A liquid handling system for the automated acquisitionof data for training, validating and testing calibration models

Edward Richards, Conrad Bessant*, Selwayan SainiCranfield Centre for Analytical Science, Institute of BioScience & Technology, Cranfield University, Silsoe, Bedforshire MK45 4DT, UK

Received 17 May 2002; received in revised form 9 September 2002; accepted 24 September 2002

Abstract

Multivariate calibration of a single sensor for many mixed analytes is useful, but generating, validating and testing the calibration models

requires large data sets which can be time consuming to collect, particularly when the number of analytes is large. In this paper, a solution to

this problem is presented, in the form of an automated liquid handling system capable of producing mixtures and triggering an external

analytical instrument to analyse them. As a case study, the electrochemical technique of dual pulse staircase voltammetry (DPSV) was used to

collect data for three analytes (glucose, fructose and ethanol) mixed in varying concentrations. Artificial neural networks (ANNs), optimised

using genetic algorithms were used to create the best possible multivariate calibration model. The liquid handling system performed 1668

experiments used in the study in approximately 60 h, compared to over 2 weeks that would be required to collect perform the experiments

manually. The best calibration models produced from the data bettered those produced from previous manually collected data [1,2] for all the

analytes concerned.

# 2002 Elsevier Science B.V. All rights reserved.

Keywords: Liquid handling; Automation; Multivariate calibration; Neural networks; Multi-analyte; Electrochemistry

1. Introduction

In a previous paper [2], we showed how an elitist genetic

algorithm could be used to optimise the parameters of a

neural network to train the best possible multivariate cali-

bration models for voltammetric data acquired from mix-

tures of ethanol, fructose and glucose. The data, 1064

experiments, was produced by hand in a laborious pipette

procedure, susceptible to variability and human error. The

1064 experiments took a suitably trained person 2 weeks to

perform. To reduce the time taken and potential human

induced error, a computer-controlled automated liquid hand-

ling system has been designed and built. The system is

capable of performing four factor nine level fully factorial

experimental designs, producing each individual sample in a

random order to negate systematic errors caused by tem-

perature variation or electrode drift.

In the case study presented here, the system was used to

produce aqueous mixtures of ethanol, fructose and glucose,

which it then passed through a flow cell. After the generation

and mixing of each sample, the system triggered a potentio-

stat to perform an electrochemical analysis using dual

pulse staircase voltammetry (DPSV) [3]. The system devel-

oped here differs from a standard flow injection system

by having the ability to automatically mix different con-

centration permutations of multiple analytes, buffer and

carrier in discrete sample volumes which are then passed

through the detector. There is no detection of concentration

gradient as the detection is performed under a stopped flow

condition after half the sample has passed through the

detection cell.

2. Materials and methods

2.1. The automated liquid handling system

The automated liquid handling system (Fig. 1) uses a

series of calibrated diaphragm pumps (Bio-Chem Valve Inc.,

USA) to create a 1-ml sample containing up to four different

analytes, buffer and water. Each of the analyte and buffer

pumps (Figs. 1 and 2) has a set volume of 20 ml. The mix and

cell pumps each have set volumes of 50 ml. Varying the

number of strokes of each of the analyte pumps from 0 to 8

Sensors and Actuators B 88 (2003) 149–154

* Corresponding author. Tel.: þ1-525-863-358; fax: þ1-525-863-540.

E-mail address: c.bessant@cranfield.ac.uk (C. Bessant).

0925-4005/02/$ – see front matter # 2002 Elsevier Science B.V. All rights reserved.

PII: S 0 9 2 5 - 4 0 0 5 ( 0 2 ) 0 0 3 1 8 - 0

times per sample allows up to 94 ¼ 6561 different mixture

permutations of the four analytes to be produced. The pump

switch interface supplies a buffered power supply to the

pumps via transistors and also powers the mixing motor

using a bi-stable latch. The pump switch interface was

controlled directly via the PC parallel port using a program

written in Cþþ Builder (Borland, USA), running under

Windows 98 (Microsoft, USA) on an Athlon 1 GHz pro-

cessor (Advanced Micro Devices, USA).

The RO water tanks were two 10 l glass vessels linked via

a siphon tube. Their combined surface area of 0.067 cm2

meant that there was very little change in pressure in the

water line over a large number of experiments. The surface

of the RO water in the tanks was maintained approximately

3 cm above the inlet of the mixing pump, so as to provide a

small positive pressure at the inlet to the pump for the

analyte and buffer pumps to work against, without changing

their performance characteristics. This would ensure that the

water line was always flooded and that the sample was held

close to the mixing pump when injected and not able to flow

out into the RO water tanks.

2.1.1. Sample production

To produce a sample, the buffer pump first injects 10

aliquots (0.2 ml) of concentrated buffer into the water pipe

(see Fig. 2), displacing the water back towards the water

Fig. 1. Photograph of the automated liquid handing system.

Fig. 2. Schematic of the automated liquid handling system, showing the major components.

150 E. Richards et al. / Sensors and Actuators B 88 (2003) 149–154

tank. Aliquots of the analytes in a concentrated form are also

injected in the same way into the water line depending upon

the mixture to be produced. Again they displace the water

and now the buffer back towards the water tank (see Fig. 2).

The maximum volume that could be displaced back towards

the tank is the maximum concentrations of all the analytes (8

aliquots each) plus the buffer (0.84 ml). The water pipe has a

1.5 mm i.d., hence a 1-ml sample will be held in 56.6 cm of

the water pipe. To ensure that the sample cannot contaminate

the water tank, the water pipe is of 1-m length. Having

loaded the water line with buffer and analytes, the mix pump

pumps a 1-ml volume (20 aliquots) containing the buffer,

analytes and water into the mixing cell. The bi-stable latch

starts the motor-driven stirrer, which mixes the sample as it

enters the mixing cell. When the 1-ml volume has been

pumped into the mixing cell and mixed for three seconds, the

stirrer is stopped and the cell pump starts to draw half of the

homogenised sample through the thin layer flow cell. After

half the sample has passed through the flow cell, pumping

stops, the potentiostat is triggered and a voltammetric

measurement is taken and stored for the sample. The rest

of the sample is then drawn through the flow cell. Two wash

cycles are then performed in tandem, which consist of the

sample production process as described above, but without

the injection of analytes or the voltammetric measurement.

Washing is performed to prevent carryover from measure-

ment to measurement.

2.2. Sample production error

The pump manufacturer’s data sheet specifies a value for

aliquot repeatability of �4% of the set volume for each

pump. A simulation of a 9 level, 3 factored, fully factorial

experimental design (729 experiments) was run in Matlab

6.1, where the assumption that the �4% of pump set

volume repeatability represented 3 S.D. of the normally

distributed aliquot population. The simulated errors were

calculated by summing, for the individual analytes and

buffer concentrations required (for each simulated sample)

the number of individual 20 ml volume aliquots the analyte

and buffer pumps would inject into the system manifold

(Fig. 2). To each simulated aliquot a random normally

distributed volume error was added, using a standard

distribution, s of �0.2666 ml (1.333% of 20 ml) and zero

mean. Having ascertained the simulated volumes injected

into the manifold, 20 aliquots of the 50 ml mix pump were

simulated (a 1-ml volume in total). Each simulated aliquot

had a randomly distributed volume error added with stan-

dard distribution, s of �0.666 ml (1.333% of 50 ml). The

simulated volume of analytes and buffer were subtracted

from the simulated volume pumped by the mix pump, to

determine the volume of additional water added to the

sample. The simulated concentrations of analytes and

buffer in the final solution could then be calculated. The

simulated RMS error for each analyte and the buffer was

then calculated by taking the square root for the average of

all the simulated concentrations subtracted from the

required concentrations squared for each analyte. The

simulated RMS error for each analyte, per sample pro-

duced, was �0.38% of required maximum analyte con-

centration. The simulated RMS error for the buffer, per

sample produced, was �0.1% of required sample buffer

concentration.

2.3. Materials

The analyte mixtures and buffer were produced in a

concentrated form so that when injected into the water pipe

they would be diluted to the concentrations of interest.

Ethanol (HPLC grade 99.97%) (Merck, USA), D-fructose

(AnalaR) (BDH, UK), D-glucose (General Purpose Reagent)

(BDH, UK) and sodium hydroxide pearls (min 98%)

(Sigma–Aldrich) were used. All reagents were dissolved

in RO water. As the volume of each solution is limited to

1 ml, the weight/volume requirement of all analytes for all

the experiments is small.

2.4. Dual pulse staircase voltammetry

The applied potential waveform is the same as that

described by Bessant and Saini [1], except that than the

duration of the electrode rejuvenation steps for both oxida-

tion (þ0.7 V) and reduction (�0.9 V) were set to 1 s. The

measurement was taken in a thin layer flow cell (20 ml)

(BAS, USA) at the surface of a 3 mm diameter platinum

electrode (BAS, USA). The counter electrode was identical

and adjacent to the working electrode (BAS, USA). The

reference electrode was Ag/AgCl (BAS, USA). The mea-

surement was taken in a stopped flow condition using an

Autolab PSTAT10 (Echo Chemie B.V., Netherlands).

2.5. Data sets

Training data sets were of a three factor five level fully

factorial design as described in Richards et al. [2]. The

maximum and minimum concentrations were 12 mM and

0 mM for ethanol, 0.68 mM and 0 mM for fructose and

0.72 mM and 0 mM for glucose. Three repeat sets of

training data were collected either side of the validation

set (6 � 125 experiments in total). The validation data set

was comprised of a repeat of the training data design (125

experiments) with an additional three factor four level fully

factored design (64 experiments) data set at concentrations

between those used for the training data. The maximum and

minimum concentrations were, for the additional part of

the validation data set, 10.5 mM and 1.5 mM for ethanol,

0.595 mM and 0.085 mM for fructose and 0.63 mM and

0.09 mM for glucose, respectively (64 experiments).

Finally, the test set was a three factor nine level design

(729 experiments) collected at equal steps between the

maximum and minimum concentrations specified for the

training data.

E. Richards et al. / Sensors and Actuators B 88 (2003) 149–154 151

2.6. Data modelling and pre-treatment

Calibration modelling was performed on a PC running

Windows 2000 (Microsoft, USA) operating system with an

Athlon 1600XP processor (Advanced Micro Devices,

USA) and 1 GB RAM. Calibration models were produced

using in-house programs written using Matlab 6.1 (The

MathWorks, USA) and the Neural Networks Toolbox 4.0.1

(The MathWorks, USA). Non-linear neural network mod-

els were used rather than linear modelling methods such as

PCR or PLS because the voltammogram response for a

mixture of the analytes is not a linear combination of the

individual analyte voltammograms. This is due to electro-

chemical deposition of the analytes onto the working

electrode surface at different potentials, so analytes depos-

ited earlier in the potential scan partially interfere with the

detection of the subsequent analytes. In the previous paper

[2] two methods for determining the optimal neural net-

work settings were compared. The first used a fully

factored parameter search approach where the number

of epochs in steps of 100, the number of principal com-

ponents input to the network, and the number of hidden

neurons was evaluated to determine the network para-

meters that produced the most accurate calibration model.

The second method employed a genetic algorithm para-

meter search to explore the same three parameters, to

evolve the best neural network. In this paper, only the

genetic algorithm neural network parameter optimisation

method was applied, as this previous proved to be the most

efficient optimisation method. Data pre-treatment (range

scaling and PCA) was performed prior to modelling as

described in Richards et al. [2]. Although seven principal

components were found to explain almost 100% of train-

ing data variance, the number of principal components

retained was 22. This was to allow for non-linearities in the

data set which had been relegated to later principal com-

ponents by the PCA algorithm, to be included for training

the networks [2]. Network performance was quantified by

using the RMS distance error, which gives an indication of

the network’s overall calibration capability. The RMS

distance (RMSD) is calculated by taking the RMS error

for the replicate training data sets (RMSR) and the RMS

error for the interpolation sets (RMSI) and combining

them as in Eq. (1).

RMSD ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiRMSR2 þ RMSI2

p(1)

This distance equation applies only to the validation and test

sets, as they were comprised of replicate training data sets

and also interpolation data sets collected at concentrations

not used for training.

3. Results and discussion

3.1. The automated liquid handing system

The automated liquid handing system proved to have

good repeatability and reliability. Fig. 3 shows six repeti-

tions of blank subtracted voltammograms at five concentra-

tions for ethanol. The data shown was produced in a random

order amongst the 6 repetitions of 3 factor 5 level fully

factored experimental design which formed the training

data.

3.2. Neural network calibration model

The results presented in this paper used the same genetic

algorithm parameter search technique as previously

empoyed [2]. The results for ethanol are used as a case

study with summarised results in Table 1 for fructose and

glucose.

Fig. 3. Blank subtracted voltammograms for five different concentrations

of ethanol. Data is taken from the training data sets.

Table 1

Best neural network calibration models obtained using validation data

Analyte Property

RMS replicate error RMS interpolation error RMS distance error Number of epochs Number of PCs Number of

hidden neurons

Ethanol 0.0248 0.0255 0.0355 472 21 9

Fructose 0.0257 0.0288 0.0386 553 21 8

Glucose 0.0231 0.0247 0.0338 357 14 8

152 E. Richards et al. / Sensors and Actuators B 88 (2003) 149–154

3.2.1. The number of training epochs

The relationship between the network accuracy and the

number of training epochs follows the same trend as pre-

viously seen in Richards et al. [2]. The effect of the training

for a given number of epochs was undermined by having

random initial weights. Only where there were short training

periods (i.e. less than 50 epochs) were there signs of under-

training. Good networks were trained within all epoch

ranges, with the best network using 472 epochs.

3.2.2. Number of PCs

In general, there was a reduction in the RMS distance

error with increasing numbers of principal components

presented to the neural networks for training, suggesting

that there was non-linear information relegated to the higher

principal components which was salient information for

modelling [2,4]. The best general network was trained using

21 principal components.

3.2.3. Number of hidden neurons

Fig. 4a shows that as the number of hidden neurons are

increased from 2 to 5 neurons there is a general improvement

in the RMS distance error. As the hidden neurons are

increased further there is generally no improvement, how-

ever, there are individual networks that are trained that are

very good general networks. The best network was trained

using nine neurons suggesting that there were many intricate

relationships that required modelling.

Fig. 4b shows how many times the genetic algorithm

chose numbers of hidden neurons to model ethanol from the

data. It can be seen that it concentrated mainly on 8 and 9

hidden neurons to produce the best networks. There may

have been many complex relationships to model or it may

Fig. 4. (a) RMS distance errors for ethanol, plotted as a function of the

number of hidden neurons. The best (*), the worst (*) and average RMS

distance errors (–) are shown. (b) Frequency the genetic algorithm chose a

number of hidden neurons to train with, for ethanol calibration.

Fig. 5. (a) Box and whisker plot showing the results obtained from the best

network when presented with the ethanol validation data. The whiskers

denote the maximum and minimum values, the boxes denote the upper and

lower quartile ranges and the line in the box denotes the median value. (b)

Box and whisker plot showing the predicted values for the best network

when presented with the test data set. The whiskers denote the maximum

and minimum values, the boxes denote the upper and lower quartile ranges

and the line in the box denotes the median value.

E. Richards et al. / Sensors and Actuators B 88 (2003) 149–154 153

have been the case that there were local areas of under fitting

that were compensated for by extra degrees of freedom

added by having more neurons [5].

3.3. Testing the best network

In the previous paper [2] the neural network models were

optimised using a validation data set, yet were not presented

with a test data set afterwards. For this piece of work, the best

model found using the validation data in the optimisation

process underwent a final test using a comprehensive three

factor nine level (729 experiments) test set. The predictive

results for ethanol from the test data can be seen displayed as a

box and whisker plot in Fig. 5b along with the results for the

validation data used to obtain the best network in Fig. 5a. At

each concentration shown in Fig. 5b there are 81 predictions.

The results for the test data set are summarised in Table 2. The

tails of the whiskers denote the maximum and minimum

predictions, the top and bottom of the box denote the inter-

quartile ranges for the predictions and the bar across the

middle of the box denotes the mean value.

4. Conclusions

The automated liquid handing system provides a fast,

reliable and accurate method of performing large data

collection activities in a random order. It was able to

improve accuracy and reproducibility with respect to

similar experiments done by hand. This was shown by

performing a previously conducted calibration method

using data collected by the automated liquid handing

system. In general, the neural network models produced

from the data collected using the automated liquid handing

system were better than those produced using manually

collected data [2]. Using the automated liquid handing

system also reduces waste and analyte reagent usage by

taking measurements from a small sample size of 1 ml.

The best model trained also provided good predictions for

a large test set incorporating many samples taken at

concentration permutations previously unused in the mod-

elling process.

In future work, we plan to produce data for the calibration

of four analyte systems. We will also be interfacing the

automated liquid handing system to alternative analytical

methods, such as optical spectrometers and gas sensors to

provide multi-analyte calibration models for different types

of analytes.

References

[1] C. Bessant, S. Saini, Simultaneous determination of ethanol, fructose

and glucose at an unmodified platinum electrode using artificial neural

networks, Anal. Chem. 71 (1999) 2806–2813.

[2] E. Richards, C. Bessant, S. Saini, Optimisation of a Neural Network

Model for calibration of voltammetric data, Chemomet. Intel. Lab.

Sys. 61 (2002) 35–49.

[3] Y. Fung, S. Mo, Application of dual pulse staircase voltammetry for

simultaneous determination of glucose and fructose, Electroanalysis 7

(1995) 160–165.

[4] F. Despagne, D.L. Massart, Neural networks in multivariate calibra-

tion, The Analyst 123 (1998) 157R–178R.

[5] S. Lawrence, C. Giles, Overfitting and neural networks: conjugate

gradient and backpropagation, in: Proceedings of the IEEE Interna-

tional Joint Conference on Neural Networks, Como, Italy, 24–27 July

2000, pp. 114–119.

Table 2

Test data set results for the best neural network calibration models

obtained using validation data

Analyte Property

RMS replicate

error

RMS interpolation

error

RMS distance

error

Ethanol 0.0459 0.0253 0.0524

Fructose 0.0304 0.0273 0.0409

Glucose 0.0306 0.0386 0.0492

154 E. Richards et al. / Sensors and Actuators B 88 (2003) 149–154