Chemometrics in Environmental Chemistry - Applications

Library search is an important tool in the computerized qualitative analysis of mixtures and in the identification of unknown substances. It involves the comparison of the sample spectrum with the entries of a spectra collection. The principles of this methods, the design and the structure of spectra libraries and the manifold of different search strategies are explained and some important applications demonstrated. Besides the identification of pure substances the characterization of mixtures with and without a separation step are mentioned.

A fast, personal-computer based pattern recognition/expert system approach for classifying, identifying and estimating molecular weights of toxic and other organic compounds from their low resolution mass spectra is described in this article. This chemometric approach can supplement manual or computer-aided techniques for identifying organic pollutants, including library searches. The system was designed for spectra of low concentration pollutants and will give some information on mixtures. A sequential design is used with a pattern recognition classifier followed by molecular weight estimators, exclusion filters and identification modules for each of six classes. The target classes are nonhalobenzenes; chlorobenzenes; bromo- and bromochloro-alkanes/alkenes; mono- and dichloro-alkanes/alkenes; tri-, tetra- and pentachloro- alkanes/alkenes and unknown. The expert system provides a chemical class, a molecular weight estimate and a lower limit to the molecular weight for a given spectrum. This system has been extensively tested with 630 reference spectra from two databases and with 37 GC/MS field spectra. Classification and identification accuracies were 95–97%. Median absolute deviations from the true molecular weights were 1–2 Da and average absolute deviations were 6–10 Da. The approach is applicable to any type of low resolution mass spectra and can be used in survey studies or for quality assurance activities.

The problem of source identification and quantitative mass apportion for airborne particulate matter commonly called receptor modeling is much the same as the spectrochemical mixture or the multivariate calibration problems commonly encountered in chemometrics. However, there are some important differences that occur for the receptor modeling problem including variability in the source profiles (corresponding to the spectral properties of the mixture components), transformation of the profiles while in transit from the emission sources to the sampling sites, and much larger errors (noise) in the sampling and analysis of the mixtures. Various mixture resolution approaches have been used including multiple linear regression (Chemical Mass Balance Model) and various type of factor analysis. The basis for these methods is presented and illustrative examples of each are given. In addition multivariate calibration methods including partial least squares, the genetic algorithm, and artificial neural networks are now being explored as ways to solve the receptor modeling problem. These methods as well as the limited applications of these approaches are also reviewed. The initial results with these methods are promising, but further development and testing are needed before they can be used as commonly as the chemical mass balance or factor analysis models.

Various applications and facets of multivariate data analysis applied to the solution of petroleum geochemical problems are described. First, some important techniques for the analysis of multivariate geochemical data are briefly discussed. Principal component, partial-least-squares regression, marker and target projections are defined within the same formal scheme of latent-variable projections. Applications of these techniques to find solutions to important geochemical problems or to gain insight into geochemical processes in certain systems of petroleum geochemistry are then discussed in detail. These applications show that latent-variable analysis represents an efficient and important tool in petroleum geochemical problem solving.

A chemometric strategy for quantitative structure-activity and structure-property relationship (QSAR, QSPR) analysis in environmental chemistry is outlined. In essence, the strategy is based upon grouping chemicals into homogeneous classes, and identifying small numbers of training set and validation set compounds that adequately represent each class. Biological and environmental response data generated for such sets of representative compounds may then be sufficient for constructing statistically sound models relating the variation in responses to the differences in chemical properties. The use of such QSAR or QSPR models may also allow the prediction of missing response data for untested compounds in the relevant class, and thus enable priority setting for further biological or environmental testing in relation to the predicted severity of these responses. Four representative examples of how QSAR and QSPR can be formulated for environmentally relevant classes of compounds and meaningful responses are addressed. Two of the illustrated applications are also developed within the framework of the chemometric QSAR strategy.

It is demonstrated elsewhere that analytical results, collected for environmental studies, must be consistent and validated in order to allow a correct statistical processing. Here, method validation is presented, and, while it requires a lot of statistics, it is a field where chemometrics have been successfully applied. A very interesting approach for method validation and laboratory evaluation is interlaboratory study. But, the classical theoretical background of repeatability and reproducibility may be insufficient to extract all necessary information, as it is based on a univariate model, whereas a multivariate approach would be more helpful.

The basic principles of interlaboratory studies and the ISO 5725 standard model are explained as it is a recognized technique for assessing method precision and trueness. In that context, measurement is split into three parts: the hypothetical true value of the sample, the bias of the laboratory (systematic error) and the experimental error (random error). From this model, two criteria are developed in order to estimate the precision of the method: the repeatability and the reproducibility, respectively noted r and R. In addition the principles of accepted tests for rejecting outliers are presented as well as some complementary applications of repeatability and reproducibility for comparing and selecting methods. The importance of using a graphical technique for interpreting the results of interlaboratory studies is also underlined as a recommended way to make decisions.

In conclusion a sophisticated example based on the analysis of mono and disaccharides by HPLC, involving 20 laboratories on 12 food samples is presented. It demonstrates that some criticism can be made of the traditional ISO 5725 standard. For example, the ANOVA model is extended for bias component decomposition. It shows that the computational method used to quantify the analyte (based on the peak area or the peak height) is very influential in some cases.

Although the primary goal of the ISO 5725 standard was to compute the precision for one method, it is also applicable to individual laboratory evaluation, using different statistics which are explained and illustrated. But laboratory ranking is much more evident when using multivariate statistical methods. Multiple Correspondence Factor Analysis is presented in detail and applied to a collaborative study on nitrates. When applying this technique, an appreciable loss of information subsequent to the binary transformation of raw data can be suspected. An optimal algorithm is presented that gives very interesting results.

Besides method validation, proficiency tests are used to evaluate the method trueness. This is obtained by comparing individual laboratory results to the conventional reference value of the sample. Different algorithms (parametric and non-parametric) classically used to compute the reference value are compared. It shows that the lack of standardization may lead to important discrepancies. On the other hand some guidelines are illustrated on a practical example for the study of sample homogeneity. For this application, four institutes collaborated on the study, including about 600 laboratories, and about 3 300 results were collected.

However, method validation raises the question of traceability in analytical chemistry. There is an evident lack of certified reference materials in comparison to the international standards used for physical measurements. It is difficult to guess how this situation will improve, but ring tests can be an element for ascertaining the metrological quality of laboratories and analytical methods. It must be underlined that standardization will be a key principle in defining recognized statistical techniques.

During the last few decades environmental studies have been performed on many occasions. Sometimes they just consist of collecting data for “accidents” occurring randomly but they can also be world-wide concerted action, aiming to monitor pollution and involving several laboratories. The author describes the scientific background to these environmental studies in order to explain the concepts that can be used to implement consistent databases which permit meaningful statistical processing and as correct as possible decision-making.

To start with, it is useful to establish a classification system for the various kinds of data (factual, referential and documentary). This classification is substantial for providing a rigorous structuring of data. This will considerably reduce the possible lack of data consistency and simplify the later statistical processing. According to the modern theory of the entity-relationship model, data can be modelled before being stored in a database. This modelling confers a fairly flexible structuring capability for data retrieval and validation. The basic theory of entity-relationship modelling is presented and illustrated for two typical environmental studies. The importance of designing efficient coding systems is largely addressed as it must be the result of several considerations: simplicity, accuracy and convenience.

All described concepts can be implemented in a Laboratory Information Management System (LIMS). Moreover, it appears that good automation of laboratory management is based on the definitions of standard procedures which relate to Quality Assurance and Good Laboratory Practices (GLP). Quality Assurance can be considered as a general policy based on international standards. But practical applications vary from one organisation to another. Its fundamental goal consists in bringing proofs of the correct functioning of the laboratory. This goal is complementary to the data modelling.

Modern Relational Database Management System (RDBMS) software contains many utilities that renders the practical development of LIMS easier. All that was described before, in general terms, must be adapted to the specific study that is planned. It is then recommended to write a document that will clearly indicate the goal and the frame work of the study. It can be totally or partially used as a requirement document for the software that will manage the data.

Statistical processing for analytical chemistry is faced with many problems where chemometrics is involved. It is usual to consider separately univariate and multivariate statistical methods. Univariate methods can be very poor when dealing with large numbers of variables. Some guidelines are presented for selecting multivariate methods according to the nature of the information, in order to avoid any misinterpretation of the results. When dealing with large data sets, several variables types may be encountered: it can then be necessary to transform some variables in order to obtain an homogeneous data array.

Two examples of multivariate environmental data processing are presented. The first consists of a study of soil pollution by heavy metals. As a preliminary step, non-linear regression was used in order to predict soil pollution around one source and delimit the hazardous area. For a restricted area this gives consistent results. Thereafter, multiple linear regression was applied in order to estimate the relative influence of different pollution sources on the total study area.

The second example deals with the problem of mercury pollution of rivers in the Alsace region. An original multivariate method is presented, called Multiple Correspondence Factorial Analysis. The double weighing technique applied allows a simultaneous comparison of different qualitative pollution mechanisms: sampling point location along the river, fish species, fish weight and sampling year. The last chapter gives general guidelines on commercially available statistical software.

The goal of this work is to demonstrate that, when the data processing phase arrives, it is evident that the fulfillment of an environmental study must be concerted. Poor conclusions are extracted from poor data and it is compulsory to orient data collection by strong initial hypothesis. Computerisation of data, if correctly done, is the best way to reach this goal.

The role of a laboratory and how laboratory automation can help achieve the aim of a laboratory are discussed. Organising scientific data and converting it into information are aims of both chemometrics and laboratory information management systems (LIMS). A discussion of scientific data and its conversion into information precedes a new approach to laboratory automation. This is necessary to see how chemometrics and LIMS can be merged effectively. A new definition of laboratory automation is proposed. The constituent groups are instrument automation, communications, data to information conversion and information management. The integration of all groups is essential for an effective laboratory.

A LIMS is presented as a conceptual model consisting of a database core surrounded by four user segments of data capture, data analysis, reporting and management. One principle of the model is that any function of a LIMS can be classified in one of these four areas. Of specific interest to chemometricians is the data analysis area of the LIMS model; here standard chemometric tools can be used such as PCA, PCR etc. to reduce data to information. The LIMS will aid the process by ordering the data and passing them to the analysis software and accepting the information into the database if required. The systems development life cycle of a LIMS is described from the project proposal, through writing the requirements specification, the selection of a system and finally the implementation and operation of the system in the laboratory.

Techniques for the automated monitoring of water quality during the treatment of potable and waste water are reviewed with a brief description of the operating principles of various types of instrumentation followed by a consideration of their integration into process control. The monitoring of river water quality is also considered from the wider point of view of analyses that give information on water quality rather than simply process control parameters.

The review also considers related topics such as instrument specification, experiences of process control in various countries and takes a brief look at how the output from automated analyzers may be integrated with information in the form of expert systems and water quality models.

Finally, the way in which laboratories are likely to change in the near future as a result of the introduction of robotic systems of analysis is discussed with illustrations from recent literature articles.