mean of this value.

The microarray data are from genomic primary fibroblast cell lines 2728 and were generated with Affymetrix oligonucleotide hgu95v2 arrays for 46 samples: 23 human (Homo sapien), 11 bonobo (Pan paniscus), and 12 gorilla (Gorilla gorilla) donors. This is a publicly available dataset within the 'fibroEset' package in R 29. It should be noted that two identical human donor arrays are in this dataset, so one was removed, reducing the dataset to 45 total samples.

The data was provided in R already normalized by the Affymetrix GeneChip MAS 5.0 algorithm. "Normalization was done by calculating multiplicative scaling factors on the basis of the median intensity of the 60th to 95th percentile of gene-expression scores" and intensities were floored to 100 fluorescent units 29. No further filtering or scaling was conducted on this dataset, which consists of 12,625 expression points for each of 45 arrays.

The spectral implementation of the weighted graph Laplacian from Lafon 34 is calculated as follows: Given a set of points X = {x₁, x₂, ..., x_n} ∈ ℝ^q, let G = (E, V) be a graph with edge weights or lines E, between pairs of vertices V. Consistent with standard terminology from graph theory, we can construct a graph, where each pair of vertices (x_i, x_j ∈ V(G)) is assigned a weight specific to the distance between them 30. The matrix, K₁ is calculated from these edge weights by a Gaussian kernel estimate

K1=K1(G)=(e−(‖xi−xj‖2/ε)) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGlbWsdaWgaaWcbaGaeGymaedabeaakiabg2da9iabdUealnaaBaaaleaacqaIXaqmaeqaaOGaeiikaGIaem4raCKaeiykaKIaeyypa0JaeiikaGIaemyzau2aaWbaaSqabeaacqGHsislcqGGOaakdaqbdaqaaiabdIha4naaBaaameaacqWGPbqAaeqaaSGaeyOeI0IaemiEaG3aaSbaaWqaaiabdQgaQbqabaaaliaawMa7caGLkWoadaahaaadbeqaaiabikdaYaaaliabc+caVGGaciab=v7aLjabcMcaPaaakiabcMcaPaaa@49D3@ if i ≠ j, where K_1ii = 0 and ε = min_i≠j ||x_i - x_j||² > 0

The epsilon term is chosen at the minimum squared Euclidean distance, as opposed to the average minimum distance (specified by Lafon) to induce maximum convergence of distance (see section on Comparison of transformed distances for an example). The vector, v is calculated from the square root of the product between matrix components from K₁ and vector components from e. The matrix P is then formed by the product of v and v_T. Then, the weighted graph Laplacian matrix, K is calculated by component division of the matrix K₁ elements

(K_1ij where i = 1, ..., q and j = 1, ..., n) by the matrix P elements

(P_ij where i = 1, ..., q and j = 1, ..., n). The calculations are given as the following:

v=K1ijei MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWG2bGDcqGH9aqpdaGcaaqaaiabdUealnaaBaaaleaacqaIXaqmcqWGPbqAcqWGQbGAaeqaaOGaemyzau2aaSbaaSqaaiabdMgaPbqabaaabeaaaaa@370E@ where i = 1, ..., q; j = 1, ..., n; and e = (1, 1, ..., 1)^T, then

P = vv^T and

K = K_1ij/P_ij where i = 1, ..., q and j = 1, ..., n.

The K matrix is decomposed by singular value decomposition (svd)

svd(K) = XHV^T (Note that for this symmetric positive semidefinite matrix K, the svd is the spectral decomposition, however, to remain consistent with the nomenclature specified by Lafon, the calculations with svd are used.)

and the n columns of the X matrix which define the left singular vectors of K are scaled by the first column of X, given by

Φj=X[,j]X[,1] MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqqHMoGrdaWgaaWcbaGaemOAaOgabeaakiabg2da9maalaaabaGaemiwaGLaei4waSLaeiilaWIaemOAaOMaeiyxa0fabaGaemiwaGLaei4waSLaeiilaWIaeGymaeJaeiyxa0faaaaa@3C4E@ where j = 1, ..., n.

This provides n - 1 characteristic roots of the matrix K given by Φ_j. It should be noted that the first column of X is scaled by itself, creating a vector of values equal to 1. By convention, this vector is designated by Φ₀. As such, the second column in the matrix X will be considered the primary Φ vector, and designated by Φ₁. Utilizing these primary Φ vectors, the data can be embedded as points in ℝ^Φ.

The spectral implementation of the weighted graph Laplacian from Ng et al. is similar up to the calculation of the K₁ matrix, however, the kernel is defined by a denominator term of 2ε² as opposed to simply ε in Lafon's method. In addition, the epsilon smoothing term (ε) (as will be demonstrated in the Results section) is not optimized at ε = min _i≠j ||x_i - x_j||² > 0, as it is in our modification to Lafon's method, and thus requires fitting for each example analyzed and classifier utilized. Following the calculation of the K₁ matrix, the matrix D is a diagonal matrix calculated from the row sums of K₁,

Dii=∑j=1nK1ij MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGebardaWgaaWcbaGaemyAaKMaemyAaKgabeaakiabg2da9maaqahabaGaem4saS0aaSbaaSqaaiabigdaXiabdMgaPjabdQgaQbqabaaabaGaemOAaOMaeyypa0JaeGymaedabaGaemOBa4ganiabggHiLdaaaa@3D91@ where i = 1, ..., q. Then the normalized Laplacian matrix, L is calculated as

L = D^-1/2K₁D^-1/2

The Laplacian matrix, L is decomposed by svd, and the n columns of the X matrix which define the left singular vectors of L have rows scaled to unit length into the matrix Y

svd(L) = XHV^T (Note the point specified in the decomposition of the weighted graph Laplacian as provided by Lafon.)

Yij=Xij(∑j=1nXij2)1/2 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacH8akY=wiFfYdH8Gipec8Eeeu0xXdbba9frFj0=OqFfea0dXdd9vqai=hGuQ8kuc9pgc9s8qqaq=dirpe0xb9q8qiLsFr0=vr0=vr0dc8meaabaqaciaacaGaaeqabaqabeGadaaakeaacqWGzbqwdaWgaaWcbaGaemyAaKMaemOAaOgabeaakiabg2da9maalaaabaGaemiwaG1aaSbaaSqaaiabdMgaPjabdQgaQbqabaaakeaacqGGOaakdaaeWbqaaiabdIfaynaaDaaaleaacqWGPbqAcqWGQbGAaeaacqaIYaGmaaaabaGaemOAaOMaeyypa0JaeGymaedabaGaemOBa4ganiabggHiLdGccqGGPaqkdaahaaWcbeqaaiabigdaXiabc+caViabikdaYaaaaaaaaa@46C2@ where i = 1, ..., q and j =1, ..., n.

Utilizing the primary Φ columns of this Y matrix, the data can be embedded as points in ℝ^Φ.

Two implementations of PCA were calculated for comparison: correlation and covariance. These calculations were conducted with the prcomp function in the 'stats' package of R 29. Since classical MDS is synonymous to PCA calculated on a Euclidean distance matrix, nonmetric MDS was performed instead (based on ranking of dissimilarities), to avoid redundant information and for an additional comparison with another nonlinear approach, using the isoMDS 31 function in the 'MASS' package of R 29. Kernel PCA was calculated with a Gaussian radial basis function kernel using the k.pca function in the 'kmethods' package of R 29. This kernel function was chosen to maintain consistency with the kernel used in both spectral methods. The weighted graphs Laplacian for the two spectral methods are anchored on some form of a Gaussian kernel (see Spectral methods section for the difference between the kernels).

In order to better demonstrate how each of the three DR methods with a kernel function transforms Euclidean distances to its respective weighting values (e.g. Gaussian radial basis function, weighted graph Laplacian L matrix-Ng et al., weighted graph Laplacian K matrix-Lafon), an example was generated with simulated data. A data matrix of five observations was generated (each composed of 10 variables) and the three methods with a kernel function were calculated on the data matrix to compare between the final transformed weight matrix that is decomposed (i.e. svd) and the standard Euclidean distance matrix (Table 3). This evaluation illustrates how distances within each method are transformed in the final step prior to decomposition, and shows where on the distribution local and remote distances converge in this transformation. For example, in kernel PCA, a Gaussian radial basis function kernel is computed from the distance matrix and these entries are plotted against their respective Euclidean distances to represent the transformed space that the eigenfunctions are calculated on, in order to provide a low dimensional embedding. For both weighted graphs Laplacian from Ng et al. and Lafon, the weight values in the L and K matrix, respectively, are compared against the Euclidean distance matrix.

The Euclidean matrix in Table 3 contains 10 unique distance values. Two values are less than the 25% of the distribution, three values are greater than the 75% of the distribution, and the remaining values are within the interquartile range. This range of distances is utilized to convey both the subtle and apparent differences within the transformed space between methods.

The results are shown in Figure 6. For each plot, the points represent the relationship between the transformed space (y-axis) and Euclidean distance (x-axis). The trend lines in each plot are determined by ordering both the x and y axis vectors, so an ordered transformation from Euclidean distance values to the weighted values would be depicted with a line that passes through each point. Those plots that do not exhibit this line pattern do not maintain the identical ordering from distances to weights. The latter statement does not imply that there is either a disadvantage or an advantage to the method. It simply provides a means of comparison for those methods that distort the ordering of some distances when transformed into a weight value. The black line in each plot is calculated based on assigning the epsilon smoothing term to the minimum non-zero squared Euclidean distance (ε = min_i≠j ||x_i - x_j||² > 0). Each subsequent line that is shaded by a portion of the rainbow spectrum (ROYGB) is then calculated by increasing this epsilon term by 1% (determined from the distribution of the squared Euclidean distances) increments up to the 50% of the distribution of squared Euclidean distances. Each line color is assigned to bins of 10% incremental values from 1% to 50% (e.g. weight value lines calculated with epsilon = 1%–10% values from the distribution of squared Euclidean distances are shaded red, epsilon = 11%–20% values from the distribution of squared Euclidean distances are shaded orange, etc.).

From the kernel PCA plot in Figure 6a, a Gaussian radial basis function demonstrates convergence of local Euclidean distances in the transformation to weight values, while remote distances exhibit more linearity with these weight values. In addition, as the epsilon smoothing term is increased at very small values, local distances converge to similar weight values much faster than remote distances. This example illustrates how local distance structure is better preserved at levels specific to the epsilon smoothing term utilized, than remote distances. Small distances are collapsed to a similar weighted value, while large distances maintain relatively stronger linearity with Euclidean distances, though this is reduced as the epsilon smoothing term is increased to values greater than the 10% of the distribution of squared Euclidean distances (denoted by the change in line shading from red to orange). The two smallest Euclidean distances converge to similar values rather quickly as the epsilon smoothing term is increased, while the points in the distance distribution converge more slowly. At epsilon values greater than the 20% (line shading of yellow, green, and blue) of the distribution of squared Euclidean distances, the weight values all converge to the same value of 1.

The Laplacian matrix, L from Ng et al., in Figure 6b shows that at minimal values of the epsilon smoothing term, there is near linearity between Euclidean distances and values in the matrix L. However, as the smoothing epsilon term is increased, the convergence occurs with both local and remote distances, while the points in the middle of this distribution maintain similar values (represented by the approximate slope of 0 from the values within the interquartile range of the plot). Both tails of the trend line quickly approach a weight value of ~0.25 with epsilon values at less than 20% of the distribution of squared Euclidean distances (denoted by line shading of only red and orange). This example illustrates that values within the middle region of the distribution (as compared the right and left tails of the distribution) are transformed to similar weight values in matrix L, with rather small values of the smoothing epsilon term, and then as this epsilon value is increased to only slightly larger values, all points in the function converge to a similar weight value. In addition, the Laplacian matrix, L does not preserve the same ordering of Euclidean distances.

The Laplacian matrix, K from Lafon (Figure 6c), in contrast to the matrix L, demonstrates convergence of weight values on both extremes of the Euclidean distance vector at very small values of the epsilon smoothing term, as illustrated by the small variance in red lines (i.e. epsilon vales at less than the 10% of the distribution of squared Euclidean distance). Additionally, opposite to the matrix L, these weight values become more linear (less converged) on both right and left tails of the distribution as the epsilon smoothing term is increased to values greater than the 20% of the distribution of squared Euclidean distances (line shading corresponding to colors of yellow, green, and blue). It is interesting to note that at maximal epsilon values (indicated by yellow, green, and blue lines in Figure 6c), the function generated by the weight values in the K matrix resemble the function generated by the weight values in the L matrix at the minimum epsilon value (indicated by the black line in Figure 6b). However, similar to the L matrix values, the ordering of Euclidean distances is not preserved. This example illustrates how the Lafon method differs from the other two, in that maximal convergence of both local and remote distances is optimized at minimal epsilon values (i.e. values at less than 10% of the distribution of squared Euclidean distance for this example). Based on the partitioned structure in the data, this reduces the dynamic range of Euclidean distances, particularly along the middle of the distribution, and transforms the distance structure to weight values (in matrix K) at both extremes of the distribution. Such a transformation, that acts to create a large gap between local and remote distances (at a partitioning threshold driven by the data) is shown to provide an optimal distance transformation for subsequent spectral decomposition for purposes of elucidating meaningful structure in image seriation and microarray species classification examples presented. Since the convergence of local and remote distances in the K matrix is highly dependent upon the primary partitioning point (for this example, between the Euclidean distances of 12.2 and 26.8), it is assumed that the most meaningful structures are defined by the difference between the smallest two Euclidean distances and the remaining eight Euclidean distances.

The k-nearest neighbor (KNN) algorithm (for k = 2 and 3) was calculated for 1,000 independent trials for each method (to address the problem of random assignment when ties occur in nearest neighbor voting) and the mean error was computed using the knn function in the 'class' package of R 29. The discriminant analysis classifiers both for separate variances (QDA) and pooled variances (LDA) were computed using the qda and lda functions, respectively, in the 'MASS' package of R 29. All classification models were trained and tested with leave-one-out cross-validation (LOO-CV).