A comparison of multivariate analysis techniques and variable selection strategies in a laser-induced breakdown spectroscopy bacterial classification

TL;DR: The partial least squares discriminant analysis was more effective at distinguishing between highly similar spectra from closely related bacterial genera and may be the preferred multivariate technique in future species-level or strain-level classifications.

About: This article is published in Spectrochimica Acta Part B: Atomic Spectroscopy.The article was published on 2013-09-01 and is currently open access. It has received 38 citations till now. The article focuses on the topics: Discriminant function analysis & Partial least squares regression.

Summary

1. Introduction

• Since the initial demonstrations of bacterial identification with laser-induced breakdown spectroscopy (LIBS) in 2003, significant progress has been made in the use of multivariate chemometric analyses to classify unknown bacterial LIBS spectra.[1-4].
• Over the last five years the authors and others have demonstrated a sensitive and specific identification of live bacterial biospecimens utilizing a discriminant function analysis (DFA) to classify LIBS spectra.[5-8].
• The intensities of strong specific elemental atomic emission lines normalized by the total observed spectral power have been utilized as independent variables in this multivariate analysis. [9].
• And this is an ongoing area of investigation.
• Model performance was quantified by calculating truth tables (and the resulting sensitivity and specificity) from the external validation tests.

2.1. Experimental Setup

• The LIBS apparatus used to obtain the bacterial spectra, as well as their bacterial sample preparation and mounting protocols, have been described at length elsewhere.
• Five spectra were acquired at each sampling location, thus twenty-five laser pulses were used to obtain this spectrum.
• The bacteria were chosen to represent a fairly wide taxonomic range.
• The 32 distinct experiments that were performed yielded the 32 data sets shown in column three of Table 1.
• No data “outliers” were omitted from their data sets and efforts were made to maximize the number of spectra from any one bacterial deposition rather than to standardize the number of spectra taken.

2.2 Models for Chemometric Analysis (Lines, RM1, and RM2)

• The three independent variable models that were tested are referred to here as the “lines” model, ratio model one (RM1), and ratio model two (RM2).
• The lines model was the simplest of the three, having been used in all their previous work.
• This approach has been utilized with success by Gottfried et al. to discriminate LIBS spectra obtained from explosives residues.
• The first thirteen variables were merely the intensities of the thirteen strong emission lines used in the lines model (indicated by an asterisk).
• It was decided that when the dimensionality of the original data was not reduced significantly then the benefits of performing a down-selection were reduced and the more appropriate model would be to use the entire spectrum.

2.3 Chemometric Analysis Techniques

• Two multivariate chemometric analysis techniques were compared for discrimination between different bacterial genera based on the LIBS emission spectra.
• This is known as external validation, because each spectrum was tested against a library where no other spectra acquired at the same time or under the same conditions were present.
• PLS-DA takes a set of independent variables as determined by their models and constructs latent variables to maximize the variance between the two groups.
• The identity of unknown spectra was then predicted based on this discrimination line in the pre-compiled library.
• All unknown samples were classified in a PLS-DA test specific for each genus, and if the test group was classified as belonging to the “no group” for each model, it remained unknown and was not classified as belonging to any genus.

3. Results and Discussion

• In each of the DFA results, four discriminant functions (DF1 through DF4) were constructed to determine the classification of each spectrum.
• The “unknown” bacterial spectra are represented by the “x” symbols and 34 of 34 unknown spectra were correctly classified as Mycobacterium, even though the model contained no other spectra from strain TA.
• An investigation of the PLS-DA was conducted to compare the number of LV’s and the corresponding rates of true positives and true negatives.

4. Discussion

• A comparison of the DFA performed with the three different models consisting of lines, RM1, and RM2 showed that RM2 yielded the overall highest true positive and true negative rates with true positive rates of 95%, 54%, 95%, and 88% for the four genera and true negative rates of 91%, 99%, 99%, and 99%.
• The sensitivity and specificity were obtained by averaging the results from the 31 tests and the standard deviation is reported as the uncertainty.
• This was merely a result of there being only two representative Staphylococci data sets to include in the analysis, as can be seen in Table 1, with one of these data sets being among the earliest experiments performed in the construction of the spectral library.
• Therefore both analyses can perform both functions, if necessary.
• It may therefore be true that a DFA is more effective in genus-level discrimination on bacterial specimens with a wide range of potential identities, but discrimination at the species- or strain-level once the genus is accurately identified may require the use of PLS-DA.

5. Conclusion

• The authors have shown that a sensitive and specific genus level classification of LIBS spectra from live bacterial specimens can be performed with a DFA or a PLS-DA using several different independent variable models.
• All results were obtained using external-validation tests.
• The number of latent variables required for efficient classification using this model was investigated, and chosen to be 20 in all subsequent tests.
• More precise identification at the species-level or strain-level may be subsequently performed with a PLS-DA, which demonstrated improved performance at discriminating highly similar spectra.
• It is likely 18 that computational processing power would easily allow such a verification, as the classification of one unknown spectrum against a pre-compiled library model is performed rapidly by both techniques.

Figures

Figure 3

Table 1 Identities of the 32 data sets used to construct a spectral library composed of 669 bacterial LIBS spectra.

Table 2 The twenty-four independent variables used in ratio model one (RM1).

Figure 1

Table 5 Truth table results for two multivariate techniques (DFA and PLS-DA) utilized in a genus-level classification of bacterial LIBS spectra.

Figure 4

Table 4 Truth table results for three independent variable models utilized in a genus-level discriminant function analysis of bacterial LIBS spectra.

Table 3 The 80 independent variables used in ratio model two (RM2).

Figure 2

Frequently asked questions

Q1. What are the contributions in "A comparison of multivariate analysis techniques and variable selection strategies in a laser-induced breakdown spectroscopy bacterial classification" ?

In this paper, the authors compared the use of three different down-selected variable models consisting of emission intensities, the sum of observed ∼4 intensities from the elements P, Ca, Mg, Na, C, and complex ratios of those intensities. 

Q2. What have the authors stated for future works in "A comparison of multivariate analysis techniques and variable selection strategies in a laser-induced breakdown spectroscopy bacterial classification" ?

Such a confirmation will need to be investigated in future work.