Open Access
Add to BibSonomy
Abstract
Pathogenic microbes contribute to several major global diseases that kill millions of people every year. Bloodstream infections caused by these microbes are associated with high morbidity and mortality rates, which are among the most common causes of hospitalizations. The search for the “Holy Grail” in clinical diagnostic microbiology, a reliable, accurate, low cost, real-time, and easy-to-use diagnostic method, is one of the essential issues in clinical practice. These very critical conditions can be met by Raman tweezers in combination with advanced analysis methods. Here, we present a proof-of-concept study based on Raman tweezers combined with spectral mixture analysis that allows for the identification of microbial strains directly from human blood serum without user intervention, thus eliminating the influence of a data analyst.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Bloodstream infections and sepsis are among the most common and often life-threatening causes of hospitalization, but can be successfully treated if both causative pathogens and their antimicrobial susceptibility and/or resistance are characterized as soon as possible [1]. Staphylococcus species, Escherichia coli, and Candida albicans belong to the most common microbes that cause bloodstream infections worldwide [2,3]. Furthermore, 20% of all C. albicans bloodstream infections are polymicrobial in nature, with Staphylococcus epidermidis and Staphylococcus aureus being one of the most common co-isolated organisms [2]. These species are part of the commensal human microflora. However, they can cause hospital-acquired infections with an extreme ability to inhabit diverse host niches, especially in immunocompromised patients and patients with co-morbidities [2]. Some of the microorganisms mentioned above occur in antibiotic-resistant forms. For example, methicillin-resistant S. aureus is the leading cause of hospital-acquired infections [4]. In addition, these microbes generally have a strong ability to form biofilms on both native and artificial substrates in the human body, providing them additional protection against antimicrobial agents and the host immune system. This leads to increased mortality, prolonged hospital stays, and higher costs of medical care. Therefore, a rapid and sensitive identification method and efficient elimination of pathogens are fundamental research goals in modern clinical microbiology and infection control.
Standard methods of microbial identification rely on time-demanding in-vitro cultivation followed by automated systems with microbe detection thresholds of approximately 10$^{7}$ to 10$^{9}$ CFU/mL (colony-forming units) [5]. The threshold varies depending on the microbial load (from several hours to days) [6–8]. Blood cultivation is commonly followed by Gram staining and biochemical testing or matrix-assisted laser desorption/ionization-time of flight (MALDI-TOF) mass spectroscopy [9]. Several protocols and procedures offer direct microbial identification from positive blood cultures, including antimicrobial resistance testing [10–12]. Other possibilities for identification include molecular methods, for example, polymerase chain reaction (PCR)-based methods [13–16] or identification based on antigen-antibody reaction [17]. However, all currently used methods have their limitations which mainly include a long time from blood draw to result, low specificity, costly consumables, personnel demand, etc.
An alternative method that can be used to identify microbes that cause bloodstream infections is the combination of Raman spectroscopy (as an analytical tool) and optical tweezers, known as Raman tweezers. They have already been proven to be reliable, rapid and cost-effective in the detection and identification of microbes that cause urinary tract infections, bloodstream infections, and infections in general [18–22].
The advantage of Raman tweezers is their ability to analyze a few living cells directly from human body fluids in a contactless way, without time-consuming in-vitro cultivation. For example, microbes spiked in human urine can be identified in less than 10 minutes at concentrations of approximately 105 CFU/mL [23]. Additionally, such fast analyses can be extended to characterize antimicrobial resistance and/or other microbial virulence factors, e.g. biofilm formation [21,24–26]. Like all methods, Raman tweezers have some limitations. One of the most important aspects is that the Raman signal is inherently weak (especially in the case of single-cell detection), which is often overcome by longer acquisition times and can lead to photodamage of the cells. This photodamage does not degrade the entire sample because Raman tweezers analysis typically uses tens to hundreds of cells, which is a negligible number in the case of clinical samples. Another major drawback of Raman tweezers analysis is the amount of time the operator has to spend searching for individual cells to analyze [27]. However, the implementation of the entire procedure in a microfluidic platform can contribute to a significant reduction in measurement times.
To our knowledge, only a few studies have focused on the identification of microbes directly in blood serum. Gu et al. developed a metagenomic next-generation sequencing (mNGS) test using cell-free DNA from body fluids to identify pathogens from clinical samples [28]. Depending on the pathogen, the sensitivity and specificity of the method ranged from 79 to 91% and 81 to 100%, respectively. As mentioned above, sequencing methods require preprocessing prior to analysis, require costly consumables, trained personnel, and time (the achievable time <6h). In addition, mNGS detects only the presence of nucleic acids (or their parts), which can be misleading. The great advantage of mNGS is its suitability for analyzing all kinds of body fluid. Therefore, genomic methods are slower but powerful and should follow up on the initial rapid screening if doubts occur.
Several studies explored methods of isolating bacteria from different types of matrices, such as blood, urine, ascitic fluid, etc. [29,30]. Pahlow et al. developed a Raman-compatible chip for isolating microorganisms from complex media [31]. The isolation of bacteria is achieved by using antibodies as capture molecules. The captured bacteria are analyzed by Raman spectroscopy and classified by chemometric methods. The authors achieved a sensitivity of 75 to 98% and a specificity of 90 to 100% for samples diluted in water. Another widely used separation method – dielectrophoresis – is often used to isolate individual pathogens for subsequent analysis [32,33]. The described isolation of bacterial cells from different matrices is a very promising tool that can be used in various body fluid media and different kinds of matrices (e.g., blood, soil, etc.). However, it represents an extra step that is not needed if Raman tweezers are used for analysis directly in blood serum.
In the scenario where Raman tweezers analysis is performed directly in body fluid (urine, blood serum, blood plasma, saliva, etc.), the identification can be complicated even though the cells cover most of the illuminated volume, because the influence of the background (response originating from the body fluid) cannot be ruled out. The background may vary significantly from one clinical sample to the other due to the patient’s condition. This makes the post-processing required for a successful microbial identification rather challenging.
This study presents an automatic Raman spectrum-based classification method. In particular, the amount of user input is minimized to reduce the possible systematic error introduced by the selection of unsuitable parameters e.g. selection and parameters of fluorescence background removal algorithm, signal smoothing strength. Therefore, only two of the often-used Raman spectrum preprocessing procedures (preparatory steps before applying an algorithm or analysis) are used – cosmic ray removal and normalization. Both can be performed automatically (without the influence of a data analyst). Building on optical methods in combination with spectral mixture analysis, we demonstrate a proof-of-concept identification of microorganisms directly in blood serum.
2. Materials and methods
2.1 Sample preparation
In the Raman tweezers experiment, we used pooled anonymized human serum samples from the collection of the Department of Microbiology, St. Anne’s University Hospital, Brno, Czech Republic. Five different blood serum samples were randomly selected and the whole experimental procedure was repeated on each of them independently. Serum was used exclusively as a matrix. Each blood serum sample was divided into five aliquots of 0.5 mL – one aliquot was without spiking (pure serum), and each of the four remaining aliquots was spiked with different microbial strains – Staphylococcus aureus CCM 3953, Staphylococcus epidermidis CCM 4418, Escherichia coli CCM 3954 and Candida albicans CCM 8261. All microbial strains were provided by the Czech Collection of Microorganisms (CCM), Brno, CZ, and stored at −80 °C (similar to the human sera). Before the experiment, all microbial strains were thawed and cultured for 24 hours at 37 °C on Mueller-Hinton agar plates. A single inoculation loop (1 µL) full of microbes was then spiked into the serum aliquot (the final concentration was approximately 10$^{7}$ to 10$^{9}$ CFU/mL depending on the size of the microbes).
2.2 Experimental setup
Experimental data were obtained using a custom-built compact system of Raman tweezers that is connected to a commercial spectrometer (Renishaw inVia Raman Spectrometer, Renishaw plc., Wotton under Edge, UK), see Fig. 1. The compact module of Raman tweezers contains optical elements that expand the laser beam to overfill a back-plane aperture of the microscope objective. The compact Raman tweezers module is mounted between a microscope objective turret and a microscope objective (see Fig. 1(c)). As a laser source, we used a 785 nm external cavity diode laser in Littrow configuration (Sacher TEC-510-0785-1000, maximum output power 800 mW, $\mathrm {M}^2$=1.7) that was connected directly to the compact Raman tweezers module via a single-mode optical fiber (optimized for 780-970 nm, cladding diameter 125 µm, core diameter 4.4 µm). Subsequently, the laser beam was focused inside the sample chamber using a high numerical aperture microscope objective (Olympus, UPLSAPO60XW, NA=1.2, WD=0.28), where the microorganism was optically trapped and analyzed. The same laser beam was used for optical trapping and Raman spectroscopy (single-beam Raman tweezers). Raman scattered light originating from trapped cells was collected by the same microscope objective and delivered to an optical path of a detection part of the commercial Raman spectrometer (Renishaw InVia spectrometer). From the parameters of the experimental apparatus, we determined that the diameter of the Airy disc is approximately 800 nm. For the microbes S. aureus, S. epidermidis, and E. coli, the size of the bacteria is comparable to the size of the focused laser beam spot. In the case of C. albicans (size 5 to 10µm), we assume a random orientation of the trapped cell.
Fig. 1. Compact Raman tweezers module installed on a Raman spectrometer, a) View of the whole setup, b) Detail of the Raman tweezers compact module mounted between the microscope objective turret and the microscope objective, c) Sketch of the Raman tweezers compact module in section (not to scale), where L1 is aspherical lens, M1 – long pass filter with edge at 785 nm (Chroma), M2 – displacement compensation window, ML – microscope lens.
2.3 Sample chamber
A sample chamber was prepared using two calcium fluoride (CaF$_{2}$) windows (Crystran, CaF$_{2}$ Raman grade polished disc, diameter 13 mm, thickness 0.2 mm) that were separated by a glass spacer (made from microscopy slide, soda lime glass, Marienfeld, Germany) with a thickness of approx. 1 mm. About 150 µL of blood sera with microbial cells were loaded into the sample chamber and closed from both sides by CaF$_{2}$ slides.
2.4 Acquisition of Raman spectra
The laser beam that passes through the compact optical tweezers module was focused by the microscope objective into the sample volume in the sample chamber. The microbes were trapped and analyzed with a laser power of 130 mW (power at the output of the fiber was approximately 170 mW) in the plane of the laser focus. The scattered light was collected for 50 s (acquisition time 1 s, 50 accumulations) per spectrum by the same laser (785 nm). Subsequently, the microscope stage was moved and another bacterium was trapped. These measurements resulted in the creation of five datasets containing five classes (four different microbes measured in serum and pure serum). The exact numbers of spectra in each dataset are described in Table 1.
Table 1. Number of measured Raman spectra in each data set. Each blood serum (A-E) was spiked with four microbes (C. albicans, E. coli, S. aureus, and S. epidermidis)
2.5 Data processing
The key steps of the Raman spectra processing and identification pipeline for the spectral mixture analysis and classification are presented in Fig. 2.
Fig. 2. Raman spectrum processing pipeline for a) identification using spectral mixture analysis and b) classification using K-nearest Neighbour algorithm. The pipeline is divided into two parts – preprocessing and identification.
2.5.1 Preprocessing for spectral mixture analysis
Cosmic-ray peak removal and normalization are necessary preparatory steps before any further analysis, which is why they are also included in our preprocessing; see Fig. 2(a). Cosmic-ray peaks were suppressed using convolution with discrete Laplace operator $D_x^2$ and corrected using linear interpolation as described in [34]. The discrete Laplace operator $D_x^2 = [\begin {smallmatrix} -1 & 2 & -1 \end {smallmatrix}]$ was convolved with maximum normalized spectra to identify areas of sharp increase or decrease in the signal intensity. The output of the convolution was then thresholded to locate the cosmic ray-peaks (threshold $t=0.1$). These peaks were removed and replaced with values obtained from linear interpolation of the edge values of the removed window with a size of 12 pixels (approximately 13 cm-1). This procedure was used repeatedly on one spectrum for 10 iterations to remove all cosmic-ray peaks in the data. The next and final step of preprocessing, data normalization, was performed using the L2 norm to scale individual samples to have a unit norm using the entire fingerprint region from 600 to 1700cm-1, without subtraction of the mean value [35].
2.5.2 Spectral mixture analysis
To identify the microbes, we assume that the measured spectrum is a linear mixture [36] of the Raman response originating from the trapped microbes as well as the surrounding medium, which is present in the Raman scattering region, as described in Eq. (1), where $\mathbf {y}$ is the measured Raman spectrum, $a_i$ is the fraction coefficient (abundance), $\mathbf {x_i}$ is the spectral component (endmember), $\mathbf {\varepsilon _i}$ is the residual between the measured and model spectrum, and $n$ is the number of spectral components (endmembers).
Without prior knowledge of the measured samples, the identification of individual spectral components can be performed using spectral unmixing algorithms, as explored in [37–39]. Because our dataset contains Raman spectra of the blood serum without any cells and individual microbes within the blood serum, an average value for each of these spectra in the training set (see below) is calculated. The averages are then used as the spectral components (endmembers) for the mixture analysis. These mean spectral responses are representative of each class with the following labels: serum without microbes (the response of the surrounding medium), S. aureus, S. epidermidis, E. coli, and C. albicans.
When the spectral components are identified, their fraction coefficients can be calculated by a least-squares optimization algorithm, which takes the mixture spectrum and spectral components (endmembers) as input. This process is known as a spectral mixture analysis or abundance mapping [38,40]. Under the condition that all coefficients sum to one and are non-negative, the coefficients denote the percentage of component abundances in the spectrum mixture. This is the basis for the Fully Constrained Least Squares (FCLS) linear spectral mixture analysis method [40]. We selected this set of constraints because the measured Raman spectrum is composed of multiple spectral responses (depending on the composition of the cell) which are overlapping. The constraints can also be modified, so that the components do not need to sum to one – Nonnegative Least Squares (NNLS), or impose no constraints to perform Unconstrained Least Squares (UCLS) spectral mixture analysis. After the fraction (abundance) coefficients are calculated, each spectrum is identified to be a member of a particular class based on its endmember with the highest coefficient value.
2.5.3 Preprocessing for classification
The commonly used preprocessing includes cosmic ray removal, smoothing, baseline correction, and normalization; see Fig. 2(b). In the spectroscopic community, Rolling Circle Filter (RCF) and Iterative Polynomial Fitting (IPF) are often used for the removal of fluorescence background [41,42]. These algorithms require a suitable selection of their respective parameters to work efficiently [43]. A data analyst should understand the data preparation process for analysis, its parameters and limitations of the methods [44].
The preprocessing used for classification consisted of cosmic ray removal and normalization with the same parameters as described above followed by smoothing using the Savitzky-Golay filter (second order, window size of ${7}{\textrm{points}} \approx {7} \textrm{cm}^-1$). The baseline correction was then performed using IPF (12th order polynomial, 10 iterations) or RCF (circle diameter of ${600}\textrm{points} \approx {600}\textrm{cm}^-1$, 10 iterations).
2.5.4 Classification
We compared the results obtained using the spectral mixture analysis with the widely used identification approach that involves Principal Component Analysis (PCA) and the K-Nearest Neighbors (KNN) algorithm for classification [45]. The classification parameters were the following – dimensionality of the data using PCA (2 components) and classification using a one-nearest neighbor KNN with the L2 metric.
2.5.5 Algorithm performance evaluation
The measured data are randomly split into training and testing sets for algorithm validation in an 80:20 ratio [46] with equal representation of each class. The spectral components were calculated as a per-class mean from the training set and evaluated on the test data. The accuracy of the identification is defined as a ratio of correctly identified spectra to the total number of classified spectra. Sensitivity and specificity are another metrics used to evaluate the proposed algorithm [47]. Sensitivity measures how well the algorithm identifies the presence of bacteria in a spiked serum Raman spectrum, and specificity describes the ability to correctly identify spectrum of a serum with no bacteria as pure serum. Because the last two metrics apply only to binary cases, we summarized the results to two cases, denoting the presence and absence of microbes in the measured spectra. The equations used for the calculation of the sensitivity and specificity can be found in Eq. S1 and Eq. S2 in Supplement 1.
3. Results and discussion
The core principle of our Raman tweezers experiment combined with advanced data analysis methods is shown in Fig. 3. An example image of optically trapped E. coli bacterium obtained using this setup can be found in Fig. S1 in Supplement 1. Raman tweezers trap microbial objects in the blood serum, and the Raman fingerprint response of the trapping volume (containing both the microbe and the blood serum) is collected.
Fig. 3. Schematic diagram of the Raman tweezers experiment with advanced data analysis. The microbes are trapped and analyzed directly in the blood serum so that the measured Raman response contains both a portion of the Raman response of the trapped object and a partial response of the surrounding medium. The spectral components of the trapped objects are then analyzed to identify individual microbial representatives.
An example of the obtained data is presented in Fig. S2 in Supplement 1 along with a plot of its PCA loadings (Fig. S3 in Supplement 1) and a list of Raman modes (Table S1 in Supplement 1). The collected response is a composition of the cell response and the surrounding medium. The distinction between the individual components of the measured Raman spectra is not trivial for two main reasons: 1) the blood serum contains components that are also present in the cell itself (mainly proteins) and 2) the intensity of the background response (serum) is a function of the size of the trapped particle(s) (covering the trapping focus). For the reasons described above, it is necessary to use advanced methods to resolve the individual spectral components.
The algorithms are evaluated by the accuracy of the identification performed on the test set; see Table 2. Under our experimental conditions, we obtained the best identification accuracy, almost 90 %, using a spectral mixture analysis with the FCLS algorithm. The FCLS algorithm achieves very good results without analyst intervention, and since it has the best results among the spectral unmixing algorithms, we will discuss the results of this method in detail. An illustrative way to present the identification results is through the confusion matrix [48], also known as the error matrix, which can describe the performance of the model in more detail than just the overall accuracy of the identification. The relative number of correct and incorrect identifications is summarized and broken down by class. Confusion matrix of the FCLS identification was calculated for each serum individually using the test set data (see Fig. S4 in the Supplement 1). The confusion matrix shown in Fig. 4 was created by averaging the individual confusion matrices. The averaged confusion matrices for NNLS, UCLS, and KNN with IPF are shown in Fig. S5 in Supplement 1.
Fig. 4. Averaged confusion matrix for identification using data from test sets classified using the Fully Constrained Least Squares (FCLS) spectral unmixing algorithm (left part) and K-Nearest Neighbors with Rolling Circle Filtering used for background correction (KNN+RCF). Data are normalized by rows (true labels).
Table 2. Accuracy of microbe identification in blood serum (from the test set), including uncertainties calculated by the Root Mean Square Error (RMSE). Serum A-E labels serum spiked with four microbes (C. albicans, E. coli, S. aureus, and S. epidermidis). FCLS labels Fully Constrained Least Squares, NNLS labels Non-Negative Least Squares, and UCLS labels Unconstrained Least Squares, KNN+IPF labels K-Nearest Neighbors Algorithm applied to spectra with background removed by Iterative Polynomial Fitting, KNN+RCF labels K-Nearest Neighbors Algorithm applied to spectra with background removed by Roling Circle Filter
The diagonal elements of the confusion matrix show the accuracy for each class (five classes – four for microbes in the serum and one for pure serum). Off-diagonal values show how the model failed in classification. The first row of the left part of Fig. 4 indicates that the FCLS algorithm correctly distinguishes the spectra of blood serum (with no microbes) from the spectra of trapped cells in the blood serum. The results of the FCLS analysis indicate that the algorithm is perfectly capable of distinguishing pure serum (having no false positives), which is not the case when using the KNN+RCF algorithm (right part of Fig. 4).
In addition, both algorithms identified some of the C. albicans and E. coli spectra as background responses (false negative). The classification of serum spiked by pathogens as pure serum can be caused by fluctuation of trapped cells or loss of trapped cells from optical tweezers during acquisition, resulting in a significantly lower Raman response of the microbial cell compared to the response of surrounding blood serum. The misclassification of S. epidermidis and S. aureus (classification of S. aureus as S. epidermidis and vice versa) can be explained by the fact that these two microbes belong to the same gram-positive bacterial family, especially in our case when the S. aureus is a non-pigmented strain. The details about the used strain of S. aureus can be found in [49]. The incorrectly identified Raman spectra using the FCLS algorithm for each serum are shown in Fig. S6 in Supplement 1.
The total performance of the system can also be evaluated based on the sensitivity and specificity; see Table 3. When comparing the sensitivity and specificity, the FCLS spectral mixture algorithm achieves results comparable to (or better than) methods that depend on preprocessing of the analyzed data (KNN in combination with IPF and RCF).
Table 3. Specificity and sensitivity of the microbe detection in each serum using the Fully Constrained Leas Squares (FCLS) unmixing algorithm, K-Nearest Neighbors Algorithm applied to spectra with background removed by Iterative Polynomial Fitting (KNN+IPF) and K-Nearest Neighbors Algorithm applied to spectra with background removed by Roling Circle Filter (KNN+RCF). RMSE is Root Mean Squared Error of the average accuracy
The cutting-edge processing and detection method described here achieves accuracy comparable with the most common procedures in the Raman spectroscopy community for samples in water, evaporated water samples, and measured directly from the agar plate [22,23,50–54].
4. Conclusion
This paper presents a proof-of-concept study demonstrating the ability to identify microbes directly in blood serum using Raman spectroscopy. In this work, we introduced a rapid and reliable method of microbial identification in blood serum using Raman tweezers together with a processing pipeline that does not require user input, reduces possible user influence, eliminates the need for designing complicated analysis, and achieves comparabe (or higher) accuracy than previously used approaches. Blood serum was spiked with microbes C. albicans, E. coli, S. aureus, and S. epidermidis. Our approach relies on the assumption that the Raman spectrum collected via Raman tweezers contains a linear mixture of Raman responses originating from trapped cells and the surrounding blood serum. We used the linear spectral mixture analysis method, together with Fully Constrained Least Squares, Non-Negative Least Squares, and Unconstrained Least Squares optimization with the accuracy of the identification (89.7 ± 0.8)%, (84.3 ± 1.6)% and (87.1 ± 1.3)%, respectively. The most significant advantage of these algorithms is that they require a relatively small amount of training data compared to the requirements of neural network-based models (1317 spectra for training; 265 spectra for testing). We compared these algorithms with a commonly used method, the K-Nearest Neighbors algorithm, in combination with Principal Component Analysis and Iterative Polynomial Fitting or the Rolling Circle Filter as background removal methods, which achieved significantly worse identification accuracy of (85.1 ± 1.4)% and (75.2 ± 1.6)%, respectively. In this experiment, we analyzed blood sera spiked with one species of microorganism at a time, but we believe that the proposed spectral mixture analysis could also allow the analysis of mixed samples containing different microorganisms. However, it would be necessary to modify the evaluation strategy in which, instead of one dominant spectrum, we determine the presence of multiple microorganisms based on several spectral components with similarly significant representations in the measured spectrum.
The use of microbe identification directly in blood serum opens the way to a new concept of biosensors that operate directly in body fluids without the analyst’s intervention during processing. Raman tweezers combined with spectral mixture analysis can rapidly identify microbes (in the range of minutes), which can be used as a preliminary screening followed by more time-consuming (and accurate) microbiological analysis. In our experimental case, the concentration of cells in blood serum is significantly higher than in clinical samples (<100 CFU/mL for sepsis, $\sim$103 CFU/mL for bloodstream infection). It is important to note that the analysis and subsequent processing described were carried out on approximately $10^2$ cells, allowing Raman tweezers to operate at concentrations relevant to clinical samples.
Funding
Ministerstvo Zdravotnictví České Republiky (NU21-05-00341); Technology Agency of the Czech Republic (FW06010453); Ministerstvo Průmyslu a Obchodu (FV40455); Akademie Věd České Republiky (RVO:68081731).
Acknowledgments
The authors acknowledge the support of the Czech Academy of Sciences with institutional support RVO:68081731 and the Ministry of Industry and Trade FV40455, Technology Agency of the Czech Republic FW06010453, Ministry of Health of the Czech Republic project NU21-05-00341.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper can be obtained from the authors upon reasonable request.
Supplemental document
See Supplement 1 for supporting content.
References
1. A. Tabah, J. Lipman, F. Barbier, N. Buetti, and J.-F. Timsit, “Use of Antimicrobials for Bloodstream Infections in the Intensive Care Unit, a Clinically Oriented Review,” Antibiotics (Basel, Switz.) 11(3), 362 (2022). [CrossRef]
2. H. Carolus, K. Van Dyck, and P. Van Dijck, “Candida albicans and Staphylococcus Species: A Threatening Twosome,” Front. Microbiol. 10, 2162 (2019). [CrossRef]
3. M. Bonten, J. R. Johnson, A. H. J. Van Den Biggelaar, L. Georgalis, J. Geurtsen, P. I. De Palacios, S. Gravenstein, T. Verstraeten, P. Hermans, and J. T. Poolman, “Epidemiology of Escherichia coli Bacteremia: A Systematic Literature Review,” Clin. Infect. Dis. 72(7), 1211–1219 (2021). [CrossRef]
4. M. Suzuki, M. Miyaki, K. Sekine, T. Kurihara, S. Abe, N. Aikawa, and N. Shinagawa, “Antimicrobial-susceptible patterns of Staphylococcus aureus isolated from surgical infections: a new approach,” J. Infect. Chemother. 17(1), 34–39 (2011). [CrossRef]
5. B. W. Buchan, “Commentary: Can Automated Blood Culture Systems Be Both New and Improved?” J. Clin. Microbiol. 60(4), e00192 (2022). [CrossRef]
6. G. Martín-Gutiérrez, C. Martín-Pérez, A. Gutiérrez-Pizarraya, J. A. Lepe, J. M. Cisneros, and J. Aznar, “Time to positivity of blood cultures in patients with bloodstream infections: A useful prognostic tool,” Enfermedades Infecciosas y Microbiología Clínica 35(10), 638–644 (2017). [CrossRef]
7. B. Lamy, S. Dargère, M. C. Arendrup, J.-J. Parienti, and P. Tattevin, “How to Optimize the Use of Blood Cultures for the Diagnosis of Bloodstream Infections? A State-of-the Art,” Front. Microbiol. 7, 125 (2016). [CrossRef]
8. B. Lamy, “Blood culture time-to-positivity: making use of the hidden information,” Clin. Microbiol. Infect. 25(3), 268–271 (2019). Publisher: Elsevier. [CrossRef]
9. J. O. Lay, “MALDI-TOF mass spectrometry of bacteria,” Mass Spectrom. Rev. 20(4), 172–194 (2001). [CrossRef]
10. M. Sauget, X. Bertrand, and D. Hocquet, “Rapid antibiotic susceptibility testing on blood cultures using MALDI-TOF MS,” PLoS One 13(10), e0205603 (2018). [CrossRef]
11. K. Ugaban, P. Pak, and R. C. She, “Direct MALDI-TOF MS and Antimicrobial Susceptibility Testing of Positive Blood Cultures Using the FASTTM System and FAST-PBC Prep Cartridges-Performance Evaluation in a Clinical Microbiology Laboratory Serving High-Risk Patients,” Microorganisms 10(10), 2076 (2022). [CrossRef]
12. P. O. Barth, E. W. Roesch, L. Lutz, A. C. de Souza, L. Z. Goldani, and D. C. Pereira, “Rapid bacterial identification by MALDI-TOF MS directly from blood cultures and rapid susceptibility testing: A simple approach to reduce the turnaround time of blood cultures,” Braz. J. Infect. Dis. 27(1), 102721 (2023). [CrossRef]
13. S. Gola, R. Martin, A. Walther, A. Dünkler, and J. Wendland, “New modules for PCR-based gene targeting in Candida albicans : rapid and efficient gene targeting using 100 bp of flanking homology region: Efficient PCR-based gene targeting in Candida albicans,” Yeast 20(16), 1339–1347 (2003). [CrossRef]
14. K. Zhang, J.-A. McClure, S. Elsayed, T. Louie, and J. M. Conly, “Novel Multiplex PCR Assay for Characterization and Concomitant Subtyping of Staphylococcal Cassette Chromosome mec Types I to V in Methicillin-Resistant Staphylococcus aureus,” J. Clin. Microbiol. 43(10), 5026–5033 (2005). [CrossRef]
15. K. A. Datsenko and B. L. Wanner, “One-step inactivation of chromosomal genes in Escherichia coli K-12 using PCR products,” Proc. Natl. Acad. Sci. U.S.A. 97(12), 6640–6645 (2000). [CrossRef]
16. S. A. Dunbar, C. Gardner, and S. Das, “Diagnosis and Management of Bloodstream Infections With Rapid, Multiplexed Molecular Assays,” Front. Cell. Infect. Microbiol. 12, 859935 (2022). [CrossRef]
17. S. Van Den Berg, H. P. Bonarius, K. P. Van Kessel, G. S. Elsinga, N. Kooi, H. Westra, T. Bosma, M. M. Van Der Kooi-Pol, D. G. Koedijk, H. Groen, J. M. Van Dijl, G. Buist, and I. A. Bakker-Woudenberg, “A human monoclonal antibody targeting the conserved staphylococcal antigen IsaA protects mice against Staphylococcus aureus bacteremia,” Int. J. Med. Microbiol. 305(1), 55–64 (2015). [CrossRef]
18. C. Xie, J. Mace, M. A. Dinno, Y. Q. Li, W. Tang, R. J. Newton, and P. J. Gemperline, “Identification of Single Bacterial Cells in Aqueous Solution Using Confocal Laser Tweezers Raman Spectroscopy,” Anal. Chem. 77(14), 4390–4397 (2005). [CrossRef]
19. W. E. Huang, A. D. Ward, and A. S. Whiteley, “Raman tweezers sorting of single microbial cells,” Environ. Microbiol. Rep. 1(1), 44–49 (2009). [CrossRef]
20. S. Dochow, C. Krafft, U. Neugebauer, T. Bocklitz, T. Henkel, G. Mayer, J. Albert, and J. Popp, “Tumour cell identification by means of Raman spectroscopy in combination with optical traps and microfluidic environments,” Lab Chip 11(8), 1484 (2011). [CrossRef]
21. S. Bernatová, K. Rebrošová, Z. Pilát, M. Šerý, A. Gjevik, O. Samek, J. Ježek, M. Šiler, M. Kizovský, T. Klementová, V. Holá, F. Růžička, and P. Zemánek, “Rapid detection of antibiotic sensitivity of Staphylococcus aureus by Raman tweezers,” Eur. Phys. J. Plus 136(2), 233 (2021). [CrossRef]
22. K. Rebrošová, S. Bernatová, M. Šiler, J. Mašek, O. Samek, J. Ježek, M. Kizovsky, V. Holá, P. Zemánek, and F. Růžička, “Rapid Identification of Pathogens Causing Bloodstream Infections by Raman Spectroscopy and Raman Tweezers,” Microbiol. Spectrum 11(3), e00028 (2023). [CrossRef]
23. K. Rebrošová, S. Bernatová, M. Šiler, M. Uhlirova, O. Samek, J. Ježek, V. Holá, F. Růžička, and P. Zemánek, “Raman spectroscopy-a tool for rapid differentiation among microbes causing urinary tract infections,” Anal. Chim. Acta 1191, 339292 (2022). [CrossRef]
24. A. R. Spencer, M. D. Klein, J. Y. Ang, T. A. Twomey, D. J. Sant, C. N. Chinkhota, R. E. Kast, and G. W. Auner, “Staphylococcus aureus identification and antibiotic resistance determination using raman spectroscopy,” Journal of the American College of Surgeons 213(3), S49 (2011). [CrossRef]
25. J. Kirchhoff, U. Glaser, J. A. Bohnert, M. W. Pletz, J. Popp, and U. Neugebauer, “Simple Ciprofloxacin Resistance Test and Determination of Minimal Inhibitory Concentration within 2 h Using Raman Spectroscopy,” Anal. Chem. 90(3), 1811–1818 (2018). [CrossRef]
26. W. Liu, J.-W. Tang, J.-W. Lyu, J.-J. Wang, Y.-C. Pan, X.-Y. Shi, Q.-H. Liu, X. Zhang, B. Gu, and L. Wang, “Discrimination between Carbapenem-Resistant and Carbapenem-Sensitive Klebsiella pneumoniae Strains through Computational Analysis of Surface-Enhanced Raman Spectra: a Pilot Study,” Microbiol. Spectrum 10(1), e0240921 (2022). [CrossRef]
27. G. Volpe, O. M. Maragó, H. Rubinsztein-Dunlop, et al., “Roadmap for optical tweezers,” J. Phys. Photonics 5(2), 022501 (2023). [CrossRef]
28. W. Gu, X. Deng, M. Lee, et al., “Rapid pathogen detection by metagenomic next-generation sequencing of infected body fluids,” Nat. Med. 27(1), 115–124 (2021). [CrossRef]
29. S. Pahlow, S. Meisel, D. Cialla-May, K. Weber, P. Rösch, and J. Popp, “Isolation and identification of bacteria by means of Raman spectroscopy,” Adv. Drug Delivery Rev. 89, 105–120 (2015). [CrossRef]
30. S. Kloß, P. Rösch, W. Pfister, M. Kiehntopf, and J. Popp, “Toward Culture-Free Raman Spectroscopic Identification of Pathogens in Ascitic Fluid,” Anal. Chem. 87(2), 937–943 (2015). [CrossRef]
31. S. Pahlow, S. Kloß, V. Blättel, K. Kirsch, U. Hübner, D. Cialla, P. Rösch, K. Weber, and J. Popp, “Isolation and Enrichment of Pathogens with a Surface-Modified Aluminium Chip for Raman Spectroscopic Applications,” ChemPhysChem 14(15), 3600–3605 (2013). [CrossRef]
32. R. E. Fernandez, A. Rohani, V. Farmehini, and N. S. Swami, “Review: Microbial analysis in dielectrophoretic microfluidic systems,” Anal. Chim. Acta 966, 11–33 (2017). [CrossRef]
33. U.-C. Schröder, A. Ramoji, U. Glaser, S. Sachse, C. Leiterer, A. Csaki, U. Hübner, W. Fritzsche, W. Pfister, M. Bauer, J. Popp, and U. Neugebauer, “Combined Dielectrophoresis-Raman Setup for the Classification of Pathogens Recovered from the Urinary Tract,” Anal. Chem. 85(22), 10717–10724 (2013). [CrossRef]
34. O. Ryabchykov, T. Bocklitz, A. Ramoji, U. Neugebauer, M. Foerster, C. Kroegel, M. Bauer, M. Kiehntopf, and J. Popp, “Automatization of spike correction in Raman spectra of biological samples,” Chemom. Intell. Lab. Syst. 155, 1–6 (2016). [CrossRef]
35. S. Z. Li and A. Jain, “L2 norm,” in Encyclopedia of Biometrics, (Springer US, Boston, MA, 2009), pp. 883.
36. S. Gawinkowski, A. Kamińska, T. Roliński, and J. Waluk, “A new algorithm for identification of components in a mixture: application to Raman spectra of solid amino acids,” The Analyst 139(22), 5755–5764 (2014). [CrossRef]
37. M. Hedegaard, C. Matthäus, S. Hassing, C. Krafft, M. Diem, and J. Popp, “Spectral unmixing and clustering algorithms for assessment of single cells by Raman microscopic imaging,” Theor. Chem. Acc. 130(4-6), 1249–1260 (2011). [CrossRef]
38. M. E. Winter, “N-FINDR: an algorithm for fast autonomous spectral end-member determination in hyperspectral data,” in Imaging Spectrometry V, Vol. 3753 (SPIE, 1999), pp. 266–275.
39. C.-I. Chang and A. Plaza, “A Fast Iterative Algorithm for Implementation of Pixel Purity Index,” IEEE Geosci. Remote Sensing Lett. 3(1), 63–67 (2006). [CrossRef]
40. D. Heinz and Chein-I- Chang, “Fully constrained least squares linear spectral mixture analysis method for material quantification in hyperspectral imagery,” IEEE Trans. Geosci. Remote Sensing 39(3), 529–545 (2001). [CrossRef]
41. N. N. Brandt, O. O. Brovko, A. Y. Chikishev, and O. D. Paraschuk, “Optimization of the Rolling-Circle Filter for Raman Background Subtraction,” Appl. Spectrosc. 60(3), 288–293 (2006). [CrossRef]
42. K. H. Liland, T. Almøy, and B.-H. Mevik, “Optimal Choice of Baseline Correction for Multivariate Calibration of Spectra,” Appl. Spectrosc. 64(9), 1007–1016 (2010). [CrossRef]
43. K. Eberhardt, C. Stiebing, C. Matthäus, M. Schmitt, and J. Popp, “Advantages and limitations of Raman spectroscopy for molecular diagnostics: an update,” Expert Rev. Mol. Diagn. 15(6), 773–787 (2015). [CrossRef]
44. O. Ryabchykov, I. W. Schie, J. Popp, and T. Bocklitz, “Errors and Mistakes to Avoid when Analyzing Raman Spectra,” Spectroscopy 37, 48–50 (2022). [CrossRef] Publisher: MJH Life Sciences.
45. D. T. Larose, C. D. Larose, L. Chrostowski, and D. Ratner, “k-Nearest Neighbor Algorithm,” in Discovering Knowledge in Data, Vol. 9166 (John Wiley & Sons, Inc., 2014), p. 149–164.
46. N. Shukla, “Machine Learning with TensorFlow, (Manning Publications, 2018), p. 91660M.OCLC: 1027223507.
47. E. Martin, S. Kaski, F. Zheng, et al., “Sensitivity and Specificity,” in Encyclopedia of Machine Learning, C. SammutG. I. Webb, eds. (Springer US, 2011), pp. 901–902.
48. T. R. Shultz, S. E. Fahlman, S. Craw, et al., “Confusion Matrix,” in Encyclopedia of Machine Learning (Springer US, 2011), pp.209.
49. K. Rebrošová, M. Šiler, O. Samek, F. Růžička, S. Bernatová, J. Ježek, P. Zemánek, and V. Holá, “Differentiation between Staphylococcus aureus and Staphylococcus epidermidis strains using Raman spectroscopy,” Future Microbiology 12(10), 881–890 (2017). [CrossRef]
50. J. Van De Vossenberg, H. Tervahauta, K. Maquelin, C. H. W. Blokker-Koopmans, M. Uytewaal-Aarts, D. Van Der Kooij, A. P. Van Wezel, and B. Van Der Gaag, “Identification of bacteria in drinking water with Raman spectroscopy,” Anal. Methods 5(11), 2679 (2013). [CrossRef]
51. S. A. Strola, J.-C. Baritaux, E. Schultz, A. C. Simon, C. Allier, I. Espagnon, D. Jary, and J.-M. Dinten, “Single bacteria identification by Raman spectroscopy,” J. Biomed. Opt 19(11), 111610 (2014). [CrossRef]
52. M. Dahms, S. Eiserloh, J. Rödel, O. Makarewicz, T. Bocklitz, J. Popp, and U. Neugebauer, “Raman Spectroscopic Differentiation of Streptococcus pneumoniae From Other Streptococci Using Laboratory Strains and Clinical Isolates,” Front. Cell. Infect. Microbiol. 12, 930011 (2022). [CrossRef]
53. K. Rebrošová, O. Samek, M. Kizovsky, S. Bernatová, V. Hola, and F. Růžička, “Raman Spectroscopy-A Novel Method for Identification and Characterization of Microbes on a Single-Cell Level in Clinical Settings,” Front. Cell. Infect. Microbiol. 12, 866463 (2022). [CrossRef]
54. K. Rebrošová, M. Šiler, O. Samek, F. Růžička, S. Bernatová, J. Ježek, P. Zemánek, and V. Holá, “Identification of ability to form biofilm in Candida parapsilosis and Staphylococcus epidermidis by Raman spectroscopy,” Future Microbiol. 14(6), 509–517 (2019). [CrossRef]
