Pathogenic bacteria, fungi, viruses, and parasites can travel through various routes (air, water, physical contact, and bodily fluids) and enter their hosts (plants, animals, and humans), causing infectious diseases [1–5]. Airborne microorganisms transmitted through human speech, coughing, sneezing, and exhalation are known to have long residence times in the environment [6–12] and are responsible for the rapid and extensive spread of diseases [13–15]. The increasing incidences of infectious diseases worldwide are a significant cause for concern, pointing out the need for better control and management of diseases. The interplay of pathogenic microbes, their hosts, and the environment, coupled with the complex dynamics of disease emergence, demand rapid, direct detection and identification of airborne microorganisms to prevent transmission and the outbreak of diseases [16]. As a result, various technologies based on nucleic acid, mass spectrometry, cell structure, and optics were reported and are now in use.

Traditional pathogen detection employs morphological and biochemical techniques involving culturing with an enrichment step to isolate the target organisms on an agar plate or to detect pathogens directly from the enriched samples using advanced, sophisticated technologies and molecular techniques [17–25]. In addition, diversified technology combinations in biosensors [18], especially optical biosensors [22], have shown improved capabilities in directly detecting airborne microorganisms. However, the current molecular technologies in use for the identification of components and signal amplification of sensors [17–19], including advanced, sophisticated technologies and techniques (genomic fingerprinting [20, 21], real-time quantitative PCR [24], mass spectrometry [25, 26]), still require on-site sampling and further testing in the laboratory, involves complex operation processes, having long detection times [1–5]. Interestingly, even though the current techniques have high detection efficiency and accuracy, the sampling strategy’s subjectivity and the sampling mechanisms employed are known to have limitations affecting the accuracy of microbe detection [2].

Although on-site air sample detection features a short detection time, flexibility, and convenience, the challenges posed by the effects of in situ environmental parameters, the low concentrations of airborne microbes, and the possibility of occurrence of a broad mix of species at a particular location, lead to practical difficulties. Hence, notwithstanding their limitations and long overturn times, conventional methods and technologies are still widely used for their high specificity and sensitivity in identifying and quantifying pathogens in the air. In addition, traditional culturing is still considered a gold standard for detecting microbes due to its ability to obtain an isolated pure form of the pathogen accurately [27], even though there are difficulties associated with the isolation and maintenance of a novel unidentified or un-culturable number of bacteria and viruses present in the real-world samples.

The growing pathogenic strain variations and antimicrobial resistances demand quick and broad prevention strategies and improved countermeasures to ensure more directed therapeutic interventions through point-of-care testing for the prevention and control of infections [28–30]. Optical technologies have generated significant interest due to their high speed, high throughput, non-destructive nature, and amplification-free measurements that require minimal sample preparation [31].

The optical technologies for the study of microbes can further be categorized based on types of interaction that occur between microbes and light: UV-Visible Spectroscopy, Fluorescence (Bulk Fluorescence Spectroscopy, Fluorescence Imaging), Flow Cytometry, Vibrational Spectroscopy (Raman, Mid-Infrared, Near-Infrared), Scattering (Elastic Scattering, Dynamic Light Scattering, Dynamic Laser Speckle (Bio Speckle), Quantitative Phase Imaging, Differential Dynamic Microscopy, Video Microscopy), and Optical Coherence Tomography.

The principles of Elastic Light Scattering (ELS), which use the characteristics of the spatial distribution of the scattered light with the same wavelength of the illuminating light source, were earlier used to find bioaerosols among complex, diverse atmospheric aerosols with high sensitivity due to its increased signal strength, and the sizeable scattering cross-section, via single-particle or multiple-particle interrogation [32–36], without using specific labelling reagents [32].

All optical technologies either detect a signal for a defined wavelength using a photodiode or other broadband detector or resolve the output as a function of wavelength using a spectral sensor. The collective molecular signature from the sample (cell or a colony) makes up the spectrum, and the level of chemical specificity within a range depends on the spectroscopy type.

Imaging systems, especially time-lapse imaging systems based on CMOS imaging sensors-were used to perform early detection and classification of bacteria. However, limitations in scanning the samples due to the constraints dictated by the field of view of the imaging sensors posed difficulties in probing the volume of samples quickly and effectively. Recent research [37] overcame the limitations by using a Thin Film Transistor (TFT) based imaging sensor to build a real-time Colony Forming Unit (CFU) detection system to count bacterial colonies automatically and rapidly identify their species using deep learning.

Addressing the need for a portable and cost-effective device for long-term monitoring and quantification of various types of pollen which affect human health, Lou et al. [38] presented a label-free sensor that takes holographic images of flowing particulate matter (PM) concentrated by a virtual impactor, which selectively guides particles of a specific size (larger than 6  μ m ) to fly through an imaging window. The inline holograms cast on a CMOS image sensor due to the flowing particles illuminated by a pulsed laser diode enable particle detection based on deep learning techniques. However, observing single particles flowing in a single file through the light beam eliminates the complexity of multiple scattering and the challenging task of deconvoluting the signals from an array of particles in size, shape, and materials [33]. Although an exciting sampling strategy, the technique adopted is still not in situ, with the inability to detect smaller (<6  μ m particles remaining a practical limitation, as the smaller particles (PM_2.5 and lesser) are known to cause many adverse effects on human health.

Optics-based technologies comprise a variety of instrumental configurations by which the signal from bacteria is probed, with or without spatial information, generally from a relatively small sample volume. However, to our knowledge, all the technologies and systems using optics for bacterial identification and characterization involve in-vitro sampling methods, not in situ ones that are desirable. Moreover, the devices are neither portable nor economical for wider deployment for emerging real-world applications. Therefore, a rapid, robust, mobile, low-cost, and in situ device is necessary for bacterial detection, discrimination, and quantification for wider deployment for operational uses in the real world.

Against this backdrop, Tatavarti [39] designed and developed a photonic system AUM (Air Unique-quality Monitor) (Figure 1) capable of detecting, discriminating, and quantifying various air molecules in situ. AUM photonic system comprises a continuous wave laser light source located at the bottom on the system’s front side to illuminate the target volume of air and a position-sensing photodetector positioned at the top on the front side, to detect the backscattered laser light from the target volume. Optical filters in front of the laser source and the photodetector ensure that only the particular wavelength of light from the source passes through and falls back on the photodetector after scattering from aerosols/molecules in the target air volume. AUM system’s backscatter data observations enabled the scatterers' (dust, particulate matter, and different air molecules - which constitute the pollutants) detection, differentiation, and quantification. In the AUM photonic system, the backscattered light from a laser (635  n m , <5  m W , circular beam, TEM₀₀, unpolarized) source is designed to fall on the small measurement area on a duo-lateral position sensing photodetector. The position sensing detector (PSD) is a Lateral Effect Diode, having a sufficiently active sensing area with a wide dynamic range, whose measurements are independent of the light spot profile and intensity distribution. The centroid of the light beam is computed and supplied as electrical output signals proportional to the displacement from the centre of the detector. The PSD’s resolution is detector/circuit signal-to-noise ratio-dependent having a resolution of better than 100 nm. Analog data from the photodetector can be digitized at any sampling frequency, up to 10 𝑘𝐻𝑧, and processed using proper digital signal processing circuitry and software. After appropriate signal conditioning, the photodetector’s analogue positional and light intensity signals are digitized and stored against their observation timestamp. The data from AUM can be ported onto storage devices with wireless connectivity through a remote cloud server or wired connection to a nearby device. Therefore, AUM photonic system can yield the backscattered beam’s position in the (x, y) plane of the detector and the intensity with a sampling frequency as high as 10 kHz.

The continuous interrogation of the laser light beam from AUM, with the microbial aerosols in a small air volume, results in optical scattering. The backscattered light characteristics induced by microbial aerosols in the small interrogation volume are captured by the photodetector and monitored by AUM as a function of time. Thus, the backscattered light falling on the PSD carries an imprint of the in situ characteristics of scatterers (bioaerosols) present in the scattering volume.

As ELS can change the spatial distribution and propagation direction of the illuminating light, the information regarding scattering could be monitored and analyzed in real-time analysis for the detection and characterization of aerosol particles (encompassing the range of sizes from 100 nm to 10 mm and concentrations from parts per trillion to parts per million) [39–43], in both point and remote-sensing systems [39, 42]. Furthermore, the principles of ELS suggest that the backscattered light signature patterns carry an imprint of the scatterers' size, concentration, molecular weight, and refractive index per se in multidimensional space. Therefore, the random patterns of points in n-dimensional space, i.e., spatial point processes [44], known to play a unique role in stochastic geometry, can be a valuable model for gaining insights into the phenomena causing them.

The identification of bacteria from optical data measurements from unknown samples generally utilizes multivariate statistical methods (Least Squares, Multiple Regression, Principal Components, Support Vector Machines, Neural Networks, etc.), along with database measurements taken from known samples. Many crucial factors, like the classification level (e.g., genus, species, and strain), the range of bacterial types, and the sample preparation’s influence on variation, affect these databases.

The method most used when seeking a linear model between two variables is the standard Minimum Least-Squared approximation (MLS), also known as linear regression. However, this method has an underlying assumption that the variable chosen as independent is assumed to be noiseless. Furthermore, the underlying belief that the base series is free of noise produces a bias in favour of the base series. Thus, there is a need for an analysis technique that considers the uncertainty in both variables under comparison. One such technique, which involves the Eigen modal decomposition of the covariance matrix, is the Empirical Orthogonal Function analysis (EOF). As conventionally applied to data, empirical orthogonal function (EOF) analysis decomposes spatial and temporally distributed data into modes ranked by their temporal variances. In addition, EOF analysis allows partitioning of large data sets into signal-like and noise-like parts.

EOF analysis is also known as the Principal Component Analysis (PCA) and is quite popular in many fields of study. There are several reasons for this popularity. Probably the most important is that this method often enables a description of the variations of a complex field with a relatively small number of functions and associated time coefficients. This property is fundamental in developing statistical prediction schemes relying on multiple linear regression. The skill and statistical confidence of prediction depend heavily upon a priori methods of reducing the number of available predictors. Secondly, EOF/PCA analysis is popular because the derived empirical functions are often amenable to physical interpretation, which may give substantial insight into complex processes.

PCA accomplishes an orthogonal linear transformation that transforms the data to a new coordinate system such that the most significant variance by some scalar projection of the data comes to lie on the first coordinate (called the first principal component), the second largest variance on the second coordinate, and so on. PCA is a linear dimensionality reduction technique that transforms a set of correlated variables into a smaller number of uncorrelated variables called principal components while retaining as much variation as possible in the original dataset.

Tatavarti and Andrade [45] have demonstrated that the Eigen decomposition technique of Empirical Orthogonal Function (EOF) analysis (or PCA), and a modified Minimum Least Square (MLS) approximation technique is the same. Furthermore, the modified minimum least square approximation technique—involving a revised definition of error, compared to the standard description of the error—is independent of bias on any variable in the analysis, suggesting that PCA is effective on linear data. Of course, one can still do a PCA computation on non-linear data. Still, the results will be meaningless beyond decomposing to the dominant linear modes and providing a global linear representation of the spread of the data.

The principles of ELS suggest that the backscattered light signature patterns carry an imprint of the scatterers' size, concentration, molecular weight, and refractive index per se in multidimensional space [32, 33, 39, 41–43]. Therefore, the random patterns of points in n-dimensional space, i.e., spatial point processes [44], which are known to play a unique role in stochastic geometry, can be a valuable model for gaining insights into the phenomena causing them.

A random medium’s scattering characteristics depend on the medium’s (air, pollutants, and microbes) composition (the geometry, morphology, number, and spatial distribution of its scatterers and their refractive indices). The systems governed by non-linear interactions are ubiquitous in nature and biology. For example, anisotropy can significantly optimize the optical scattering efficiency of a volume of air laden with diverse pollutants and microbes of varying concentrations [42].

The elastic scattering of light by air laden with diverse pollutants and microbes is a function of their size, shape, molecular structure, refractive index, and concentrations of various microbes in the measurement volume and the distance of scatterers from the detector. The interactions of the scattered radiation due to all these parameters are non-linear and can only be considered linear at best under overly simplistic optimal conditions [42]. Therefore, it becomes more prudent to consider the higher-order statistical moments of scatter data to explore the relationships between optical scattering parameters and the scatterers.

Tatavarti [39] demonstrated the real-world application and use of the AUM technology to detect, discriminate and quantify the various pollutants in the air when AUM was collocated with EPA-approved state-of-the-art air quality monitoring stations (comprising multiple sensors for monitoring different parameters) in India at various geographical locations during varying environmental conditions over a long period. The outputs from the collocated AUM and the EPA-approved stations yielded comparable results indicating the effectiveness and utility of real-world applications of AUM in detecting, discriminating, and quantifying the various pollutants in the air.

Recent statistical and machine-learning methods for detecting interactions among features include decision trees and their ensembles: CART [46], Random Forests (RFs) [47], Node Harvest [48], Forest Garrote [49], and Rulefit3 [50], as well as methods more specific to gene-gene interactions with categorical features, such as logic regression [51], multifactor dimensionality reduction [52], and Bayesian epistasis mapping [53]. However, except for RFs, the above tree-based procedures grow shallow trees to prevent overfitting, excluding the possibility of detecting high-order interactions without affecting predictive accuracy [54]. RFs are an attractive alternative, looking at high-order interactions to obtain state-of-the-art prediction accuracy, even though interpreting interactions in the resulting tree ensemble remains challenging.

The AUM photonic system monitors backscattering characteristics in real-time. It determines the attributes of in situ microbes, utilizing a unique algorithm that combines concepts of stochastic geometry with those of artificial intelligence and machine learning. An extended backscatter data matrix is constructed to represent various statistical features of backscattered light data from the observations of the AUM photonic system. From an initial data matrix of M observations of three parameters [i.e., data matrix of size (M × 3)] of the backscattered light Intensity and positional information at desired sampling; N observations of backscattered light data of seven statistical parameters (four statistical moments, and the three statistical features—minimum, maximum and range), over a chosen averaging interval, are computed to form an extended, scattered data matrix of size (N × 35), containing multidimensional information. Data analytical tools and a robust machine learning algorithm (Random Forests algorithm) [47, 54, 55] enabled the extraction of information from large and high-dimensional datasets, thereby identifying structures and enabling detection and discrimination of the various microbial characteristics present in the volume of interrogation of light and matter, which were responsible for the scattering process.

We employed an innovative experimental configuration (Figure 2), where AUM was positioned in an anaerobic chamber to make optical backscattering observations from a small interrogation volume, into which microbial aerosols were introduced continuously as a function of time. To deliver microbial aerosols into the interrogation volume, we used a piston compressor nebulizer which converts the liquid in the machine’s carousel (PBS solution laden with microbes) into a fine mist due to controlled pressurized air. Ensuring that the liquid sample is homogeneously mixed and that the nebulization rate is constant, we determined the change in concentration of bioaerosols injected into the interrogation volume as a function of time.

The microbes for this study comprised nine bacterial strains and two mixtures of five bacterial species combinations. Table 1 summarizes the pathogenic bacteria, pre-injection concentrations, and data acquisition details. Table 2 highlights the characteristics of pathogenic bacteria considered for this study. The selected strains and mixtures are known to cause various infections and are resistant to antimicrobial drugs. The samples of pathogenic bacteria were cultured by adopting the gold standard protocols and procedures of cell culture. The selected pathogenic strains were brought to 37°C after retrieving from -80°C and placing them in sterile autoclaved distilled water. Luria-Bertani (LB), a widely used bacterial culture medium for various facultative organisms, was used for this study. LB broth medium was prepared and autoclaved at 121°C for 15  min . The sterile medium was inoculated with individual stains under aseptic conditions using a Laminar Airflow cabinet and incubated aerobically in the incubator at 37°C for 18–24 h. 2% of the overnight grown seed culture was inoculated in a sterile medium and set further to count the cell numbers. The grown cultures, harvested by centrifugation at 10,000 rpm for 10  min ⁡ , and the cell pellets were washed twice with Phosphate Buffered Saline (PBS) solution and resuspended in 2  m l of PBS to measure optical density (OD_600nm ) using UV Spectrophotometer (Microplate Reader, 800 TS, United States). Approximate cell numbers were calculated by the standard value of OD₆₀₀ (0.1 OD = 10⁸ cells/ml). Table 1 details the individual bacteria strains and the mixtures of bacterial species, their initial pre-injection concentrations (final cell numbers measured after 24 h incubation), cell culture particulars, protocols, and procedures adopted for their classifications using gold standards, along with durations of each trial run (Table 1). Pathogenic bacterial species and their initial concentrations for aerosol injection cover a wide range of detection sensitivity and specificity [2].

Tests were conducted by continuously injecting different microbial samples (nine different bacterial strains and two different mixtures of five bacterial species) through the nebulizer for each trial run. During that time, the AUM photonic system recorded the optical backscattered data at a sampling frequency of 1  k H z , for a duration of 20 min resulting in 12,00,000 observations for each trial. The large number of observations ensured tests for statistical repeatability, robustness, and a high degree of statistical confidence. As a result of using a nebulizer during each run, the microbial concentrations inside the anaerobic chamber continuously increased with time during each run. The increase of concentrations as a function of time was computed based on the initial concentrations of the samples, the injection rate of the nebulizer, the pre-injection volume of the sample in the carousel after dilution in PBS solution, and the time of the trial durations. It was established that the increase of microbe concentration in the anaerobic chamber was linear with time. For each sample, trials with each sample run for a duration of 20 min were conducted. All the experiments were video recorded to gain additional insights during analysis. During the tests with individual microbe strains, as the pre-injection concentrations of the samples were different, the minimum concentrations of bacterial strains injected were <10 CFU, while the maximum concentrations injected ranged from 1.6 × 10⁷ CFU to 2.3 × 10⁸ CFU, whereas in trials with two mixtures comprising five different bacterial species with the individual injection concentrations ranging from 10 CFU to 10⁸ CFU, the average initial pre-injection mixture concentrations varied from 3.1 × 10⁸ CFU to 3.2 × 10⁸ CFU. In summary, the whole range of aerosol bacterial strains, generally reported in various environments (from low concentrations to high concentrations and low specificity to high specificity), were covered during the trials.