Introduction

Inductively coupled plasma-mass spectrometry (ICP-MS) can be traced back to 1980 [1] and it continues to be an invaluable tool for a plethora of analytical sub-fields and applications [2]. One of these sub-fields is the investigation of small entities, as a whole, in the nano- and micrometre scale: the so-called single-particle ICP-MS (sp-ICP-MS) and single-cell ICP-MS (sc-ICP-MS) [3]. sp-ICP-MS is an efficient tool in the determination of the size, number concentration, and elemental content determination of nanoparticles [4]. sc-ICP-MS on the other hand is usually focused on the study of the heterogeneity within cell populations. Through the study of intrinsic elements and/or metal tags within the cells, the study of genes, proteins, and their expression is enabled, down to a single-cell level [5]. Despite their different objectives, the principle of operation is quite similar between sp- and sc-ICP-MS: After suitable dilution, the suspension is introduced into the ICP-MS, at which point the entity is atomized into the so-called ion cloud. The ion cloud passes through the instrumentation, reaching the detector in its entirety, in a short time window. The detector is set on a comparable collection window per produced data point, generating, in the presence of a single entity within the collection window, what is called an event. Several reviews have since directly addressed the principle of operation, its benefits and shortcomings, the wide range of application fields as well as its critical aspects, and its importance to the future of micro- and nanoscale analysis [3, 6,7,8,9].

The rising interest of the community led to extensive fine-tuning and tailoring of the technique to the needs of the samples. The temporal resolution was one of the first aspects that was further explored. As a result, lower dwell times (monitoring time per data point, per m/z) were incorporated, aiming to increase the signal-to-noise ratio of the detected events (from ms to µs, even ns range) [10, 11]. As the dwell time became shorter than the duration of an ion cloud (300–1000 µs) [12, 13], events stretched from one data point per event into several. The term “transient signals” was allocated for the collection of data points on lower dwell times. From a treatment point of view, data points belonging to transient events needed not only be identified (e.g., separated from the background), but also be correctly grouped to their respective events. In that manner, the particle number can be more accurately estimated. Moreover, it was observed that the determination of the critical values used in the identification of the signals, meaning, the decision threshold, was also occasionally inaccurate during the use of lower dwell times. This was at the time attributed to the different baseline distributions exhibited for low-intensity signals, which is vital for the calculation of the critical values. That fact added one more step to the decision of the critical values, e.g., identifying the background distribution [10, 11, 14]. In addition, the quantification process of the individual events changed. In the case of a single data point per event, the height of the intensity signal is correlated to the size of the entity detected. In the case of transient events, the integral (or sum) of the data points comprising the event needs to be considered instead [10, 15,16,17]. Furthermore, the study of multiple elements in single entities highlighted the need for the acquiring of multi-elemental profiles, as this would enable the fingerprinting of the detected entities. For those purposes, the time-of-flight (TOF) mass analyser was incorporated, using a multichannel plate (MCP) detector assembly, for the quasi-simultaneous detection of multiple elements in individual events (sp/sc-ICP-TOFMS). However, due to the nature of the MCP, the baseline distribution does not follow the normal, lognormal, or Poisson distributions that were previously observed for other detector types, like electron multipliers used for quadruple-based instruments. Therefore, the process for the determination of the critical values for the detection of single entities used up to that point would prove insufficient. Later on, the MCP-generated background distribution was successfully modelled as a compound Poisson distribution through the use of a Monte Carlo simulation approach, and a more appropriate way to determine the critical values was described [18,19,20]. In addition, that process needs to happen individually for every m/z of interest, increasing either the time of the data-treating process, or the processing power needed, or both. Similarly, the information sought via a multi-elemental analysis is often derived by the correlation of the results of different isotopes, via processes like Principal Component Analysis (PCA), which adds to the workload of the data treatment approach. Finally, the size of the datasets is increasing with the addition of more isotopes of interest, further indicating the need for the upgrade of the equipment involved or the process of the data treatment [19]. Following, therefore, along the evolution of the sp-ICP-MS over the years, one can easily spot the difficulties stemming from the increasing relative size of the acquired datasets, as well as the increase in the complexity of the necessary data treatment and evaluation. Even though the results and applications have been highlighted in the previously published reviews, up until recently, little attention has been allocated to pointing out the extensive efforts of the analytical community, in terms of keeping up with the increasing demands of the acquired data. Within this article, the progression and advancements of data treatment and evaluation tools in the field of sp/sc-ICP-MS, including sp/sc-ICP-TOFMS are highlighted.

Advancements in data treatment tools

Overall, the features of the data treatment tools include either sample information, like particle number density, mass or size, or signal deconvolution options, or both. After the introduction of the multi-element single-particle analysis, the data treatment tools were likewise upgraded. In order to avoid any kind of conflicting interest, the vendor software options have been neglected within this article. However, it is noteworthy that most vendors have their proprietary options for the treatment of single-particle and single-cell-related data. They all prioritize user-friendly interfaces, smoothing their learning curve. On the other hand, they usually lack in terms of transparency, as it is unlikely for a company to divulge the inner workings of their algorithms, and flexibility, as the available options are usually specific.

Text-based interface approaches and non-specialized options

Microsoft Excel has been utilized in the early stages, owing to its universal availability and popularity [21]. In addition, it is purposefully user-friendly, enabling easy copy-and-paste solutions which are familiar to most users. Furthermore, much has been proven achievable through the incorporation of its numerous functions, and/or VBA (Visual Basic for Applications), a coding-like implementation into the Microsoft environment. “SPC” is one of the first tools that was widely available to the community and is based on Microsoft Excel. The nowadays called “SPC,” for “single-particle calculation” spreadsheet (last updated in 2015, available for download: https://www.wur.nl/en/show/single-particle-calculation-tool.htm), was developed by RIKILT of the Wageningen University and Research Centre. While it needed a manual input for the threshold of the particle detection, meaning it did not include signal treatment options, it offered a step towards automatization of the data treatment process by calculating relevant information about the input data. More recently, Laborda et al. developed a different spreadsheet that includes all the calculations necessary for the estimation and prediction of event identification threshold criteria (available for download: directly from the journal’s page on the article) [14].

The aforementioned increased complexity of the generated data led to other available options being explored, that featured higher processing power or ability to handle larger datasets, which has always been a known limitation with Excel. For instance, Igor Pro 6.37 has been incorporated in the treatment of low-dwell time acquisitions (75 µs), complementary to Microsoft Excel, for peak identification, integration, and calculation of the necessary sample information [22]. In the case of more complex treatment, like the multivariate analysis of multi-elemental datasets, OriginPro has also been reportedly incorporated in the field of single-cell analysis, as a complementary tool after the signal deconvolution has been done by vendor software [23].

Programming languages have also played a major role in modern signal analytics, owing to their flexibility and wide variety of capabilities. The implementation of that versatility enables even the potential automatization of the data-treating process, saving in terms of time and energy. The general preference turns to high-level languages (farther from the “machine” language and closer to human experience), like Python or C. As an example, for the needs of the fast data generated by their home-built data acquisition unit, Engelhard et al. have been utilizing “C” [11, 24, 25]. A dedicated command line tool has been developed for the processing of the raw data, designed to extract any information related to particles. Taking advantage of the aforementioned flexibility, programming is becoming key in the field of multi-elemental analytics. Mehrabi et al. performed their sp-ICP-TOFMS analysis solely on MATLAB, on a quite complex experimental paradigm, involving a microdroplet generator on-line coupled to a conventional ICP-TOFMS introduction system [26]. The program was designed to identify microdroplet signals by their (known) multi-elemental fingerprint, while also running the signal deconvolution procedure for the nanoparticles, using the compound Poisson modelling mentioned earlier. Python code has also been reported to be used exclusively as a signal deconvolution and sample information tool in the multi-elemental analysis of single cells, in an AF4/ICP-TOFMS approach [27]. The code was designed to extract events for every chosen element, identify events that contained more than one element, and give information about number density, relative ratios of elements, etc.

Graphic user interface approaches

The first graphic user interface (GUI) option, NanoCount, was developed by Cornelis et al., during a process that stretches over several years (available for download: https://blogg.slu.se/nanocount/). The first version of NanoCount came to be back in 2014, when it was not a GUI yet, but a piece of code written in MATLAB, yet one of the first automated tools for signal deconvolution [28]. The first interactive version of the software came after the publication of a drift-correction approach by Cornelis et al. [29]. Since 2017, the third and improved version has been available to the public, featuring options for signal deconvolution in both single- and multiple-data points per nanoparticle datasets, as well as background drift correction and transport efficiency determination. The output options include treated data, as well as histograms and relevant values like particle concentration, etc.

SPCal is a more recently developed option for the handling of single-particle data. It is a Python-based interactive platform, freely available to users as a self-contained executable file (download: https://github.com/djdt/spcal). The open-source logic promotes transparency in terms of the incorporated algorithms, to go along with the user-friendliness of a graphic user interface. Since its beginning, it accepted a variety of relevant file types, and it featured signal deconvolution options in accordance with the latest developments in the field. Regarding the outputs, it returns all relevant sample information, as well as optional exported images of the relevant histograms. There are also options for the optimal fitting of the data (normal, lognormal, etc.), as well as the exporting of the treated data. With its newest version available, it is rapidly evolving, and can nowadays accept sp/sc-ICP-MS as well as sp/sc-ICP-TOFMS data, while its processing can give out relevant histograms and fitted distributions, scatter and PCA plots, and more [30].

TOF-SPI (“TOF Single Particle Investigator”) is specialized to sp/sc-ICP-TOFMS generated data. It is developed in LabVIEW, and it is freely available as a self-contained executable file (available for download: https://github.com/TOFMS-GG-Group). It is capable of batch analysis and designed for optional calibration with an online microdroplet generator. Developed by the same group that published the newest approach for the determination of the baseline as a compound Poisson distribution in TOF instruments, it uses that same algorithm to calculate threshold values for the identification of particle events. It features several useful functions, like identifying and correcting split particle cases, background-correcting the data, etc. It generates output on several related variables, including absolute sensitivities, background information, as well as information about the mass of the user-chosen analytes. It is also able to handle the size conversion, in the case of particles, following the established procedures [31]. The approaches mentioned in this chapter are summarized in Table 1.

Table 1 Grouped data treatment approaches, with their subjective advantages and disadvantages
Full size table

Machine learning approaches

The field of single-particle analytics has been rapidly evolving, at an increasing rate, and the development of data treatment tools and options is likewise evolving as best as humanly possible. It is not, however, unrealistic to believe that there might soon come a day when information might go unnoticed, due to inadequacy in terms of extracting it from the acquired data, simply due to human inability to properly identify the emerging pattern. For that purpose, the analytical community has already begun taking advantage of the possibilities being made available through the use of machine learning algorithms [36]. Machine learning (ML) is an umbrella term, part of the evolving field of artificial intelligence (AI). Its purpose has always been to solve problems in a way that would be more time and cost efficient than a human alternative, basically by developing its own algorithmic processes. This happens by first “training” the program in the way that the problem needs to be solved, and then allowing it to apply its acquired “knowledge” to specific examples. After appropriate training, the developed algorithm should also be able to eliminate the possibility of human error related to the performing of necessary, repeated tasks. Based on this paradigm, there are three broad categories into which ML approaches can be divided: (i) In the so-called “supervised learning” approach, the program is given examples of input data, with their respective desired outputs, and its goal is to make the logical connections and develop its way to extract similar outputs of similar inputs. (ii) “Unsupervised learning” differs in the way that there is no guidance from the user, as the name suggests. Rather, the data are given to the program, which is designed to identify the patterns on its own, which sometimes is itself the task at hand. (iii) As a third approach, there is “reinforcement learning,” which includes the program operating on a task-feedback loop, to navigate a situation, trying to optimize or maximize a specific aspect of that work (f.i. driving a car). While it is true that no approach is fit for every situation, there is one thing that AI will always perform better at than the vast majority of humans: pattern recognition [37].

Supervised approaches have already been used in a variety of ways in the single-particle data treatment process. For instance, a supervised gradient decision tree boosting classification (GBC) has been used in the differentiation process between engineered and natural nanomaterials [38]. The training process of the model starts with giving it sp-ICP-TOFMS data of reference material, at which point it “learns” the process of identifying them, by choosing the parameters that contribute to the differentiation the most. The configuration of the model is considered optimum when it produces the highest median performance, using the lowest number of parameters. The then-trained model is validated through the pristine sample analysis and then applied to the real samples. Another similar approach has been incorporated in the distinction between engineered TiO2 from natural Ti nanomaterials in soil, using binomial logistic regression (BL) [39]. It was similarly trained by giving reference samples of mixed engineered and natural material. Supervised learning has also been incorporated in the analysis of airborne particles [40], while another TOF-related approach utilized a laser ablation introduction system to analyse road dust samples. In that case, a supervised classification approach was used to create a potentially fully automatized data processing pipeline [41].

On the other hand, unsupervised machine learning, as was already mentioned, is a very suitable option for identifying previously unknown patterns. For that reason, it has been incorporated into fingerprinting natural nanomaterials through the analysis of their multi-elemental profile (sp-ICP-TOFMS) [42]. Advancing the currently used methods, unsupervised clustering analysis has been incorporated in the classification of particles in wastewater treatment samples, focusing on the immense potential that the full automatization of the process exhibits [43]. In both of those cases, data have been given to the developed model, which was itself tasked with identifying the emerging patterns, among other things. Furthermore, a multi-staged, semi-supervised strategy has recently been developed for the classification of Ce-containing particles as natural, incidental, or engineered [44]. The ultimate goal of this study was to mitigate the systematic discrimination that TOFMS instruments exhibit when it comes to the moiety’s elemental composition, related to their size. Not only was the implemented strategy able to make the distinction with confidence, but it also provided the tools for the better identification of individual events that were previously not safely extracted from their background.

Despite their numerous potential advantages, machine learning approaches are highly dependent upon their training process. ICP-MS, and especially sp/sc-ICP-MS, is a complex technique, with numerous variables involved. Understanding the nature of the baseline signal can be critical, and it is a vital part of the threshold criterion choosing process, which is the cornerstone of the data-treating process [10, 14, 19, 45]. In addition, factors like variable sensitivities, elemental composition of the entity, and even similarities between training “samples” or datasets with real ones, etc. should all be considered during the model training process. Furthermore, the computer running the training lacks any kind of inherent critical thinking. That bears the risk that the in-process feedback can be misinterpreted, as the computer will only check for the parameters it is instructed to, before moving to the next iteration. Therefore, caution must be exercised when developing a machine learning approach, and a strict process for the validation of the training method should be incorporated before the results are considered dependable.

Outlook

The options of data-treating tools, when it comes to single-particle and single-cell ICP-MS analysis, have been steadily evolving over the past few years. The ever-increasing incorporation of the technique in a variety of fields has led to a proportionate increase in the demands of the community regarding their capabilities. The limitations of traditional means of data evaluation, like spreadsheets and other commercial multi-purpose software, have led to the development of more specialized solutions. This manifested in the development of both novel tools, specially designed for the respective task, and the incorporation of programming into the data treatment and evaluation process.

Nowadays, however, it is becoming clear, that the evolution and demands of the field might soon outpace the capabilities of human-driven analysis, regarding the obtaining of the maximum amount of information from the analytical process. As a result, the first artificial intelligence options have long since made their appearance and have occasionally proven superior to their manual counterparts. On the other hand, the successful incorporation of machine learning approaches, in the field of data treatment in the field of single-particle and single-cell ICP-MS, raises concerns about the quality management of the AI-generated results, and they need to be addressed before this tool can be further implemented in the analytical process.

The ultimate goal seems to be the automatization of an increasing portion of the whole analytical process, from the sample obtaining to treating the data. Thus, the capacity for the currently irreplaceable human scientific sight, intuition, and improvisation can be increased, allowing for the focusing on generating the questions, instead of attending to laborious and repetitive processes. As the analytical field evolves, so will the capabilities, and therefore, the potential benefits to every other field and discipline, both current and potentially still unrealized. The field of multi-elemental single-entity analytics, in particular, is at the moment largely underappreciated. The single-cell metallomics field has been gaining in momentum, highlighting its importance, and indicating its potential future applications. For instance, model cell organisms have been incorporated in the ecotoxicological assessment of natural systems. Furthermore, the root cause of many diseases can be linked to their elemental content, and the relative levels of different target analytes. A streamlined nanomaterial fingerprinting process, from sample introduction to data evaluation, could be of great benefit in numerous fields like forensics or archaeometry; for example, the multi-elemental profiling of archaeological samples can lead to the identification of previously unobserved patterns and correlations. The successful and dependable fingerprinting of gunpowder residue can be an invaluable tool in the field of forensics. In all of the above-mentioned cases, the treatment of the data will prove to be equally as important as the sound technical knowledge and application of the instrumentation, so that the integrity of the results can be safeguarded. Therefore, the evolution of the available data treatment options should be closely monitored, in order to create dependable tools that address the needs of the analytical procedures.