Method for antimicrobial susceptibility testing using dynamic holographic laser speckle imaging

Abstract

The present disclosure provides label-free dynamic holographic (or interferometric) speckle imaging techniques (systems and methods), which can be used for rapid antimicrobial susceptibility testing with high sensitivity, enabling healthcare professionals to provide effective treatment to patients. Embodiments utilize a time series of holographic or interferometric images of a first bacterial sample to determine susceptibility of the bacteria. Embodiments may utilize frequency domain data. Some embodiments may utilize a machine language classifier trained with an input data set from a plurality of samples to determine a susceptibility of the bacteria.

Description

METHOD FOR ANTIMICROBIAL SUSCEPTIBILITY TESTING USING DYNAMIC HOLOGRAPHIC LASER SPECKLE IMAGINGCross-Reference to Related Applications

[0001] This application claims priority to U.S. Provisional Application No. 63 / 593,221, filed on October 25, 2023, now pending, the disclosure of which is incorporated herein by reference.Field of the Disclosure

[0002] The present disclosure relates to susceptibility testing of microbes, and in particular, susceptibility testing of microbes using holographic or interferometric imaging.Background of the Disclosure

[0003] Antimicrobial resistance (AMR) poses a critical global threat to public health, food security, and economic development, necessitating urgent advances in diagnostic methods to curb its spread. It is estimated that bacterial AMR was directly responsible for 1.27 million deaths worldwide in 2019. This number is projected to rise to 10 million by 2050 according to the World Health Organization’s estimation. AMR impacts individuals across all age groups and geographies, placing a profound burden on healthcare systems worldwide. While AMR occurs naturally in microorganisms, the misuse of antibiotics in humans and animals accelerates this process. To combat AMR, two primary' approaches can be envisioned: developing brand-new antimicrobial agents and enforcing antimicrobial stewardship. The development of new antimicrobial drugs is extremely challenging due to factors such as short drug life expectancy, substantial financial investment, stringent regulations, and unpredictable industry' responses. Given these difficulties, the proper use of existing antimicrobial agents, along with efforts to minimize the spread of resistant bacteria through appropriate infection control measures, is crucial. Addressing AMR necessitates the effective use of existing antibiotics, underscoring the critical role of antimicrobial susceptibility testing (AST) in guiding therapeutic decisions and curbing resistance.

[0004] AST methods are the collective set of diagnostic processes that facilitate the phenotypic or genoty pic assessment of antibiotic susceptibility, which is crucial for assessing bacterial drug resistance. Traditional AST methods, such as broth microdilution, disk diffusion,and gradient diffusion assays, suffer from the drawbacks of time-consuming and labor-intensive enrichment processes. These methods entail a prolonged incubation period of 6 to 24 hours for growth and an additional 1 to 3 days for isolation and confirmation through morphological, molecular, biochemical characterizations. Once an isolate is obtained, antimicrobial susceptibility testing using standard culture-based methods requires an additional 16 or more hours. This delay hinders early diagnosis and timely treatment, highlighting the urgent need for faster diagnostic approaches. While some automated systems can deliver results within 6 to 12 hours, they necessitate extensive preparation, including the use of labeling agents, large sample volumes, and incur high operational costs. Consequently, there is a significant demand for developing label-free sensing platforms that can provide rapid and sensitive AST.

[0005] Additionally, most rapid AST methods require a bacteria concentration of at least 105CFU / mL, which is too high for early diagnosis in conditions such as uncomplicated, singlespecies, urinary tract infections (UTIs) or in blood stream infections (BSI) which are the leading cause of sepsis. For example, UTIs are among the leading causes for treatment in adult primary care and account for a considerable percentage of antibiotic prescriptions. Although 105CFU / mL from urinary sample is considered as the cut-off value in laboratory testing, lower concentrations of bacteria (1O2-1O5CFU / mL) are considered clinically relevant for early diagnosis in many reports. Developing a sensitive and rapid AST method for direct sample analysis with a lower detection limit can thus accelerate the diagnosis of bacterial infections and reduce the frequency of unnecessary antibiotic prescriptions.

[0006] While many studies have demonstrated results on rapid or low-concentration AST, they often require complicated sample preparation steps such as cell isolation and device fabrication, which can be time-consuming. In previous work, we developed a machine learning based dynamic laser speckle imaging (DLSI) system capable of performing AST at a bacterial concentration of 105CFU / mL within 2 hours. DLSI captures changes in bacterial motion characteristics under antibiotic exposure, enabling quick and accurate predictions of antimicrobial susceptibility and minimum inhibitory concentrations (MIC). The MIC is defined as the lowest concentration of an antibiotic that successfully prevents bacterial growth, determining the categorical susceptibility of the bacteria to the drug. While our DLSI system meets the concentration requirements of conventional AST methods for microbial isolates, achieving rapid AST with lower bacterial concentrations remains a challenge. Developing arapid AST method that functions at lower bacterial concentrations is crucial for early disease treatment and diagnosis.Brief Summary of the Disclosure

[0007] The present disclosure provides a dynamic holographic (or interferometric) laser speckle imaging (DhLSI) system designed to improve the detection limits and reduce the time required for AST. The system can perform direct / rapid AST at bacterial concentrations as low as 103CFU / mL. To confirm the ability of DhLSI to perform standard tests, an analysis was performed with bacterial concentrations of 105CFU / mL and showed that the results were consistent with conventional methods. Embodiments of the DhLSI system use a reference beam in a holographic or interferometric detection setup to increase sensitivity to weak scattering signals, which may be advantageous for assessing bacterial responses to antibiotics at lower concentrations. The DhLSI system was validated using two model bacterial strains (Gramnegative E. coll and Gram-positive E. faecalis) and two antibiotic treatments (ampicillin and gentamicin), over multiple independent experiments. By utilizing machine learning algorithms such as, for example, Artificial Neural Networks (ANN), embodiments of the system may analyze dynamic speckle patterns resulting from bacterial motion under antibiotic exposure, allowing for determination of MIC and bacterial susceptibility. This approach offers the potential of rapid and label-free AST directly at low bacterial concentrations, shortening incubation times compared to traditional AST methods. DhLSI can also enable direct AST, i.e.. without bacterial isolation, from patients with, for example, uncomplicated UTI or sepsis with a single causative infectious agent.

[0008] The present disclosure provides label-free dynamic holographic (or interferometric) speckle imaging techniques (systems and methods), which can be used for rapid antimicrobial susceptibility testing with high sensitivity, enabling healthcare professionals to provide effective treatment to patients. Compared to existing antimicrobial susceptibility testing methods, which are typically performed at I05CFU / mL (Colony Forming Units per milliliter), embodiments of the present disclosure can be used to perform antimicrobial susceptibility testing of bacteria at low concentrations of only 10?CFU / mL (or lower). This is 2 orders of magnitude lower compared to standard broth microdilution (BMD). This enables direct detection and early diagnosis with faster antimicrobial susceptibility testing and minimal sample preparation.Methods and systems according to the present disclosure promise to facilitate bacterial infection treatment by providing an efficient and effective solution for healthcare professionals.Description of the Drawings

[0009] For a fuller understanding of the nature and objects of the disclosure, reference should be made to the following detailed description taken in conjunction with the accompanying drawings.

[0010] Figure 1. A schematic of a dy namic holographic (or interferometric) laser speckle imaging (DhLSI) system according to an embodiment of the present disclosure, (a) The laser beam interacts with (b) the bacterial culture in the cuvette, which contains varying concentrations of antibiotic, (c) The signal beam combines with (d) a reference beam to form a hologram, which is recorded with (e) a CMOS camera over (f) a total time of 10 seconds, comprising 600 frames. Following preprocessing, (g) Fourier analysis extracts information regarding bacterial motion, (h) Pnncipal Component Analysis (PCA) is utilized for validation and visualization. Subsequently, the Fourier results are inputted into an Artificial Neural Network (ANN) for (i) predicting minimum inhibitory concentration (MIC). This comprehensive embodiment integrates holographic imaging or interferometry, Fourier analysis, PCA, and ANN prediction to facilitate rapid antibiotic susceptibility testing.

[0011] Figure 2. (a) Raw speckle images captured using the DhLSI system depicting 103CFU / mL E. coll exposed to 0.25* MIC, 0.5 x MIC, l x MIC, and 2 MIC of ampicillin for 3 hours. Distinct speckle behaviors among the samples reveal bacterial motion, (b) Raw speckle images obtained with the conventional dynamic imaging system showing 103CFU / mL E. coll exposed to 0.25 x MIC, 0.5x MIC, 1 x MIC, and 2x MIC of ampicillin for 3 hours. The speckle patterns among the samples show minimal differences due to weak scattering fields from the low bacterial concentration.

[0012] Figure 3. Principal Component Analysis results of 105CFU / mL E. coll exposed to 0.25 x MIC, 0.5 x MIC, lx MIC, and 2x MIC of ampicillin for (a) 0 hours, (b) 1 hour and (c) 2 hours. The PCA analysis reveals distinct data clusters indicated generally on the figure.

[0013] Figure 4. Machine learning prediction results for (a) initial concentrations of 103CFU / mL, 104CFU / mL, and 105CFU / mL of (b) E. faecalis and (c) E. coli treated withampicillin, and (d) E. coll treated with gentamicin. Samples with initial concentrations of 105CFU / mL and 104CFU / mL were treated with antibiotics for 2 hours, while samples with an initial concentration of 103CFU / mL were treated for 3 hours.

[0014] Figure 5. Frequency response plots for (a) E. coli and (b) E. faecalis treated with ampicillin at initial bacterial concentrations of (c) 105CFU / mL, (d) 104CFU / mL, and (e) 103CFU / mL. Bacteria at 105CFU / mL and 104CFU / mL were treated with the antibiotic for 2 hours, while those at 103CFU / mL were treated for 3 hours.

[0015] Figure 6. PCA plots for (a) E. coli and (b) E. faecalis treated with ampicillin at initial bacterial concentrations of (c) 105CFU / mL, (d) 104CFU / mL, and (e) 103CFU / mL. Bacteria at 105CFU / mL and 104CFU / mL w ere treated with the antibiotic for 2 hours, while those at 103CFU / mL were treated for 3 hours. Squares represent broth data, circles represent 2x MIC data, dots represent l x MIC data, plus signs represent 0.5 x MIC data, and crosses (‘x’) represent 0.25 x MIC data.

[0016] Figure 7. Machine learning prediction results for E. coli treated with varying concentrations of gentamicin. The initial bacteria concentrations are (a) 105CFU / mL, 104CFU / mL. and 103CFU / mL. Testing was performed on samples from (b) set 1, (c) set 2, and (d) set 3. Samples with initial concentrations of 105CFU / mL and 104CFU / mL were treated with antibiotics for 2 hours, while samples with an initial concentration of 103CFU / mL were treated for 3 hours.

[0017] Figure 8. Frequency response plot for 105CFU / mL E. coli treated with varying concentrations of gentamicin for 2 hours, shown (a) before z-score normalization and (b) after z- score normalization.

[0018] Figure 9. (a) Schematic of the optical system featuring components: Beam Splitters (BS1 and BS2), Mirrors (Ml and M2), and Neutral Density Filters (NDF1 and NDF2). (b) Schematic of Raman system. HWP: half waveplate. PBS: polarizing beam splitter.QWP: quarter waveplate. M: mirror. DM: dichroic mirror. BS: beam splitter. OBJ: 100x objective lens. FL1, FL2 are focusing lenses. LPF: long-pass filter with 650 nm cut-off wavelength.

[0019] Figure 10. A diagram of a DhLSI system according to an embodiment of the present disclosure.

[0020] Figure 11. A chart showing a method according to another embodiment of the present disclosure.Detailed Description of the Disclosure

[0021] Figure 1 illustrates the working principle of an embodiment of the presently disclosed DhLSI system for rapid AST. When a coherent laser beam illuminates a bacterial suspension, which is a turbid medium, scattered light is produced. When the bacterial concentration is low, the weak scattering field is hard to detect. Thus, an intense reference beam is introduced to enable holographic or interferometric acquisition of speckles, providing a homodyne detection alternative with enhanced sensitivity compared to the direct imaging. An imaging sensor e.g, a camera) then records a sequence of holographic speckle patterns (for example, a sequence of 600 frames). A holographic speckle pattern remains unchanged in a static medium. However, in a dynamic scattering medium, as in the case of a living bacterial suspension, the holographic speckle pattern changes. This variation may be exploited for assessing the susceptibility of bacteria when subjected to antibiotics, as the evolution characteristics of the holographic patterns is linked to the bacterial motion, which is in turn is related to the inhibition status of the bacteria. In some embodiments, machine learning algorithms may be used to analyze the changing characteristics of the holographic speckle patterns, treating each pixel as an observation, with each observation containing a sequence of (for example. 600) time points.

[0022] More specifically, a reduction in bacterial motion due to the application of antibiotics leads to alterations in holographic speckle dynamics, compared to non-inhibited bacteria. Qualitatively speaking, more rapidly changing speckle patterns indicate higher levels of bacterial motion, suggesting potentially limited bacterial inhibition by the applied antibiotic. These temporal trends can be quantified in the frequency domain by performing the Fourier transform along the time axis for each pixel, and may be followed by calculating the average frequency response across all pixels (Figure 1 g). If bacteria are exposed to insufficient antibiotic concentrations, elevated frequency response in the high-frequency region (e.g., > 5 Hz) is expected. Conversely, if bacteria are exposed to sufficient levels of antibiotics, they are inhibited, resulting in lowered high-frequency response. These different behaviors in frequencyspectra can be visualized through dimensionality-reduction techniques, such as a principal component analysis (PCA) plot of the first two components (Figure 1 h). This plot shows that data from non-inhibited bacteria cluster together, while data from inhibited bacteria cluster with the broth-only medium which may be used as a negative control. By training one or more ANN models using results from, for example, a broth microdilution method as the ground truths, the dynamic speckle patterns may be analyzed through machine learning and the MIC may be determined. The ANN model(s) may categorize pixel-level observations as “inhibited” or “noninhibited.” Once a model is trained, performance may be evaluated using an unseen dataset to validate the accuracy of the susceptibility prediction. The final inhibition status of the sample may be determined by applying a majority rule, where the classification result for each pixel is aggregated, and the predominant label is assigned as the overall prediction for the sample.

[0023] While DLSI measures the speckle intensity 1 oc |Sj2, where S stands for the scattering field, the holographic system detects |7? + S|2= |7? |2+ |S |2+ RS* + R*s, with R representing the reference field and * indicating the complex conjugate. DhLSI thus benefits from the interference terms RS* and R*S , which can dominate | j2in weak scattering cases, such as low bacteria concentration scenarios (|S| « |7? |). Figure 2 depicts raw speckle images obtained using (a) the DhLSI system and (b) the system without the reference beam for imaging E. coli at 103CFU / mL exposed to ty MIC, MIC, l x MIC, and 2x MIC concentrations of ampicillin (MIC values were validated by broth microdilution methods). At the presence of the reference beam, holographic speckle patterns are clearly observed in all bacteria suspensions. In comparison, when the reference beam is blocked, raw speckle images obtained with the conventional dynamic imaging system barely detect any signal and show almost no differences among the samples.

[0024] With reference to Figure 11, in a first aspect, the present disclosure may be embodied as a method 100 for antibacterial susceptibility testing of bacteria. The method 100 may include exposing 103 a first sample of the bacteria to an antimicrobial agent. The first sample is exposed to the antimicrobial agent at a first concentration of the agent. At a predetermined time (t) after exposure, a series of holographic or interferometric images of the first sample is captured 106 over time (i.e., an imaging period). For example, a series of holographic or interferometric images may be captured 60 minutes after exposure to the antimicrobial agent and over a 10 second imaging period at a sampling rate of 60 Hz (60 imagesper second). The time after exposure may range from 0 seconds ( / .<?., immediately upon exposure) to 24 hours day or more. For example, the predetermined time after the exposure may be 0, 30, 60, 90, and 120 min, 4 hours. 6 hours, 8 hours. 12 hours. 18 hours, 24 hours, or more, or times therebetween. The sample may be imaged more than one of these times (see discussion below). The imaging period may be a predetermined time. For example, the imaging period may be < 1, 2, 5, 10, 20. 30 seconds, or more, or any time therebetween. The sampling rate may be a predetermined rate, such as, for example, 10, 20, 30, 50, 60. 100, 120, 200 Hz, or more, or any sampling rate therebetween. It will be recognized that the maximum frequency in the frequency domain may depend upon the sampling rate (z.e., according to the Nyquist sampling theorem). Each holographic or interferometric image may be a two-dimensional (2D) image — i.e., extending in an x-dimension and ay-dimension. The disclosure may generally refer to holographic imagery for convenience, but the present techniques should be understood to be applicable to and include interferometric imagery.

[0025] Susceptibility of the bacteria is determined 109 based on the series of holographic or interferometric images of the first sample. For example, the machine learning classifier may be trained to determine susceptibility of a bacteria using a plurality of bacterial samples (further described below).

[0026] The method 100 may include transforming 112 the time-based series of holographic or interferometric images to frequency domain data (a frequency spectrum). For example, the time series data at each pixel location of the holographic images may be transformed to frequency domain data. In some embodiments, a subset of the pixel locations may be transformed. In some embodiments, the time-based series of holographic or interferometric images is transformed to the frequency domain using Fourier transform.

[0027] The susceptibility may be determined by comparing the frequency domain data of the first sample with frequency domain data of a second sample. The second sample may contain the bacteria exposed to the same antimicrobial agent (as the first sample) at a different concentration, exposed to the same antimicrobial agent as the first sample and the imaging captured at a different predetermined time after exposure, or exposed to a different antimicrobial agent from that of the first sample, etc. In some embodiments, the second sample contains filtered broth (i.e., no bacteria). In this way, susceptibility may be determined based on adifference in the frequency domain data from frequency domain data of the second sample (as further discussed below).

[0028] In some embodiments, the method 100 includes normalizing 115 the frequency domain data to a DC (i.e., 0 Hz) component. In some embodiments, the method includes normalizing the frequency domain data using z-score normalization.

[0029] In some embodiments, the method includes repeating 118 the steps for one or more additional samples. Each of the one or more additional samples includes bacteria exposed to a different concentration of the antimicrobial agent from the first sample and the second sample.

[0030] In some embodiments, determining 112 a susceptibility of the bacteria includes performing 121 a principal component analy sis of the frequency domain data of the first sample, and one or more additional samples. In some embodiments, determining 112 a susceptibility of the bacteria includes inputting the frequency domain data of at least the first sample into a pretrained machine learning model which generates output data, and determining susceptibility based on the output data.

[0031] With reference to Figure 10, in another aspect, the present disclosure may be embodied as a system 10 for antimicrobial susceptibility testing of bacteria. The system 10 includes a sample holder 20 and a holographic or interferometric image sensor 30 positioned to obtain one or more holographic or interferometric images of a sample held in the sample holder. A processor 40 is in electronic communication with the holographic or interferometric image sensor 30. The processor is configured to perform any of the methods disclosed herein. For example, the processor may be configured to: receive from the holographic or interferometric image sensor a series of holographic or interferometric images of the first sample over time after exposure, wherein the series of holographic or interferometric images are captured during an imaging period; determine a susceptibility of the bacteria based on the series of holographic or interferometric images of the first sample.

[0032] The holographic or interferometric image sensor 40 may comprise a laser source 32 and a beam splitter 34 to split the beam from the laser source into a reference beam A and a sample beam B. A second beam splitter 36 may be configured to recombine the reference beam and sample beam such that the combined beam impinges on the image sensor 30.

[0033] In another aspect, the present disclosure may be embodied as a non-transitory computer readable medium storing a program configured to instruct a processor to execute any of the methods described herein. For example, the medium may include processor instructions to: obtain from a storage location a series of holographic or interferometric images of a first sample of bacteria over time after exposure to an antimicrobial agent, wherein the series of holographic or interferometric images were captured during an imaging period; determine a susceptibility of the bacteria based on the series of holographic or interferometric images of the first sample.

[0034] The term processor is intended to be interpreted broadly. For example, in some embodiments, the processor includes one or more modules and / or components. Each module / component executed by the processor can be any combination of hardware-based module / component (e.g., graphics processing unit (GPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a digital signal processor (DSP)), software-based module (e.g., a module of computer code stored in the memory and / or in the database, and / or executed at the processor), and / or a combination of hardware- and software- based modules. Each module / component executed by the processor is capable of performing one or more specific functions / operations as described herein. In some instances, the modules / components included and executed in the processor can be, for example, a process, application, virtual machine, artificial intelligence, neural network, and / or some other hardware or software module / component. The processor can be any suitable processor configured to run and / or execute those modules / components. The processor can be any suitable processing device configured to run and / or execute a set of instructions or code. For example, the processor can be a general-purpose processor, a central processing unit (CPU), an accelerated processing unit (APU). a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a digital signal processor (DSP), graphics processing unit (GPU), microprocessor, controller, microcontroller, and / or the like.Experimental (Example) Method

[0035] To visualize how DhLSI can perform rapid AST, Figure 3 shows the PCA results after E. coli (103CFU / mL; see Supporting Information (below) for results of other concentrations) are exposed to ampicillin for 0, 1, and 2 hours. The figure highlights behavioraldistinctions among bacterial samples, using the first two principal components. Within one hour of exposure, samples treated with 1 x and 2 x MIC of antibiotics closely cluster with the broth, indicating bacterial inhibition akin to the broth's motionless state. In contrast, samples treated with - 2xand - 4x MIC of antibiotics separate away from the broth, suggesting higher level of bacterial motion due to insufficient antibiotic exposure. After two hours, this pattern intensifies, demonstrating a greater separation between the inhibited and non-inhibited groups.

[0036] To validate the versatility and practicality of the presently disclosed system and methods independent expenments were conducted with bactena incubated under standardized conditions. Specifically, E. coll samples with an initial concentration of 103CFU / rnL were tested on day 1 and the same experiments were repeated nearly two months later, on days 55 and 57. All samples were treated with ampicillin for three hours. We analyzed averaged frequency responses within the 0. 1 Hz to 30 Hz range, where differences between inhibited (sufficient concentration of ampicillin) and non- inhibited (insufficient concentration of ampicillin) bacteria are most pronounced (see Supporting Information - Frequency Response Spectrum). Results revealed higher frequency responses (or elevated bacterial motion) in samples exposed to sub- MIC of ampicillin compared to inhibitory concentrations. Similarly, PCA can facilitate the visualization and initial assessment, with samples treated with l x or 2x MIC ampicillin clustering near the broth, indicating minimal bacterial activity, while those treated with non- inhibitory concentrations clustered away, suggesting heightened activity due to insufficient concentration of ampicillin (see Supporting Information - Principal Component Clustering Plot). Next, an ANN model was trained using the frequency response data within the 0. 1 Hz to 30 Hz range, comprising a total of 300 features. Considering the substantial time lapse between independent experiments on day 1 versus days 55 and 57, the model was trained using data from days 55 and 57 and tested with the unseen data from day 1, to demonstrate the robustness of the approach.

[0037] To demonstrate the generality of DhLSI, the present methodology was applied to additional bacterial species and antibiotic combinations: E. faecalis treated with ampicillin and E. coli exposed to gentamicin. Figure 4 shows the machine learning prediction results from these experiments, including the MIC values. E. faecalis samples at initial concentrations of 103CFU / mL treated with ampicillin for 3 hours were tested on days 1, 41, and 42. Samples at 104CFU / mL treated for 2 hours were tested on days 1, 43, and 46, and samples at 105CFU / mLtreated for 2 hours were tested on days 1, 22, and 23. Similarly, E. coll samples at initial concentrations of 103CFU / mL treated with ampicillin for 3 hours were tested on days 1, 55, and 57. Samples at 104CFU / mL treated for 2 hours were tested on days 1, 55, and 56, and samples at 105CFU / mL treated for 2 hours were tested on days 1, 55, and 57. Additionally, E. coli samples at initial concentrations of 103CFU / mL treated with gentamicin for 3 hours were tested on days 1, 96, and 103. Samples at 104CFU / mL treated for 2 hours were tested on days 1, 93, and 99, and samples at 105CFU / mL treated for 2 hours were tested on days 1, 93, and 99. For each scenario, a model was trained using data from two of the three days and then tested using the data from the third day. We explored all possible training / testing combinations, training the models with data from different pairs of days and testing with the remaining day. The models presented in Figure 4 were trained using data collected from the closer latter two days, and only the unseen data from day-1 samples were used as the test sets. Given that the day-1 samples were prepared months apart from those on the latter two days, using day 1 samples for testing and the other two days for training represents more distinct behavioral differences between the training and test sets. This validation approach demonstrates a ‘worst-case’ scenario, showing that even with samples prepared months apart, the models can still yield accurate MIC values in AST. All other combinations that yielded better performance are presented in the supporting information section (see Supporting Information below - Training and Testing Combination Results of Machine Learning). Transitioning from ampicillin to gentamicin revealed different bacterial responses within the two- to three-hour timeframe. Variations in responses across different days were more pronounced. To mitigate this variability, we used a normalization method similar to Z-score normalization (see Supporting Information - Z-score Normalization). This method was applied to individual frequencies within four groups corresponding to 0.25x MIC, 0.5zMIC, 1 x MIC, and 2xMIC for each day, enhancing the robustness and reliability of our models.

[0038] Table 1 compares the performance of an embodiment of the present DhLSI method with the broth microdilution method and other rapid AST methods, focusing on required cell concentration, AST time, system cost, and sample preparation complexity'. DhLSI can work with a lower initial cell concentration compared to some other rapid AST methods while maintaining a low system cost and minimal sample preparation compared to single-cell AST techniques. Additionally, embodiments of DhLSI may use only a few- optical components, compared to other optical techniques such as inverted microscopy or Raman spectroscopy.Table 1. Comparison of the present DhLSI method with other RAST and identification (ID) methods

[0039] Materials and Methods

[0040] Bacterial Culture:

[0041] E. coll K-12 (ATCC 10798) and Enterococcus faecalis ATCC 47077 were used as Gram-negative and Gram-positive model strains. Cultures were stored as frozen stocks at -80 °C and were resuscitated by streaking onto Mueller Hinton agar (MHA) and incubating streaked plates for 20 ± 2 h at 37°C. Subsequently, one isolated colony of each strain was suspended in 10 mL Mueller Hinton cation-adjusted broth (Mueller Hinton II Broth, Sigma- Aldrich, Burlington, MA) and incubated for 20 ± 2 h at 37°C statically. The optical density' (ODeoo) of each strain was measured using an Eppendorf Biophotometer (Eppendorf. Hamburg,Germany) and adjusted to 0.2, which was equivalent to approximately 108CFU / mL. Following this, the culture was diluted in MH-II broth to achieve the final concentration of 2-8 x 105CFU / mL in the broth microdilution assay plate.

[0042] Antimicrobial Susceptibility’ Assay:

[0043] Stock solutions of ampicillin (Sigma-Aldrich, CAS# 6953-4) and gentamicin (Sigma-Aldrich, CAS# 1405-41-0), were prepared by dissolving antibiotic in a sterilized Milli-Q ultrapure water. The stock solutions were aliquoted and stored at -20 °C and each aliquot was used once to prevent deterioration of antimicrobial activity' due to freeze-thaw cycles. The minimal inhibitory concentration (MIC) of ampicillin and gentamicin for the two strains was determined using the broth microdilution method and the resistance was interpreted using the breakpoints outlined in the CLSI guideline M100-ED34:2024. Briefly, wells of a 96-well microtiter plate (VWR, Radnor, USA) were filled with 50 pL of MH-II. Antibiotics were added to the wells of the first row and sequentially diluted two-fold down the row, followed by the addition of 50 pL of culture prepared as described. For each test microtiter plate, the bacterial culture in MH-II broth without antibiotics was used as a positive control and MH-II broth without antibiotics and bacterial culture w as used as a negative control. Microtiter plates were incubated at 35°C for 16 - 20 hours. The MIC for each antibiotic was defined as the lowest concentration of an antimicrobial that resulted in no microbial growth and was determined by visual inspection and measurement of the ODeoo using a microplate reader (BioTek Synergy Neo2 Hybrid Multi-Mode Microplate Reader, Thermo Fisher Scientific, Waltman, NJ). Each test was conducted in at least three independent biological replicates, each with two technical replicates.

[0044] AST Sample Preparation:

[0045] The culture for the rapid AST using DhLSI method was prepared as outlined above. Briefly, 10 mL MH-II broth was added to the culture tube and the antibiotic was spiked into the broth at the targeted concentration. Then the bacteria overnight culture population w as determined used ODeoo value and added to the broth containing antibiotics to obtain the desired concentrations of 5* 103- 5*105CFU / mL. Cultures were incubated at 37 °C and 2.5 mL of each culture was sampled every hour for Dy namic Holographic Laser Speckle imaging. Samples were discarded after imaging.

[0046] Dynamic Holographic Laser Speckle Imaging Setup:

[0047] Figure 10 is a schematic diagram of the experimental Dynamic Holographic Laser Speckle Imaging (DhLSI) system. A low-power helium-neon laser (HNLS008L, Thorlabs Inc., USA) emitting at 632.8 nm with 0.8 mW power was selected to minimize the impact on bacterial samples and prevent damage. The laser beam was expanded using a beam expander (including a pair of lenses) to achieve a 1 cm diameter. It was then split into a signal beam and a reference beam using a nonpolarizing beam splitter (BS1), with the beam diameters and the intensities further adjustable via additional beam expanders and neutral density filters (NDF1, NDF2). The signal beam passed through a cuvette with 2.5 mL bacterial suspension, capturing bacterial motion information (via modulation of the field amplitude and phase). The reference beam was combined with the signal beam using another nonpolarizing beam splitter (BS2) to generate dynamic holographic speckles, which were recorded using a CMOS camera (Zyla, Andor) for AST analysis.

[0048] Data Collection, Pre-processing, and Machine Learning Model Development:

[0049] The overnight results from the broth microdilution method were utilized as labeled data to determine the bacterial inhibition status and identify the MIC value. The MIC is defined as the lowest concentration of an antibiotic that inhibits the visible growth of bacteria. This value was determined through the broth microdilution method, which involved exposing bacteria to a range of antibiotic concentrations and observing their grow th after an incubation period, typically 16 - 20 h. Once the MIC was determined using the broth microdilution method, AST was performed with the DhLSI system using bacteria at concentrations of 103, 104, and 105CFU / mL, treated with four specific antibiotic concentrations (0.25x MIC, 0.5* MIC, 1 x MIC, and 2x MIC). For each cuvette containing the bacteria-antibiotic mixture, DhLSI measurements were performed after the bacteria and antibiotics had been mixed for 1 hour, 2 hours, and 3 hours. During each measurement, a holographic speckle video was recorded, capturing 600 frames (with 500x500 pixels per frame).

[0050] A Fourier transform was performed on the datasets along the time axis, converting the time-domain data at each pixel into the frequency domain with a resolution of 600 frequency points per pixel, spanning from -30 Hz to +30 Hz. Post-transformation, we focused solely on the positive frequency components (0 Hz to +30 Hz) since negative frequencies, being the complex conjugates of their positive counterparts, do not contribute additional information toour application. The intensity of the captured holographic speckle signal can be highly sensitive to optical setup parameters (laser power, exposure time, camera gain, etc.), making calibration and generalization challenging. To account for variations in laser power, optical system, and sample opacity, the Fourier spectra were normalized by dividing each spectrum by its DC (0 Hz) component, thereby highlighting changes in the spectral shape. The analysis utilized only the spectral intensity — representing the magnitude of the frequency components — while excluding the spectral phase information. We then applied PC A to explore variations among datasets from broth, inhibited, and non-inhibited samples, enabling both visualization and preliminary analysis. After applying a Fourier transformation along the time axis and focusing solely on positive frequencies, the data was structured as 500x500 pixels by 300 frequency components per video, resulting in 250,000 data points per cuvette.

[0051] To demonstrate the reliability’ of the results, three replicate cuvettes were prepared for each bacteria-antibiotic combination and the procedure was repeated on three different days. In each experiment, all four bacteria-antibiotic combinations were tested. The collected data from these experiments were then analyzed using an ANN model implemented with MATLAB’s Classification Learner. The ANN architecture included an input layer with 300 neurons, representing frequency components from 0 Hz to 30 Hz, a fully connected hidden layer with 300 neurons, and a binary' output layer with 2 neurons, classifying the data as ‘inhibited’ or ‘noninhibited.’ The ReLU activation function was employed in the hidden layer neurons. The iteration limit was set to 1000. and the ’Standardize Data’ option was disabled. The model was trained using data from two of the three experiments, totaling 6,000,000 data points across all antibiotic concentrations, with 80% of the data allocated for training and 20% for holdout validation.

[0052] During testing, the trained and validated model was applied to the unseen 3,000,000 data points from the third experiment. Each pixel in a cuvette w as individually assessed by the ANN to predict whether it corresponded to “inhibited” or “non-inhibited.” The final classification for each cuvette was determined through a majority voting process, aggregating the predictions from the 250,000 pixels. A classification was assigned based on whether the majority (greater than 50%) of pixels w ere predicted as “inhibited” or “noninhibited.” For example, if 70% of the pixels w ere predicted as “non-inhibited” and 30% as “inhibited,” the cuvette would be classified as “non- inhibited.”

[0053] Supporting Information

[0054] Frequency Response Spectrum

[0055] Figure 5 shows the Fourier transform spectrum results from experiments with E. coll and E. faecalis at varying initial bacterial concentrations and durations of exposure to ampicillin. The experiments were conducted for 2 hours at initial concentrations of 105CFU / mL and 104CFU / mL, and for 3 hours at 103CFU / mL. The spectrum patterns differentiate between samples exposed to antibiotic concentrations exceeding l x MIC and below l x MIC. Samples subjected to antibiotic concentrations above 1 x MIC (plotted with dots and circles) exhibit spectrum patterns similar to those of the broth, characterized by minimal high-frequency components. Conversely, samples exposed to antibiotic concentrations below l x MIC (plotted with lines and bars) display a higher intensity of high-frequency responses. This distinction underscores the utility of the frequency spectrum as a crucial indicator of bacterial activity7and its response to antibiotic treatment.

[0056] Principal Component Clustering Plot

[0057] Figure 6 illustrates PCA results obtained from experiments involving both E. coli and E. faecalis. conducted with varying initial bacterial concentrations and durations of exposure to ampicillin. The experiments were conducted for 2 hours at initial concentrations of 105CFU / mL and 104CFU / mL, and for 3 hours at 103CFU / mL. Different antibiotic concentrations are represented by distinctive markers: green squares for the broth, purple circles for 2x MIC samples, yellow dots for lx MIC samples, orange plus signs for 0.5x MIC samples, and blue crosses for 0.25 x MIC samples. PCA was used for visualization and initial assessment. Samples exposed to inhibitory concentrations (> 1 x MIC) clustered closely with the broth, indicating minimal bacterial activity. In contrast, samples exposed to sub-inhibitory concentrations (< 1 x MIC) clustered farther away, suggesting increased bacterial activity due to insufficient ampicillin. This visualization highlights bactenal activity levels across various antibiotic concentrations, with higher concentrations resulting in clusters closer to the broth and lower concentrations leading to greater dispersion.

[0058] Training and Testing Combination Results of Machine Learning

[0059] Figure 7 illustrates the machine learning prediction results for E. colt treated with vary ing concentrations of gentamicin. The initial bacterial concentrations are (a) 105CFU / mL, 104CFU / mL. and 103CFU / mL. To ensure robustness, we trained our models using data from two sets and tested them on the third set exploring all possible combinations. Notably, samples from set 1 were prepared months apart from those in sets 2 and 3, which were collected on different days. When set 1 samples were used for testing and the other two sets for training, distinct behavioral differences betyveen the training and test sets led to poorer results. This scenano represents the worst-case prediction, demonstrating that the models still yielded accurate MIC values even with significantly temporally separated samples. The remaining combinations, which produced better scores, are detailed in (c) training yvith data from sets 1 and 3 while testing with data from set 2, and (d) training with data from sets 1 and 2 while testing with data from set 3.

[0060] Z-score Normalization

[0061] We extended our investigations to include E. coli and gentamicin. Figure 8 illustrates the frequency response of E. coli exposed to gentamicin for three hours at an initial concentration of 105CFU / mL. Cultures yvere prepared on days 1, 93, and 99, revealing distinct responses compared to ampicillin. Notably, variations in response across different days were more pronounced than those observed across antibiotic concentrations. To address this variability, some embodiments of the present disclosure may use a new calibration method using z-score normalization. This method is applied to individual frequencies within four groups corresponding to 0.25* MIC, 0.5x MIC, lx MIC, and 2x MIC for each day. The findings, depicted in Figure 8, show that samples exposed to sub-inhibitory concentrations of gentamicin exhibit increased intensity in the high-frequency range compared to those exposed to concentrations exceeding l x MIC. This systematic approach enhances the clarity and consistency of our results.

[0062] Optical System

[0063] Figure 9 presents the optical systems used in this experimental embodiment. Figure 9(a) illustrates the schematic diagram of a Dy namic Holographic Laser Speckle Imaging (DhLSI) system. A low-power helium-neon laser (HNLS008L, Thorlabs Inc., USA) emitting at 632.8 nm with 0.8 mW power was selected to minimize the impact on bacterial samples and prevent damage. The laser beam yvas expanded using a beam expander (including a pair oflenses) to achieve a 1 cm diameter. It was then split into a signal and a reference beam using a nonpolarizing beam splitter (BS1), with beam diameters and the intensities further adjustable via additional beam expanders and neutral density filters (NDF1, NDF2). The signal beam passed through a cuvette with 2.5 mL bacterial suspension, capturing bacterial motility information (via modulation of the field amplitude and phase). The reference beam was combined with the signal beam using another nonpolarizing beam splitter (BS2) to generate dynamic holographic speckles, which were recorded by a CMOS camera (Zyla, Andor) for AST analysis. Figure 9(b) illustrates the experimental configuration of the customized Raman system. The excitation laser (HL63163DG, THORLABS) operated at a wavelength of 633 nm and had an output power of 100 mW. The laser beam first passed through a prism pair to convert its elliptical beam profile into a circular one. The laser power used can be adjusted using, for example, a half-wave plate (HWP) in combination with a polarizing beam splitter (PBS). Additionally, the beam passed through a quarter-wave plate (QWP), used to suppress unintended back reflection. Subsequently, the laser beam was directed to a lens pair to expand the beam so that it filled the back aperture of a 100X Nikon CFI60 objective lens with a numerical aperture of 0.8, achieving a diffraction limited focusing spot size on the sample. For sample monitoring, an imaging subsystem was employed, comprising a beam splitter (BS), a focusing lens (FL1), and a charge-coupled device (CCD) camera.

[0064] A dichroic mirror (DM) (RET6381p, Chroma) was utilized to reflect the laser towards the sample while transmitting the Raman signals to the spectrometer (SpectraPro-25001) equipped with a liquid nitrogen cooled CCD camera. A 650 nm long-pass filter (650 LP) was employed to reject the residual 633nm excitation light. All measurements were conducted under the condition that the laser power on the sample was 8 mW and the spectrometer integration time was set to 30 seconds.

[0065] Additional Examples

[0066] Example 1. A method for antibacterial susceptibility testing of bacteria, comprising: exposing a first sample of the bacteria to an antimicrobial agent at a first concentration; capturing a series of holographic or interferometric images of the first sample over time after exposure, wherein the series of holographic or interferometric images are captured during an imaging period; and determining a susceptibility of the bacteria based on the series of holographic or interferometric images of the first sample.

[0067] Example 2. The method of example 1 , wherein determining a susceptibility of the bacteria comprises using a machine learning classifier applied to input data comprising the series of holographic or interferometric images of the first sample.

[0068] Example 3. The system of example 2, wherein the machine learning classifier has been trained with a plurality of samples of the bacteria exposed to antimicrobial agents.

[0069] Example 4. The method of any one of examples 1 or 2, wherein determining the susceptibility of the bacteria further comprises transforming the time-based series of holographic or interferometric images to frequency domain data; and wherein the machine learning classifier is applied to the frequency domain data.

[0070] Example 5. The method of example 4, wherein the time-based series of holographic or interferometric images is transformed to the frequency domain using Fourier transform.

[0071] Example 6. The method of example 4, further comprising normalizing the frequency domain data to a DC component.

[0072] Example 7. The method of example 4, further comprising normalizing the frequency domain data using z-score.

[0073] Example 8. The method of any one of examples 1 or 2, wherein the steps are repeated for one or more additional samples, each of the one or more additional samples exposed to a different concentration of the antimicrobial agent from the first sample.

[0074] Example 9. The method of example 8, wherein determining a susceptibility of the bacteria comprises performing a principal component analysis of the frequency domain data of the first sample, the second sample, and the one or more additional samples.

[0075] Example 10. The method of example 1, wherein the imaging period is taken at a predetermined time after exposure.

[0076] Example 11. A system for antibacterial susceptibility testing of bacteria, comprising: a sample holder; a holographic or interferometric image sensor positioned to obtain one or more holographic or interferometric images of a sample held in the sample holder; and a processor in electronic communication with a storage device, wherein the processor isconfigured to: receive from the holographic or interferometric image sensor a series of holographic or interferometric images of the first sample over time after exposure, wherein the series of holographic or interferometric images are captured during an imaging period; determine a susceptibility of the bacteria based on the series of holographic or interferometric images of the first sample.

[0077] Example 12. The system of example 11, wherein the processor is further configured to determine a susceptibility of the bacteria using a machine learning classifier applied to input data comprising the series of holographic or interferometric images of the first sample.

[0078] Example 13. The system of example 12, wherein the machine learning classifier has been trained with a plurality of samples of the bacteria exposed to antimicrobial agents.

[0079] Example 14. The system of any one of examples 11 or 12, wherein the processor is further configured to determine the susceptibility of the bacteria by transforming the timebased series of holographic or interferometric images to frequency domain data and applying the machine learning classifier to the frequency domain data.

[0080] Example 15. The system of example 14, wherein the processor is further configured to transform the time-based series of holographic images to the frequency domain using Fourier transform.

[0081] Example 16. The system of example 14, wherein the processor is further configured to normalize the frequency domain data to a DC component.

[0082] Example 17. The system of example 14, wherein the processor is further configured to normalize the frequency domain data using z-score.

[0083] Example 18. The system of any one of examples 11 or 12, wherein the processor is further configured to repeat the steps for one or more additional samples, each of the one or more additional samples exposed to a different concentration of the antimicrobial agent from the first sample.

[0084] Example 19. The system of example 18, wherein the processor is further configured to determine a susceptibility of the bacteria by performing a principal componentanalysis of the frequency domain data of the first sample, the second sample, and the one or more additional samples.

[0085] Example 20. The system of example 11, wherein the imaging period is taken at a predetermined time after exposure.

[0086] Although the subject matter is presented in terms of certain examples and embodiments, these examples and embodiments are intended to be non-limiting, and other examples and embodiments, including examples and embodiments that do not provide all of the benefits and features set forth herein, are also within the scope of this disclosure. Various structural, logical, process step, and other changes may be made without departing from the scope of the disclosure.

Claims

What is claimed is:

1. A method for antibacterial susceptibility testing of bacteria, comprising: exposing a first sample of the bacteria to an antimicrobial agent at a first concentration; capturing a series of holographic or interferometric images of the first sample over time after exposure, wherein the series of holographic or interferometric images are captured during an imaging period; and determining a susceptibility of the bacteria based on the series of holographic or interferometric images of the first sample.

2. The method of claim 1, wherein determining a susceptibility of the bacteria comprises using a machine learning classifier applied to input data comprising the series of holographic or interferometric images of the first sample.

3. The system of claim 2, wherein the machine learning classifier has been trained with a plurality of samples of the bacteria exposed to antimicrobial agents.

4. The method of claim 2, wherein determining the susceptibility of the bacteria further comprises transforming the time-based series of holographic or interferometric images to frequency domain data; and wherein the machine learning classifier is applied to the frequency domain data.

5. The method of claim 4, wherein the time-based series of holographic or interferometric images is transformed to the frequency domain using Fourier transform.

6. The method of claim 4, further comprising normalizing the frequency domain data to a DC component.

7. The method of claim 4, further comprising normalizing the frequency domain data using z- score.

8. The method of claim 2, wherein the steps are repeated for one or more additional samples, each of the one or more additional samples exposed to a different concentration of the antimicrobial agent from the first sample.

9. The method of claim 8, wherein determining a susceptibility of the bacteria comprises performing a principal component analysis of the frequency domain data of the first sample, the second sample, and the one or more additional samples.

10. The method of claim 1, wherein the imaging period is taken at a predetermined time after exposure.

11. A system for antibacterial susceptibility testing of bacteria, comprising: a sample holder; a holographic or interferometric image sensor positioned to obtain one or more holographic or interferometric images of a first sample held in the sample holder; and a processor in electronic communication with a storage device, wherein the processor is configured to: receive from the holographic or interferometric image sensor a series of holographic or interferometric images of the first sample of bacteria over time after exposure to an antimicrobial agent, wherein the series of holographic or interferometric images are captured during an imaging period; determine a susceptibility' of the bacteria based on the series of holographic or interferometric images of the first sample.

12. The system of claim 11, wherein the processor is further configured to determine a susceptibility of the bacteria using a machine learning classifier applied to input data comprising the series of holographic or interferometric images of the first sample.

13. The system of claim 12, wherein the machine learning classifier has been trained with a plurality of samples of the bacteria exposed to antimicrobial agents.

14. The system of claim 12, wherein the processor is further configured to determine the susceptibility of the bacteria by transforming the time-based series of holographic or interferometric images to frequency domain data and applying the machine learning classifier to the frequency domain data.

15. The system of claim 14, wherein the processor is further configured to transform the timebased series of holographic images to the frequency domain using Fourier transform.

16. The system of claim 14, wherein the processor is further configured to normalize the frequency domain data to a DC component.

17. The system of claim 14, wherein the processor is further configured to normalize the frequency domain data using z-score.

18. The system of claim 12, wherein the processor is further configured to repeat the steps for one or more additional samples, each of the one or more additional samples exposed to a different concentration of the antimicrobial agent from the first sample.

19. The system of claim 18, wherein the processor is further configured to determine a susceptibility of the bacteria by performing a principal component analysis of the frequency domain data of the first sample, the second sample, and the one or more additional samples.

20. The system of claim 11, wherein the imaging period is taken at a predetermined time after exposure.

21. A non-transitory computer readable medium storing a program configured to instruct a processor to execute the method of any one of claims 1-10.

Images