Mary Mulholland

Linearity and the limitations of least squares calibration

The magnitude of errors that can arise in practice from the limitations of the least squares method of calibration is estimated. Data generated from y = xn (0.7 < or = n < or = 1.3 and 1 < or = x < or = 30, or < or = 60) was analysed by least squares regression. Each y-value was then presented to the linear model and an x-value predicted. The relative errors on small x-values reached 70% of the concentration value even when r2 exceeded 0.999. Estimates of the errors on each predicted x-value, determined from the standard errors of the slope and intercept failed to reveal large errors at small x-values. Reducing the range over which linear regression is performed improved the errors. Other data sets with a heteroscedastic error distribution show that linear regression by least squares can also lead to the rejection of methods that performed sufficiently well for their application. Heteroscedastic data may be treated by repeated measurements at the lower end of the range. Data from a validation of an HPLC method for isoflavones in legumes is used to show the errors in recovery when a check sample is presented to the instrument using a calibration which satisfies the linearity tests. It is recommended that both y- and relative x-residuals are inspected. It is proposed that over-reliance on linear calibration supported by r2 may make a major contribution to large, hitherto unexplained, inter-laboratory errors.

Knowledge Acquisition without Analysis

This paper suggests that a distinction between knowledge acquisition methods should be made. On the one hand there are methods which aim to help the expert and knowledge engineer analyse what knowledge is involved in solving a particular type of problem and how this problem solving is carried out. These methods are concerned with classifying the different types of problem solving and providing tools and methods to help the knowledge engineer identify the appropriate approach and ensure nothing is omitted. A different approach to knowledge acquisition focuses on ensuring incremental addition of validated knowledge as mistakes are discovered (validated knowledge here means only that the earlier performance of the system is not degraded by the addition of new knowledge). The organisation of this knowledge is managed by the system rather than the expert and knowledge engineer. This would seem to correspond to human incremental development of expertise. From this perspective task analysis is a secondary activity related to explanation and justification not acknowledge acquisition. Ripple Down Rules is a limited example of this approach. The paper considers the possibility of extending this approach to make it more generally applicable.

Forensic classification of paper with infrared spectroscopy and principal components analysis

Fourier transform infrared (FT-IR) spectra of six document papers were classified by the chemometric technique of soft independent modelling of class analogy (SIMCA) using principal component analysis (PCA). The data were split into two regions, mid infrared (MIR, 2500–4000 cm−1) and the near infrared region (NIR, 4000–9000 cm−1). The raw data failed to produce an appreciable separation; hence it was pre-processed using derivatives and the multiple wavenumber ratios technique. The aim of this research was to determine the spectral region with the best discriminating power and evaluate the impact of data pre-processing techniques on the classification of paper. It was found that MIR (both raw and ratios) had higher discriminating ability than the NIR. The first derivatives and two ratios produced the best classifications. These results indicate that IR spectroscopy coupled with chemometric analyses can be effectively employed for the forensic classification of paper.

A COMPARISON OF CLASSIFICATION IN ARTIFICIAL INTELLIGENCE, INDUCTION VERSUS A SELF-ORGANISING NEURAL NETWORKS

Abstract Three methods of classification (machine learning) were used to produce a program to choose a detector for ion chromatography (IC). The selected classification systems were: C4.5, an induction method based on an information theory algorithm; INDUCT, which is based on a probability algorithm and a self-organising neural network developed specifically for this application. They differ both in the learning strategy employed to structure the knowledge, and the representation of knowledge acquired by the system, i.e., rules, decision trees and a neural network. A database of almost 4000 cases, that covered most IC experiments reported in the chemical literature in the period 1979 to 1989, comprised the basis for the development of the system. Generally, all three algorithms performed very well for this application. They managed to induce rules, or produce a network that had about a 70% success rate for the prediction of detectors reported in the publication and over 90% success for choosing a detector that could be used for the described method. This was considered acceptable due to the nature of the problem domain and that of the training set. Each method effectively handled the very high noise levels in the training set and was able to select the relevant attributes.

Expert system for ion chromatographic methods using dynamically coated ion-interaction separation

Abstract The development of an expert system is described for ion chromatographic methods which use ion-interaction chromatography as the separation technique. The object of the system is to help define appropriate starting conditions for the analysis of a desired group of ions. The system is implemented in a rule-based expert system development tool, Xi-Plus. Rules are used which act on certain properties of the sample and on the availability of instrumentation and accessories. With this information the method conditions can be defined for the column, detector and mobile phase. The expert system incorporates a module which allows the user to modify some of the rules in order to avoid problems which arise with expert systems which are too rigid. Many laboratories have their own preferences for columns, etc., which would be difficult for an expert system to predict. The rule change module therefore allows users to customise the system to their own requirements. Two approaches to the knowledge engineering process were employed. The first used the conventional approach of interrogation of an expert (in this case, P. R. Haddad). In the second approach, statistical analyses were applied to a previously compiled database of published ion chromatographic methods. The conclusions from these searches were then examined by the expert to define rules for the expert system. This paper describes this process of knowledge acquisition and some preliminary results on the use of the expert system.

Application of the C4.5 classifier to building an expert system for ion chromatography

A classifier based on the philosophy of induction is used to create rules for the choice of a detector for ion chromatography (IC). A data base of over 4000 instances was used to obtain suitable classifications, that were initially compiled for use in Haddad and Jacksons recent monograph on IC. The properties of an ion chromatography method that were used for rule building were solute name, separation mechanism, application, number of solutes, ion class, halide type, sulphate presence, nitrate presence, and whether a suppressor or post column detection was used. The classifier, C4.5, chooses the property that maximises the amount of information gained. By examining properties that relate to a given class, C4.5 operates a top—down induction process. A tree with 62 rules was derived from this data base. The evaluation of these rules against two test sets showed that 70% of the proposed methods were an exact match with the published methods and a further 22% gave practical methods as assessed by our expert.

Practical evaluation of ion chromatography methods developed by an expert system

As ion chromatography (IC) has matured as an analytical technique, it has become more automated. IC has not seen the abundance of automated method optimisation techniques that are provided to conventional chromatography. The authors have previously attempted to fill this gap by developing an expert system that can give comprehensive advise on IC method conditions for a variety of IC separation mechanisms. The expert system can give advice on several IC method conditions, including mobile phase, column, pH, mechanism, post column reactors, suppressor use and gradient applicability. The work in this paper describes the evaluation of the expert system including a practical evaluation of the methods, suggested by both the expert and the expert system, by running the full methods on an ion chromatograph and validating the methods. One of the features of IC is that more than one method can be suitable for a given set of analytes, differences were therefore expected in the methods suggested by the expert and those suggested by the expert system. The aim of the work presented here was to find if the expert system methods could perform in practice as well as those of the expert. Results of the validation of sensitivity, precision and limits of determination are given. The paper highlights some of the problems with expert systems developed using a database, as opposed to one developed by an expert.

Teaching a computer ion chromatography from a database of published methods

As ion chromatography (IC) has matured as an analytical technique it has become more automated. Most instrument control and data handling is now handled by computers. However, IC has not seen the abundance of automated method optimisation techniques which are provided to conventional chromatography. To a certain extent this was because IC differed greatly in the approach required to optimise selectivity and sensitivity. There was quite a diverse range of chemistries (or separation mechanisms) applicable to IC, such as ion exchange, ion interaction, etc. This paper describes an effort to fill this gap by developing an expert system which can give comprehensive advise on suitable method conditions for a variety of IC mechanisms. To build this system we applied an approach known as induction by machine learning, which was developed within the field of artificial intelligence (AI). A database of over 4000 published methods using IC, where the sample information and the chromatographic conditions were recorded, was used to train an expert system (ES). Both induction and a neural network model were applied to this task and an expert system which can advise on the following IC method conditions: mobile phase, column, pH, mechanism, post-column reactors, suppressor use and gradient applicability, was successfully developed. This paper presents a summary of the most pertinent conclusions from this study. A test set of different methods was extracted from the database and they were not applied in the training of the expert system. These were used to test the expert system and different amounts of information were used as inputs. The resulting outputs of the expert system were evaluated by the expert, who decided whether the method would work or not and if it was a good method or the ideal method for the application. Over 85% of methods were found to work and almost 62% of the methods were considered ideal. These were acceptable results when one considers the limitations of using a database of published methods as a learning set and the time saved by the use of machine learning.

Chapter 5 Ruggedness tests for analytical chemistry

Publisher Summary This chapter describes an investigation into statistical methods for the determination of the ruggedness of an analytical method as part of an overall method validation strategy. Ruggedness testing is carried out as part of a precision study and the goal is to establish the effect of small changes in the method conditions (such as temperature or instrumental settings) on the qualitative and quantitative abilities of the method. A ruggedness test allows (1) the identification of conditions, which are critical to the overall method performance, (2) the method to be documented in an unambiguous format, and the specification of system suitability criteria. The design of an overall strategy for method validation and the role of a ruggedness test in this strategy are also described in the chapter. The implementation of a ruggedness test includes aspects such as the selection of factors to test, selecting an experimental design, interpreting the results, and the final documentation of the validated method.