WO2000032091A2

WO2000032091A2 - Diagnosis of gastric and lung disorders

Info

Publication number: WO2000032091A2
Application number: PCT/GB1999/003981
Authority: WO
Inventors: Antony Peter Francis Turner; Alexandros K. Pavlou; Hugh Barr
Original assignee: Cranfield University; The Gloucestershire Royal Hospital Nhs Trust
Priority date: 1998-11-27
Filing date: 1999-11-29
Publication date: 2000-06-08
Also published as: GB2361872B; GB2361872A; GB0115844D0; AU1289300A; WO2000032091A3; GB9825904D0

Abstract

For diagnosis or monitoring of gastric or lung conditions, particularly H.pylori infection or tuberculosis, a gas sample is passed to a multiplicity of chemical sensors which generate electrical outputs. These outputs are passed to a data processing system, preferably a hybrid intelligent system employing a search optimisation engine of genetic algorithms and a multiplicity of ne ural networks. This determines distinctive patterns characteristic of particular disease states.

Description

DIAGNOSIS OF GASTRIC AND LUNG DISORDERS

Technical Field

This invention relates to a system for use in diagnosing and/or monitoring gastric and lung disorders by analysing gas samples. These may be sampled from patient directly (e.g. gas sample from stomach or breath) or generated from a sample affected by a said disorder (e.g. sputum) e.g. via enzyme treatment. H. pylori (HP) , infection is known as the most common gastrointestinal bacterial disease world-wide. It is now accepted as the major cause of gastroduodenal ulceration in over 80-90% of patients, and chronic atrophic (type B) gastritis (1) . It has been also recognised as type 1 human carcinogen and suggested as a co-factor in the development of gastric adenocarcinoma, and mucosa-associated lymphoid tissue (MALT) lymphoma (2) . Two patterns of HP infection have been recognised: a. in developing countries where a large proportion of children are infected and almost all adults in different age groups have a chronic HP infection, and b. in western developed countries, where the prevalence of infection increases from age 20 onwards (3) .

Current tests used to detect HP are either invasive (endoscopy, biopsy, histology, culture, rapid urease test, Polymerase Chain Reaction) or non-invasive (serology, ¹³C urea breath test) . All the above tests vary in their sensitivity and specificity and the choice of test will depend whether the aim is to detect infection or to test the success of eradication treatment. Unfortunately, despite the introduction of new serological and molecular biological techniques, contamination effects, high cost and the need for high skilled personnel can severely limit their diagnostic efficiency. Most of the above techniques are not able individually to: a. Characterise and discriminate between different gastric disease stages, focusing only on HP detection and not recognition of disease level, b. offer a complete diagnosis in the form of a rapid bedside technique (for ¹³C-urea breath test it generally takes more than 48hrs to deliver results due to the lack of a mass spectrometer in the endoscopy unit, and the need for additional personnel to perform the test) , c. offer an easy-to-use and inexpensive test (PCR and serology need additional skilled personnel, and contamination effects lead to the need to repeat tests, urea-labelled pill is expensive to produce, histology and culturinrequire highly skilled personnel) (4,5), d. facilitate modern pattern recognition methods and artificial intelligence to store, analyse and predict patient data, e. provide individually a single "gold standard" technique and to encompass several aspects of gastric disease. There is always a need to combine several diagnostic tests like the rapid urease test, endoscopy, culture, histology and PCR in order to detect infection and characterise a certain disease stage. However in most UK hospitals, endoscopy and culture remain the standard technique.

Tuberculosis (TB) is caused by Mycobacterium tuberculosis (MTB) and is a major public health problem in many countries world-wide with particular significance in developing countries. Approximately one third of the world population (1.9 billion) is infected with MTB. Globally about 10 million new cases of TB are detected every year and 3 million deaths occur annually due to the disease (6) . It is also estimated that 16 million people are currently infected with HIV and one third of them will eventually progress to Tuberculosis. Multidrug resistance, poverty, and AIDS have all contributed to global TB resurgence. Globally bronchoscopy and conventional microbiological methods such as culture and microscopy are still considered the most specific and sensitive diagnostic techniques. In most developing countries detection of TB is based exclusively on time consuming and skilled observation, under the microscope, of stained mycobacteria. New methods based upon nucleic acid amplification, such as PCR, strand displacement amplification (SDA) , and serology are considered a step forward, however they are laborious very expensive, and suffer from specimen contamination, and low sensitivity. The introduction of Gas Chromatographic techniques in the early 60' s prompted investigation of the diagnostic potential of breath or skin volatiles related to several disorders such as cancer, liver cirrhosis and certain biochemical human tissue states (lipid peroxidation) (7,8,9,10,11). Additionally workers in the former Soviet Union have also reported the existence of volatile fatty acids and other biomarkers in the exhaled breath of TB patients (12) , and very recently Wang C-H et al. have detected a significant amount of exhaled nitric oxide in active pulmonary TB (13) . Although all of the above methods are characterised as laborious, expensive and unable to give reproducible results due to lack of powerful artificial intelligence software or equipment inefficiency, they showed the existence of abnormal concentrations of endogenous volatile mixtures due to the onset of infection. The diagnostic power of odours has been recognised since the 4^th century BC. In recent years several workers have investigated the production of certain odours due to infection and destruction of human tissue (14, 15) . The first model of an artificial electronic odour detection system was described during the early 1980' s and attempted to mimic some functional characteristics of the human olfactory system. Since then, a significant amount of research has been undertaken to design new integrated sensor array systems and apply them, principally, in the food industry (16) . Moreover in the past two years scientists have reported novel applications in microbiology (17), diabetes diagnosis (18) , and intrapulmonary volatile pattern discrimination (19) . In vi tro experimental work completed in our facilities has revealed the existence of certain volatile patterns of MTB, HP and other gastroeosophageal pathogens. The invention can also be applied to analysis of gas samples generated in vitro from samples obtained from patents, e.g. sputum samples. Disclosure of Invention

In a first aspect the invention provides apparatus for use in diagnosing and/or monitoring gastric and/or lung disorders comprising

(a) a sampling system for collecting a gas sample generated by a patent or generated from a sample taken from the patient;

(b) an array of gas sensors each having a different pattern of sensitivities to potential components of the gas sample and being adapted to provide an electrical output signal in response to detection of one or more of said components; and means for passing gas from said gas sample to said array;

(c) a data processing system arranged to receive said electric output signals, said data processing system being adapted to analyse the output signals to detect patterns indicative of the presence of predetermined disorders and/or stages of predetermined disorders.

In a second aspect the invention provides a method of diagnosing and/or monitoring gastric and/or lung disorders comprising (a) collecting a gas sample generated by a patent or generated from a sample taken from the patient;

(b) passing gas from said gas sample to an array of gas sensors each having a different pattern of sensitivites to potential components of the gas sample and being adapted to provide an electrical output signal in response to detection of one or more of said components; and

(c) passing said output signals to a data processing system which analyses output signals to detect patterns indicative of the presence of predetermined disorders and/or stages of predeterminred disorders.

For generating the gas to be sample we may use a natural odour generation system. For gastric analysis this may involve giving a pill or a solution of non- radioactive and unlabelled urea (10 times cheaper than labelled alternatives and harmless) to the patient suitably 30 minutes before endoscopy after a period, e.g. 16 hrs, of fasting. For lung air analysis, a period, e.g. 16 hrs, of fasting may suffice, optionally with administration of harmless biochemical inducers able to concentrate in the lungs and metabolised by mycobacteria or cancer cells.

A preferred type of embodiment uses a sniffing endoscope comprising: (1) A bag sampling system to collect and equilibrate the intrapulmonary and intragastric volatiles (the latter optionally in the presence of gastric juice) , which will be connected directly to one endoscopic channel by using non-toxic, e.g. Teflon, tubing. After each gas sample or "sniff", clean (e.g. carbon filtered) air passes through the channel to avoid contamination from patient to patient. Volatile substances are transferred very rapidly into an inexpensive sampling bag (one bag per patient) by applying a suction pump (suitably 0.7 barr) to avoid losing any volatiles.

(2) An array of gas sensors that interact in a unique way with individual gas molecules or complex odour mixtures and transform physicochemical interaction to electronic signals captured by specific data acquisition software (electronic nose) . - 1 -

(3) Software and microprocessor for fast odour recognition: A hybrid intelligent system that controls automatically a search optimisation engine of genetic algorithms to identify the best sensor parameters and the most reliable configurations of a group of back propagation neural networks (NN) which finally learn volatile profiles and identify disease patterns.

Expert systems employ a type of hybrid information processing which is now replicated in a new generation of adaptive machines. At the heart of these adaptive machines are intelligent computing systems some of which are inspired by natural mechanisms . NNs can learn to recognise patterns by repeated exposure to many different examples. They are good at recognising complex patterns, from financial training to medical imaging. Genetic algorithms are also naturally inspired and based on the biological principle of "survival of the fittest". A genetic supervisor evolves a problem's solution over many generations, with each generation having better solution than its predecessor. Both techniques are good at explaining their decisions but they cannot automatically acquire the rules they use to make those decisions. These limitations favour the use of hybrid intelligent systems. Like Kobe Steel Plant Japan, which uses a hybrid of different intelligent techniques to solve sub-tasks of the problem (21) .

The "sniffing" endoscope offers the following advantages:

Fast and inexpensive collection of volatile samples, which are representative of the actual physical and biochemical status of the human lung and the stomach.

Recognition and discrimination of different disease stages in the stomach and the lung before and after treatment . Storage, and rapid analysis (10 min) of patient data in a bedside diagnostic system.

Considerable reduction in the cost of tests.

Some embodiments of the invention will now be described with reference to the accompanying drawings. Brief description of drawings

Fig 1 is a schematic drawing of a data processing system in which several parallel NNs optimised by a specific genetic algorithm can be trained, tested and run automatically by an expert intelligent system which will be able to apply certain rules extracted from experimental results and laboratory experience (20,21).

Fig 2 is a schematic view of a "sniffing" endoscope which chararacterises HP atmospheres in artificial stomach by using a hybrid intelligent model: (1A&1B) 2L artificial stomach, (2A&2B) bag sampling systemcollection of volatiles for odour analysis, (3) 2-way valve polypropylene stoppers, (4) fibre optic endoscope, (5) 3- way valve stopper, (6) gas sensor array and microprocessor unit, (7) activated carbon filter, (8) control sample, 15ml of RO water, (9) Vacuum pump, (10) Data capture software, (11) Hybrid intelligent model comprising genetic algorithm system 12, NN back propagation analysis system 14, and multivariate analysis system 16. Fig 3 : an actual sensor-response curve of a H.pylori-enriched media (HPE) headspace. Four parameters have been selected to study the sensor response: Absorption (Ab: maximum rate of change of resistance), Desorption (DS: maximum negative rate of change of resistance) , Divergence (DIV: maximum step response) , Area (AR: Area under the curve) .

Fig 4 : graphs showing Twenty-five sample responses "sniffs" characterised by 19 genetically selected sensor parameters. Three graphs can be seen; (N) sterile artificial stomach, (HPN) HP positive article stomach and interaction with certain natural biochemical inducers.

Fig 5: discriminant analysis scores and formation of three separate clusters. An artificial stomach atmosphere containing H. pylori and biochemical inducers (HPE) has produced a completely different odour profile. Although sterile artificial stomach (N) and H. pylori normal growth (HPN) are closer, still there is a clear distinction between them.

Fig 6; schematic diagram of a flow injection bubbling system that was applied for bacterial odour delivery and detection (WB: water bath, SP; sampling point, SU: sensory unit, CF: carbon filter, F: bio- filter, AF: air flow; CS: control sample.

Fig 7; graphical representation of DA scores between; (av) M. svi υm, (c) control, (p) P, aevroginosa, (tb) Tuberculosis and (sc) M. scrofulace .

Fig 8: graphical representation of thirty-eight sensor parameters showing a clear discrimination between; (m) M, aviυm and M, scrofυlaceυm_r (p) P. aeu.rogi-io.sa (tb) MTB and (c) control sterile cultures. Fig 9: graphical representation of GA-NN prediction confidence (λ «1) of 10 sputum samples;

Fig 10: graphical representation of non-linear nature and complexity of 46 sputum pulses and 5 groups of patterns;

Fig 11: graphical representation of DA-cv separation and correct classification;

Fig 12: graphical representation similar to Fig 11 for a different experiment. Modes for carrying out the invention

Example 1 "Sniffing" the static headspace of an artificial stomach infected with H.pylori 1.1 Odour generating system

Following successful isolation of HP from gastric biopsies and growth in tissue culture flasks (75cl

Corning) , containing 70ml brain heart infusion broth (Oxoid) , 5% serum bovine (Oxoid) and antibiotic supplement (Vancomycin lOmg IT¹, Trimethorpim lactate 5mg L^"1, Cefsulodin 5mg L^"1, Amphotericin B 5mg L^*1) (Oxoid) , three separate treatments were prepared, adjusted to 10⁷ cells ml^-1 in media of pH 7.3 and inoculated into 2L urine drainage bags (inflated with carbon filtered air) (Simpla) , each containing an anaerocult C sachet (Merck) to create a microaerophillic atmosphere which constitutes the in vivo HP microbiotic environment and favours its metabolic activation: (I) H. pylori (HPE) in 80ml of medium containing BHI-5% serum bovine, 5ml of 15% sterile urea solution (Oxoid), 0.75mg ml^"1 L-asparagine, and L- glutamine (Sigma). (II) H. pylori (HPN), normal growth in 80ml BHI-5% serum bovine with no additives. (Ill) Sterile medium (N) containing 80 ml BHI-5% serum bovine plus urea, glutamine and asparagine as described above. Replicate samples of each treatment were incubated for 100 minutes at 37^° C and then placed randomly, one each time, in a 6L plastic container (Gio Style) (1A, IB: artificial stomach and 2A, 2B: Bag sampling system figure

2) .

1.2 Odour delivery system

Following 100 min of incubation each bag was placed in a 6L plastic container and connected via a 2-way valve stopcock (BDH) and Teflon tubing (Tygon) to the anterior end of a fibre optic endoscope (Olympus) . The other endoscope' s channel end was connected via Teflon tubing and a 3-way stopcock to a bag sampling system (plastic container & 2L sampling bag) which was also connected to a vacuum pump

(10 L min^"1, Patterson) (Figure 2) . The volatile collection system is based on a continuous open airway channel between the artificial stomach (1A &1B figure 2), the endoscope biopsy channel (4 figure 2) the sampling bag (2B figure 2) and a vacuum pump (9 figure 2) that is capable of rapid transfer (30 sec) of the intragastric atmosphere to the electronic nose apparatus, for direct odour analysis (6,7,8 figure 2). 1.3 Odour detection system

An electronic nose (Bloodhound Sensors, Leeds, UK) , which employed 12 conducting polymer sensors was used. Specific selection and tailoring of polymers, doping materials and precise manufacturing process can make each gas sensor consistently responsive to different volatile groups. Physicochemical interaction between the volatiles and the conducting polymer surface produces a change in resistance, which can be amplified and analysed through a data capture software (Bloodhound Sensors, 10 figure 2) . The sensory unit employed a control sample container (8, figure 2) that produces two calibration reference points, a baseline and a control sample. Activated carbon- filtered air is passed over the sensor surface and generates the baseline (flow: 4ml min^"1) (7, figure 2) . The control sample unit contained 15ml of sterile water and was used to confirm that the reference point was not affected by drift. A specific sampling profile used 4 seconds of absorption time and 12 seconds of desorption time.

1.4 Odour recognition system and data analysis

Figure 3 describes a real-time sensory response curve taken from a H. pylori volatile headspace. Twelve sensor responses and 4 parameters created a set of 48 normalised input variables. Twenty-five samples were collected: a. 8 from sterile atmosphere (N) , 8 from HP (normal growth, HPN) atmosphere and 9 samples from HP (enriched media, HPE) atmosphere of enhanced volatility (Figure 4) . The previous data was divided into two groups randomly: a. Training data (5 samples N, 5 samples HPN, 5 samples HPE) 60% of all data, b. Test data "unknown" (3 samples N, 3 samples HPN, 4 samples HPE) 40% of all data. The latter was kept out of Neural Network (NN) training. A novel hybrid intelligent system of Genetic Algorithms- NNs (Neuralyst, Brain Maker, USA) and multivariate techniques (XlStat, France) was employed. Genetic training uses a special type of optimisation technology, which consists of models of a selective evolutionary process. This used an evolutionary combination of all the three strings and an addition of a mutation rate, which eventually produces a "phenotype", a new more evolved NN architecture. Evolution towards the most successful NN configuration is processed through successor generations. Each NN structure in the generation is evaluated by the lowest error achieved after a certain number of epochs. The input parameter set is represented by an inclusion rate of 0.77. Furthermore a population size of 4 was set to determine the number of phenotypes evolved in each generation. Immigration pool mode was also used to replace the weakest phenotypes. A set of 1 for cross breeding determines the frequency of intermingling of features on the same string. Finally a mutation rate of 0 . 6 was applied to evolve new NN structures.

A genetic algorithm-NN back-propagation employing a specific architecture of 19 input neurones (sensor parameters), a learning rate of 0.94878 and a momentum of 0.354654 after 10 generations of neural evolution, achieved a 93% prediction rate. Nine out of 10 "unknown" samples were identified correctly and only HPN3 was confused with (N) sterile artificial stomach (Table 1) . A subset of the genetically input sensor parameters was used to perform Discriminant analysis (DA) . DA identified a set of sensor parameters that best discriminated between the three tested classes. For maximum discrimination the following two conditions had to be satisfied: a. the distance between the bacterial clusters should be as far as possible and b. distances within each bacterial cluster should be as close as possible. Eventually DA identifies a new axis Z such that a new variable from rough sensor data could provide the maximum discrimination. The above multivariate linear technique showed a complete distinction between HPE and the other two classes. Additionally both HPN and N samples, although being very close, have formed two distinctive clusters (Figure 5) .

Example 2. Sniffing" Tuberculosis in vi tro .

2.1 Clinical isolates

All bacteria were isolated from patients at the chest unit and assigned a Hospital number: Mycobacteri um tubercul osis (MTB) (RIVM myc 4514), M. avi um (RIVM myc

3875), M. scrofulaceum (RIVM myc 3442) and Pseudomonas aeuroginosa (AMC 23123) .

2.2 Cultural system/Odour-generating model

The above bacterial isolates were cultured in tissue flasks (75cl Corning) containing Tween albumin medium

(Oxoid) to a final volume of 70ml until they reached their stationary growth phase (4 weeks for MTB, 2 weeks for the other two mycobacterial (Myc) species and 36 hrs for

P. aeuroginosa (P) at the same time. Conventional diagnostic microbiology and optical density measurements confirmed satisfactory growth of each species. The vent-caps of the bottles were sealed with paraffin film (Nesco) to concentrate the metabolic production of bacterial volatiles

(7 days for Myc and 12 hrs P) . A number of sterile Tween- albumin samples (C) were also prepared and cultured for 4 weeks under the same conditions with the pathogenic cultures .

2.3 Odour delivery (Flow injection-bubbling system) and detection (Fig 6) When bacterial cultures reached their stationary growth phase, they were transferred, together with some control replicates in a 37^°C water bath and left there for 15 min to equilibrate. Each flask was connected with a specifically designed air-filtered sparging system, which consisted of Teflon tubing, a hydrophobic bio-filter (0.45μm, PTFE Whatman Hepavent) to reduce humidity over the sensor surface and an activated carbon filter to ensure clean air above the bacterial headspace. The sampling point was adjusted to a set height above the culture. A flow rate of 200ml^"1 was set automatically by the sensory unit (Figure 6) . Additionally, environmental conditions, at the sampling point, in the water bath and in the laboratory were monitored continuously in order to establish a standardised sampling regime. The volatile sensing system was the same as in example 1, except that a tissue culture flask (75cl Corning) containing 70ml of sterile water was used as control sample to prevent a drift at the reference point and a new sampling profile was used of 6 sec of absorption and 14 sec of desorption time. 2.4 Odour recognition and data analysis

Five parameters (AR: area, DV: divergence, DS : desorption, AB: absorption and Ratio: AB/DS) and 14 conductive polymer sensors extracted 70 sensor parameters, that carried 38 input/train-normalise signals and 15 "unknown" randomly selected "sniffs". Data processing employed a hybrid intelligent model of genetic training, optimisation of back propagation NNs and multivariate techniques. Genetic inclusion rate of 0.88, a mutation rate of 0.6 and a cross breeding of 1 were used together with an immigration pool mode to evolve new NN phenotypes.

After 100 generations of neural evolution, a 3-layer back propagation NN (63-26-5) carrying a learning rate of 0.414, a momentum of 0.9262, an input noise of 0.0365 and a testing tolerance of 0.5 achieved a NN prediction rate of 96%. Fourteen out of 15 samples were identified correctly

(Table 2) . Two subsets of genetically selected input sensor data were used to perform DA. The first subset of normalised data (48 sensor parameters) demonstrated a clear distinction between all bacterial classes (Figure 7) . Additionally the second data set (38 sensor parameters) managed to discriminate between (c) : control, (tb) : MTB,

(m) : M. avium & M. scrofulaceum and (p) : P. aeuroginosa

(Figure 8) .

Example 3. TB-Sense: Human sputum enzymatic treatment enables Mycobacterial species volatile pattern recognition. 3.1 Introduction

Analysis of sputum may provide valuable clinical and physiological information.

The identification of Mycobacteri um sp. in sputum is laborious and time-consuming.

The initial investigation requires staining and microscopic examination of sputum.

It takes a minimum of 30min preparation of each sputum sample followed by microscopic examination. An experienced laboratory technician would normally expect to examine no more than 6-10 sputum samples per hour. Furthermore, sputum microscopy is positive in a minority of cases and only when there is a heavy bacterial load and the technique can only identify the presence of Mycobacteri um as a group and cannot differentiate between different strains. This can only be achieved by culture of sputum, which takes a minimum of two weeks and usually 4-6 weeks to correctly identify the sub-species and provide appropriate antibiotic sensitivities. Recently, the rapid molecular examination of species utilising PCR-based technology has become available, however this is technically demanding, requires significant laboratory resources and clean facilities and is expensive (unit test cost approx. £60) .

We are currently investigating the hypothetical existence of 2 non-linear dynamic (chaotic) systems: "a" complex metabolic changes during active Tuberculosis (TB) and pulmonary infection and their expression as volatile pulses generated after enzymatic treatment over the headspace of fresh human sputum and "b" the interaction of chemosensory surfaces and sputum volatile groups and their complex pattern recognition by using a hybrid intelligent system. To understand the complexity of in vivo TB-sputum patterns and according to our hypothesis we developed a model of four interdependent factors (Example 1, 2) . However in this example the odour generating mechanism consisted of an enzymatic cocktail treatment of sputum which generated within 5 hrs some unique and species- specific headspace volatile patterns.

Two are the enzymatic target groups here, the complex mycobacterial lipid-cell wall and necrotic tissue substances and other TB metabolic products present in sputum. Certain enzymes like upases can interact with complex biological substances such as long chain fatty acids and create novel flavours and volatile compounds. There have been several applications of flavour enzymatic generation in food technology. However in this study we introduced the idea of enzymatic cocktails instead of single enzyme treatments. The introduced novel diagnostic test introduces a unique biochemical "dialogue" with respiratory pathogens and Mycobacteria and TB itself de profundi s . It also forces the respiratory infection and TB to reveal their active metabolic pulses and express them as non-linear complex patterns generated on the surface of an array of 14 conducting polymer gas sensors. 3.21 Materials & Methods

3.2.1 Sputum volatile generation.

Forty-six 5ml sputum samples infected with M. tuberculosis (tb:10 samples), M. avium (av: 10 samples), Pseudomonas aeruginosa (p: 10 samples), mixed infection (tb- p, m: 8 samples) and normal/control sputum (c: 8 samples) have been collected from patients attending the Chest Unit clinic of Amsterdam Medical College and Royal Tropical Institute. From each patient sample 1 ml of sputum was mixed with 2ml of an enzymatic cocktail (in Phosphate Buffer Saline 7.7 pH) containing lmg of porcine pancreas lipase (Sigma) and 4mg of Aspergill us niger lipase (Sigma) . Forty-six sputum treatments were incubated for 6 hrs at room temperature (23°C) . 3.2.2,3 Volatile delivery and detection systems were the same as presented in example 2 except that a new sampling profile of 7sec of absorption time and 21sec of desorption time was set.

3.2.4. Intelligent pattern/pulse recognition. Fifty-six sensor parameters and 46 sputum pulses generated a matrix of 2576 sensor data items. Two data groups have been selected randomly: "a" training data (tb: 8, av: 8, m:6, c: 6, p: 8) 78% of all data, "b" test data of 10 "unknowns" (c: 2, tb: 2, m: 2, p: 2, av: 2) 22% of all data. The latter was kept out of Neural

Network (NN) training. A hybrid intelligent model of genetic algorithms- neural networks (GA-NNs) and multivariate techniques (discriminant analysis-cross validation: DA-cv) . The Genetic supervisor used an evolutionary combination of an inclusion rate of 0.93, a population size of 5 (number of NN phenotypes evolved per generation) , an immigration pool mode (to replace the weakest NNs in each generation) , a set of 3 cross- breedings (frequency of intermingling of NN features in the same phenotype) and a 0.743 mutation rate. After 10 generations of evolutionary training, the Genetic Supervisor selected a 4-layer (51 input-13-21 hidden-5 output) back-propagation NN which employed a sigmoid function, an adaptive learning rate, a momentum of 0.174 and achieved a prediction rate of 96% (Table 3, below) .

Nine out of 10 samples were identified correctly and only one normal sputum sample (c2) was not classified correctly. However its real output (0.48) was very close to preset test tolerance limit of 0.5 (Fig. 9) By extracting the 51 "genetically" selected sensor parameters, 4 groups of complex non-linear patterns have emerged (Fig 10) . DA-cv also used that set of "genetically" selected parameters and managed to discriminate between: 1. (av, c, m, p, tb: 2 "unknowns" were recognised correctly by cross validation, Fig 11) and 2. (c,b: Pseudomonas patterns as bacterial, mb : for Mycobacteria av-tb and m for mixed infection tb-p, where another two "unknowns" were classified correctly, Fig 12) .

Table 1 : Adual output performance and architecture of a hybrid genetic algorithm- optimised back propagation NN in discrimination between H.pylori in enriched media (HPE), sterile artificial stomach (N) and H.pylori normal growth (HPN). A corred identification of nine out of ten "unknown" samples has been achieved (1 for true and 0 for false).

NEG HPN HPE MF

N1 1.016624 0.004214 0.001396 TRAIN N1 NEURAL NETWORK STATISTICS

N2 1.032813 -0.03177 0.00032 TRAIN N2 0.246791 RMS Error

N3 1.026024 -0.02479 -0.00436 TRAIN N3 30 Number of Data Items

N6 1.026494 -0.0262 -0.00166 TRAIN N5 28 Number Right

!N7 1.017941 0.000767 -0.00022 TRAIN N7 2 Number Wrong

HPN1 0.00173 0.998638 -0.00237 TRAIN HPN1 93% Percent Right

HPN2 -0.01606 1.010791 -0.00018 TRAIN HPN2 7% Percent Wrong

HPN4 -0.0347B 1.037109 45.00079 TRAIN HPN4 4903 Training Epochs

HPNS -0.01002 1.006666 -0.00042 TRAIN HPN5

HPN7 -0.01932 1.014484 -O.00045 TRAIN HPN7 Network Parameters

HPE1 -0.01123 -0.01667 0.997363 TRAIN HPE1 0.94878 Learning rate

HPE2 0.000723 -0.01448 1.006091 TRAIN HPE2 0.364654 Momentum

HPE4 0.003376 -0.02288 1.007233 TRAIN HPE4 0.079867 Input Noise

HPEG 0.00116 -0.01247 1.006293 TRAIN HPE6 0.03 Training Tolerance

HPE8 0.00022 -0.01908 0.9884 TRAIN HPE8 0.6 Testing Tolerance jN4 1.018076 -0.01841 0.002704 TEST N4 Genetic Training Statistics

IN6 1.036639 -0.03796 0.006769 TEST N6 10 Generation Count

N8 1.043061 -0.04376 0.007101 TEST N8 4 Structure Count

HPN3 1.046207 -0.04601 0.002032 TEST HPN3 0.023432 Least RMS Error

HPN6 -0.00646 1.000386 -0.00045 TEST HPN6 6000 Least Epochs

HPN8 0.086268 0.928009 0.009686 TEST HPN8 NETWORK ARCHITECTURE

HPE3 -0.01006 -0.01411 0.994107 TEST HPE3 # Layers 3

HPE5 -0.00163 -0.01213 1.000989 TEST HPE6 # Neurons per Layer

HPE7 0.00069 •0.00676 0.966876 TEST HPE7 19 INPUT

HPE9 0.000926 0.019791 0.999277 TEST HPE9 16 HIDDEN

3 OUTPUT

Table 2: Real output performance of a hybrid genetic algorithm optimized back propagation NN in discri ination between headspace atmospheres created by the following clinical Isolates: P.aeurogmosa (pε), M. tuberculosis (tb), M.avlum (»V), M.scrofulaceυm (SC) and control (no growth) (C). A 96% preldction rate has been achieved and 14 out 15 "unknown" samples have been identified correctly.

PSEUDO TB AVIUM MSCROF CONTR MF

1.001727295 0.001898193 0.000119019 0.000253296 -0.00266724 TRAIN _Ps1

1.003372192 0.000991621 0.004046631 -0.004245 -0.00115662 TRAIN ps2

1.004547119 -0.00635986 0.005692944 -0.00082092 -0.00115662 TRAIN ps4

0.993569946 -0.00145874 0.007772827 -0.00840759 -0.00333862 TRAIN ps15

0.999142456 -0.00508423 0.000421143 0.001428223 -0.00135803 TRAIN ps17

-0.00672913 1.004278564 -0.01008606 -0.01411438 -0.0169342 TRAIN t 1

-0.0092804 0.998303223 0.009619141 -0.01089172 0.003677368 TRAIN lb 3

-0.00760193 0.99541626 0.011364746 -0.0085083 0.003375244 TRAIN tb5

-0.01011963 0.997698975 0.000354004 -0.0144165 -0.00018311 TRAIN tb7

-0.00531921 1.00045166 -0.00196228 -0.01320801 0.008276367 TRAIN tb10

-0.01089172 1.005487061 -0.00206299 -0.02002258 -0.00975037 TRAIN tb14

0.000756836 1.010388184 -0.00323792 -0.01508789 -0.00236511 TRAIN tb16

0.000588989 0.987963867 0.016030884 -0.00948181 0.009988403 TRAIN tb17

-0.00558777 0.996994019 0.002502441 -0.00817261 -0.00803833 TRAIN tb18

-0.05 -0.00847473 1.013543701 -0.00313721 -0.01522217 TRAIN av3

-0.05 -0.00354004 1.010791016 -0.00501709 0.003643799 TRAIN av5

0.007269287 0.002368164 0.998806763 -0.02583008 -0.00518494 TRAIN av7

-0.00404358 0.001159668 0.998739624 0.000253296 -0.00528564 TRAIN βv 1

-0.00055237 -0.00172729 1.009884644 -0.00834045 -0.00333862 TRAIN av13

-0.01220093 -0.00820618 1.008172607 0.005255127 - -00..0000333300550055 T TRRAAIINN av15

-0.00209656 -0.01656494 1.019082642 -0.02344666 - -00..0000886677661155 T TRRAAIINN av19

-0.0067627 -0.0101532 1.013577271 -0.01149597 0.000253296 TRAIN av21

-0.00058594 -0.01784058 1.013577271 -0.01988831 0.01321106 TRAIN av23

0.004180908 -0.00337219 -0.00018311 0.990515137 -0.00155945 TRAIN sc1

^' -0.02163391 -0.00921326 0.011364746 0.995248413 -0.00860901 TRAIN sc3

8.54492E-05 -0.00978394 -0.00256653 0.998370361 -0.00515137 TRAIN sc5

0.000756836 -0.01505432 0.000387573 0.998269653 0.004986572 TRAIN sc7

-0.00082092 -0.01025391 1.83105E-05 1.000854492 -0.00484924 TRAIN SC11

0.002401733 0.001092529 -0.00085449 1.001794434 -0.00562134 TRAIN sc13

-0.00625916 -0.00058594 0.O05758667 0.99944458 -0.00189514 TRAIN sc15

-0.00350647 -0.00924683 0.003408613 0.997900391 0.003710936 TRAIN sc17

-0.0076355 -0.00713196 0.010055542 0.991320801 -0.00753479 TRAIN SC23

-0.00797119 -0.00783691 -0.00162659 -0.05 0.996524048 TRAIN c1

-0.01841125 -0.00397644 -0.00072021 -0.0118988 0.897061157 TRAIN c3

-0.0118988 -0.00018311 -0.00508423 -0.00273438 0.992764282 TRAIN c7

0.003308105 -0.0021637 0.008578491 0.010357666 1.008877563 TRAIN c9

-0.01048889 -0.00588989 0.000387573 -0.01079102 1.001391602 TRAIN cU

0.007000732 -0.00746765 1.83105E-05 0.004684448 1.004782104 TRAIN c17

0.β71572β7β 0.001629639 0.001126099 -0.01008606 0.250881958 TEST p«6

0.176089478 -0.03177185 0.141009521 0.011129761 -0.02113037 TEST p«13

0.987191772 0.003677368 0.0015625 0.005892944 0 0..0000550088772288 TTEESSTT ps19

-0.02324524 0.791174316 0.032345581 -0.01498718 --00..0011555555778866 TTEESSTT tbβ

-0.0026001 0.92941894S -0.02039185 -0.02455444 --00..0000447744885544 TTEESSTT tb12

0.052755737 0.974182129 -0.01347656 0.06625061 00..009999440044990077 TTEESSTT tb19

-0.05 0.092736816 0.583822632 0.319229126 0.010726929 TEST • v1

0.008108521 -0.00921326 0.977590942 -0.01807556 -0.00535278 TEST «v9

•0.0046814 -0.03761292 0.791543579 0.130938721 - -00..0011222266880077 TTEESSTT «v17

-0.00354004 0.049969482 0.011297607 0.939053345 --00..0011225577001199 TTEESSTT sc9

0.065142822 -0.01119385 0.021032715 0.89887085 --00..0000004455116666 TTEESSTT •de

0.009619141 0.018045044 -0.01324158 0.934957886 --00..0000226666772244 TTEESSTT • c21

-0.04812012 0.049197388 0.188577271 -0.05 0.921563721 TEST cS

-0.03633728 -0.01344299 -0.00776978 -0.00961609 0.S74099731 TEST c11

-0.02149963 -0.02512512 -0.00266724 0.132315063 1.014147949 TEST CIS Table 2 cont .

HYBRID GA-NN STATISTICS

Network Run Statistic* 0.171020064 RMS Error

76 Number of Data Items 72 Number Right 3 Number Wrong 96% Percent Right 4% Percent Wrong 24563 Training Epochs

Network Parameter*

0.414 Learning rate 0.9262 Momentum 0.0365 Input Noise

0.02 Training Tolerance 0.5 Testing Tolerance 1 Epochs per Update 0 Epoch Limft 0 Time Limit (Hrs) 0 Error Limit (Increase)

Genetic Training Statistics

100 Generation Count 3 Structure Count 0.00429669 Least RM S Error 0 Least Epochs NETWORK ARCHITECTURE

# Layers 3

# Neuron* per Layer

63 INPUT 26 HIDDEN 6 OUTPUT

CAPTURED DATA: 53 SAMPLES TRAIN DATA: 38 samples.72% "UNKNOWN": 15 samples, 28% for each species: P.aeurogmosa : 5 train, 62.5%

3 test, 37.5% M. tuberculosis: 9 train, 75%

3 test, 25% M.a um : 9 train, 75%.

3 test.25 % M.scrofulaceum : 9 train, 75%

3 test, 25% Control: 6 train, 66%

3 test, 33% Table 3

000958557 -00476501 -00411713 102965698 -005 tram TB10 Genetic Training Parameters

-00162628 -0008609 -00444946 098M8499 -005 tram TB 7 0935 Inclusion Rate

-00097504 -00416748 00014954 -00147858 LO 14685059 tram Ml 4 Max Layers

-00017273 -00390228 00087799 001196899 0988769531 tram M2 30 L2 Neuron Umit

-00038757 -00381165 -00193512 000945129 099944458 tram M3 25 L3 Neuron Lmit

000317383 -00383514 -00352295 00052887 L004446 11 tram M5 02 Mm Learnng Rate

002217407 -00365387 -00280792 -005 099558 106 tram M8 02 Max Momentum

-00361023 -00434204 00068329 -00469116 L031167603 tram M9 003 Max Input Noise

0.831497 -00010559 -00338531 -00256622 0036776733 test AV7 5 Population Size

0.999545 02789795 00997864 -00499329 -005 teit AV10 Immigrate Population Mode

0.072662 04834875 01062042 017498779 -004771729 teit C2 3 Crossovers

009753723 0.596948 00615845 -00469452 -003472595 tes t C5 0743 Mutation Rate

046864624 -00497986 0.5 7 -00338867 -004684448 test PS2 I Neurons per Layer

-00466095 0 K10827 1.04473 -00439575 -005 test PS7 51

012835388 -00485229 01013977 0.533771 -005 test TB2 13

•00043457 00256317 -O0IM96 0.959363 -002304382 test TB4 21

0.001328 -0.04258 -0.03355 -0.02956 1.0343903 teit M4 5

-0017572 -00336517 0062793 007638855 0.91203 test M6

Table 3 cont

Table 3 GA-NN statistics pA avRim Control Pseudo TB Mixed mf

L03979492 -00489594 -00078369 -00459717 -004879 L5 tram AVI

098440552 -00042114 00035767 -00238159 -002539368 tram AV2 Neural

099524841 00106262 -00411377 -00365051 -001955261 tram AV3 Network Run Statistics

099605408 00129425 -00448975 -00152222 -0000183 II tram AV5 0.2187962 RMS Error

096879578 O0K15591 -00440247 -00259979 0016534424 tram AV6 50 Number oT Data kerns

098799744 00121704 -00380493 -0024353 0020965576 tram AV8 48 Number Right

L02492371 -0033316 00155273 -00497986 -005 tram AV9 2 Number Wrong

L00538635 00322449 00041138 -00498993 -005 tram V4 96% Percent Right

004382629 09317688 -00414398 003687744 -003096619 tram Cl 4 % Percent Wrong

-00470795 LOI26038 -00477173 -00471802 -003506165 tram C3 1440 Training Epochs

-00349945 10058563 00109619 -00380157 -004808655 tram C4

-0002533 09942078 -00008545 -00377808 -004798584 tram C6 Network Parameters

-00103546 09702057 -00453003 001455383 -004939575 tram C8 0506918546 Learning rale

001965637 09922607 00107941 000804138 0014956665 tram C9 0174077578 Momentum

-00007202 -005 L007434I -00254272 -000098877 tram PS1 0002990204 Input Noise

•0002063 -005 0990918 -00185455 0018817139 tra PS3 003 Training Tolerance

•00016266 -001700D 09928314 -00202576 -004986572 tram PS4 05 Testing Tolerance

-00235809 -00459381 10462067 -0047818 -005 tram PS5 1 Epochs per Update

-00353638 -00439575 L0484558 -00479858 -005 tram PS6 Enhanced Parameters

-00075684 -0006897 09938049 000636292 -004993286 train PS8 Sigmoid Function

0002771 00135468 L0441925 -00498322 -005 tram PS9 A aptive LR

-00217682 -00458038 09913879 -00028351 -005 tram PS10

-00025665 00333191 -00175049 09855D31 -003076477 tram TB 1 Genetic Training Statistics

-00019623 00286865 -00088776 L00078735 -003274536 tram TB3 K⁾ Generation Count

-00140472 001875 00250275 L01293945 -003841858 tram TB5 5 Structure Count

•0016095 00264038 00171051 100108948 -003942566 tram TB6 0019599055 Least RMS Error

-00069305 00340576 -00009552 09757782 -003123474 tram TB8 5000 Least Epochs

000847778 -00481873 -00383179 L02636719 -005 tram TB9 REFERENCES

1. Malfethciner, P. et al. Helicobacter pylori: an Atlas. Science Press, (1996), London.

2. Forman, D. et al. H.pylori and gastric cancer. Lancet, (1994), 343, 243-244.

3. Pounder, R.E. et al. The prevalence of H.pylori in different countries. Aliment Pharmacol., (1992).

4. Megraud, F. The most important diagnostic modalities for H.pylori now and in the future. Euro. J. of Gastr/gy and Hep/gy, (1997), 9(suppl), S13-S15.

5. Megraud, F. Advantages and disadvantages of current diagnostic tests for the detection of H.pyIori. Scand. J. of Gastro/gy, (1996), 215, 57-62.

6. Koclii, A. The global TB situation and the new control strategy of the World Health Organisation. Tubercle, (1990), 72,1-6.

7. Eriksen, S. Studying the composition of human breath. New scientist, (1964), 381, 608-61 1.

8. Chen, S. et al. Volatile fatty acids in the breath of patients with cirrhosis of the liver. J.Lab. Clin. Me , (1970), 75, 622-627.

9. Brown, D.J. A method for determining the excretion of volatile substances through skin. Meth. & Find Exptl. Clin. Pharmacology, (1985), 7, 269-274.

10. Gordon, S.M. Volatile organic compounds in exhaled air from patients with lung cancer. Clin. Chem, (1985), 31, 1278-1282.

11. Kneepkens, CM. , et al. The potential of the hydrocarbon breath t st as a measure of lipid peroxidation. Free Radical Biology & Medicine, (1994), 17,127-160.

12. Protsiuk, R.G., et al. Characteristics of the fatty acid composition of surfactant phospholipids in the condensate of the expired breath of patient with pulmonary tuberculosis. Vrach Delo., (1988), 77, 9-10.

13. Wang, C-H. , et al. Increased exhaled NO in active pulmonary TB due to iπducible NO synthase up regulation in alveolar macrophagcs. Euro. Respi. Journal, (1998), 11, 809-815.

14. Hayden, G.F. Olfactory diagnosis in medicine. Postgraduate Medicine, (1980), 67, 110-116.

15. Lidell, K. Smell as a diagnostic marker. Postgraduate Med. Journal, (1976), 52, 136-138.

16. Di Natale, C, et al. An electronic nose for the recognition of the vineyard of a red wine. Sensors & Actuators, (1996), B33, 83-88.

17. Gibson, T.D., et al. Detection and simultaneous identification of micro-organisms from headspace samples using an electronic nose. Sensors & Actuators, (1997), B44, 413-422.

18. Wang Ping, et al. A novel method for diabetes diagnosis based on electronic noses. Biosensors & Bioelectronics, (1997), 12, 1031-1036.

19. Hanson III, C.W., et al. The use of novel electronic nose to diagnose the presence of intrapulmonary infection. Anaesthesioiogy, (1997), 87(3A), A269.

20. Paviou, A., et al. An in vitro artificial model of early odour generation, rapid detection and recognition inn discrimination of H.pylori, and other gastrocosophagcal pathogens. World Congress Biosensors, 1998, Berlin.

21. Otsuka, Y., et al. Applications of Neural Network models to iron and steel manufacturing process. Proceedings of the 2^nd Int. Conf. On Fuzzy logic & Neural Networks, Lizuka, Japan, 1992, pp815-8I8.

Claims

1. Apparatus for use in diagnosing and/or monitoring gastric and/or lung disorders comprising: (a) a sampling system for collecting a gas sample generated by a patient or generated from a sample taken from the patient;

2. Apparatus according to claim 1 wherein said data processing system comprises a hybrid intelligent system that controls a search optimisation engine of genetic algorithms and a multiplicity of neural networks arranged to analyse said output signals using predetermined rules and thereby to determine a said pattern.

3. Apparatus according to claim 1 or claim 2 wherein the sample system is adapted to collect samples of gas generated in the patient's stomach or lung.

4. Apparatus according to claim 3 wherein the sampling system comprises an endoscope having a tubular probe adapted to be inserted into a body cavity of a patient.

5. Apparatus according to claim 1 wherein the sampling system is adapted to collect samples of gas generated in vitro from samples of material taken from a patient, said sampling system including a vessel defining a sample receiving volume and a headspace, and means for withdrawing a gas sample from the headspace.

6. A method of diagnosing and/or monitoring gastric and/or lung disorders comprising:

(a) collecting a gas sample generated by a patient or generated from a sample taken from the patient;

(b) passing gas from said gas sample to an array of gas sensors each having a different pattern of sensitivities to potential components of the gas sample and being adapted to provide an electrical output signal in response to detection of one or more of said components; and

(c) passing said output signals to a data processing system which analyses the output signals to detect patterns indicative of the presence of predetermined disorders and/or stages of predetermined disorders.

7. A method according to claim 6 which employs apparatus according to any of claim 1 to 5.

8. A method of claim 6 or claim 7 wherein the gas sample is generated in vitro from a sample of material obtained from a patient.

9. A method of claim 8 wherein the gas sample is generated by a process comprising treating said material sample with at least one enzyme.

10. A method according to claim 9 wherein said at least one enzyme comprises a lipase.

11. A method according to claim 9 or claim 10 wherein said material is treated with at least two different enzymes .

12. A method according to claim 11 wherein said at least two enzymes comprise a mammalian lipase and a fungal lipase.

13. A method according to claim 6 or claim 7 including a step of administering non-labelled urea to a patient and subsequently collecting a gas sample from the stomach.

14. A kit of parts for carrying out a method of any of claims 9 to 13 including apparatus according to claim 1 and at least one enzyme for use in the method of any of claims 9 to 12 or urea for use in the method of claim 13.