GB2364571A

GB2364571A - Diagnosing and/or monitoring urinary tract infection

Info

Publication number: GB2364571A
Application number: GB0008349A
Authority: GB
Inventors: Anthony Peter Francis Turner; Alexandros Pavlou
Original assignee: Cranfield University
Current assignee: Cranfield University
Priority date: 2000-04-06
Filing date: 2000-04-06
Publication date: 2002-01-30
Also published as: GB0008349D0

Abstract

A urine sample is incubated with suitable growth medium. Volatiles produced are then extracted from the medium, eg by passing air through it, and passed to an array of sensors having different responses to the different volatile materials. The sensor outputs are analysed, preferably by neural networks and genetically, and the results are used to characterise the infection. The sensors may be conducting polymers. The growth medium may include amino acids.

Description

2364571

Introduction

Urinary tract infections (UTI) are a significant cause of morbidity with 3 million UTI cases each in the USA alone'. Thirty-one percent of nosocomial infections in medical intensive care units are attributable to UT12, and it is estimated that 20% percent of females, aged of 20 and 65 years suffer at least one episode per year. There are also links to other complicated or 3,4 chronic urological disorders such as pyelonephritis, urethritis, and prostatitis Approximately 80% of uncomplicated UTI are caused by Ecoli and 20% by enteric pathogens such as Enterococci, Klebsiellae, Proteus sp., coagulase (-) Staphylococci and 5,6 fungal opportunistic pathogens such as Candida albicans. Current diagnostic techniques require 24-48 hrs to identify pathogenic species in urine midstream specimens (: 105c ml-1) and apply antibiotic sensitivity tests. Despite the introduction of molecular tests, microscopy and culture remain the gold standard in every day clinical practice.

During the 1970s there were some early attempts to develop automated urine volatile analytical devices using advanced gas-chromatographic (GC) techniques and more than 200 volatile organic compounds (VOC) were measured 7. Additionally, others studied the stimulatory effect of certain biochemical precursors in bacterial volatile production and rapid 8 GC discrimination between Ecoli and Proteus sp., or the generation of specific urinary volatile profiles of patients with metabolic disorders, postulating the existence of certain urine biomarkers9. However lack of advanced computerised techniques to analyse specimen complexity, laborious techniques, high cost and the need for highly skilled personnel, prevented these methods from being introduced to clinical practice. Recently Phillips and co-workerslo have revealed the role of VOC as important markers in lung cancer diagnosis.

OR In this study we report the analysis of 70 specimens of human urine by the application of an intelligent diagnostic model based on novel generation, detection, and rapid recognition of urinary volatile patterns within 5 hrs of receipt of specimens in the laboratory. Methods 1" UTI recognition problem Urine samples and volatile generating kits (VGK) Twenty-five anonyinised mid steam urine samples (each approx. 20ml) were obtained from patients with a clinical diagnosis of acute uncomplicated urinary tract infection and submitted for standard microbiological analysis. From each sample, after aseptic filtering extraction of eukaryotic cells, 5ml of urine were inoculated into specially made polypropylene centrifuge bottles (50ml, Sterilin) each containing a complex medium of 60% v/v brain heart infusion (BHI) broth (Oxoid), 40% cooked meat broth (Oxoid), O.Olmg ml-1 Benzamide (Sigma), lmg ml-1 Urea, Lactose and Arabinose (Sigma), 0.05mg ml-1 Porcine/Gelatine (Oxoid), 0. 001mg ml-1 Tween 80 (Sigma), 0.75mg ml-1 of a series of amino acids (LSerine, L-Methionine, LLeucine, L-Proline, L-Valine, L-Histidine, LTryptophan, Sigma) mixed to a final volume of 20ml and incubated for 4 1/2 hrs at 37'C aerobically. Flow injection analysis (FIA) of urinary volatiles After 4 1/2 hrs of incubation to coincide with the logarithmic phase of growth, 25 VGK were placed in a 37'C water bath and directly connected with a specifically designed air-filtered sparging (bubbling) system. This consisted of Teflon tubing (Tygon), a hydrophobic biofilter (0.45Lrn PTFE, Whatman-Hepavent) and an activated carbon filter (Whatman) to provide clean air- flow above the urine headspace. A flow rate of 200ml min-1 was set automatically and environmental conditions at the sampling point were continuously monitored. The actual urine sampling time and baseline recovery per specimen was 3 min. Urine Sensorial Analysis A 14 gas-sensor array (model BH1 14-Bloodhound, Leeds, UK) employed a set of specifically tailored conducting polymers which may physically interact with urinary volatile compounds to produce a change in electrical resistance. A data capture software analysed complex volatile patterns liberated over the headspace of 25 UTI-VGK specimens and extracted multiple sensor parameters from each patient profile for ftirther artificial intelligence analysis. Intelligent UTIpattern recognition system According to a standard PHLS diagnostic protocol 20 UTI out of 25 randomly selected patient samples were identified. Nine samples were found to be infected with E. coli (e), 5 with Proteus mirabilis (p), 6 with mixed infection (ep) of Proteus sp. and Gram (+) cocci (normal flora) followed by 5 normal (n) urine samples. Microscopy and culture on selective media confirmed the diagnosis in every VGK sample. In order to improve pattern classification a set of 56 sensor parameters was extracted from 14 gas-sensor responses. Two patient sample groups were randomly selected for AI analysis: a "training set" of 16 UTIVGK (e: 6 samples, p: 3, n: 3, ep: 4, 64% of UTI patient data) and an untrained group of 9 "unknown" UTI cases (e: 3, p:2, ep:2, n:2, 36% of data). Sensor data processing employed a hybrid intelligent system of Genetic Algorithms, Back Propagation Neural Networks (GANN) and multivariate techniques such as non-parametric Principal Components Analysis (PCA), parametric Discriminant Function Analysis and cross validation (DFA-cv). Genetic supervising uses an optimisation engine, which consists of models of evolutionary combination of all input sensor parameters. A number of NN features, an addition of crossing over and a mutation rate can be used to evolve the "fittest" NN solution for a specific diagnostic problem 11,12. Evolution towards the "best" NN configuration -to learn the training set and recognise the "unknown" UTI- was processed through 10 generations. An "immigration" mode evolved a NN population pool (5NNs /generation) and replaced the "weakest" phenotypes. A setting of 2 cross-overs determined the frequency of intermingling of NN features within the same topology and a mutation rate of 0.6 created new NN phenotypes. "Genetically" selected sets of sensor parameters were also used as input variables in multivariate analysis (PCA and DFA-cv). 2 "d UTI recognition problem Urine samples and volatile generating kits (VGK) Forty-five 5ml urine samples (following eukaryotic cell filtering extraction) were collected from randomly selected patients admitted in Gloucestershire PHLS and inoculated into specially made centrifuge bottles (50ml, Sterilin) each containing 95% BHl broth (Oxoid), 5% serum bovine (Oxoid), 0.70mg ml-1 of a series of amino acids (L-Leucine, L- Alanine, LSerine, L-Valine, L-Asparagine, L-Glutamine, L-Methionine, Sigma), ling ml-1 Urea (Sigma), 0.75mgml-l Lactose (Sigma), O.Img ml-1 Casein (Oxoid), 0.3mg ml-1 Acetylcholine (Sigma) to a final volume of 20ml per VGK and incubated aerobically for 5 hrs at 37C. The urine volatile delivery and sensorial systems were the same as described previously (I st UTI problem). Intelligent UTIpattern recognition system Thirty cases of UTI were identified from 45 randomly selected samples by standard microscopy and culture: 13 patients were infected with E coli (e) , 9 with Proteus sp. (p) and 8 with coagulase (-) Staphylococcus sp., (st) . Two genetic training algorithms processed urine V data through a parallel evolutionary succession process towards competent NN solutions. The first GA analysed patient data which had been randomly divided into a "training" group of 31 urinary samples (e: 9, p: 6, st: 5 and n: 11) and a group of 14 "unknowns" (e: 4, p: 3, st: 3 and n: 4, 3 1 % of patient collected data). An evolutionary process of 5 generations (3 NNs/generation) was carried out employing I crossover and a mutation rate of 0.5. Additionally the second GA performed a much broader evolutionary optimisation analysis of 100 generations. It also attempted to analyse the same amount of patient data but with a higher ratio of "unknown" proportion (42% of collected patient data) including 26 training samples (e: 8, p:4, st:4, n:10) and 19 "unknown" UTI (e:5, p:5, st:4, n:5). A population of 600 NNs was evolved using an immigration mode, 2 crossovers and a mutation rate of 0.7 towards the "fittest" NN solution. Both "genetically" selected sensor parameters were also used to perform PCA and DFA-cv. Results I" UTI recognition problem The Genetic supervisor selected a 3-layer back-propagation NN of 35 input (normalised sensor parameters), 14 hidden and 4 output neurones. A log-sigmoid activation function was used to discriminate between 4 UTI classes. For every input urine sample a numerical target describing confidence of prediction was set as k-- I with a UTI output tolerance limit of 0.5.

A momentum of 0.35 determined the lifetime of a correction term as the training process of 16 urine VGK took place. An input noise of 0.0 16 provided a slight random variation to each UTI pattern for every training cycle. Additionally an adaptive learning rate speeded up the training procedure. The GA-NN managed to "learn" all samples after 3840 epochs and achieved a UTI prediction rate of 100%. All 9 previously untrained "unknown" UTI cases were identified correctly with a confidence of prediction ranging from 0. 81 to 1.02. Thirtyfive "genetically" selected input neurones (sensor parameters) were used to perform PCA and DFA-cv so that all UTI clusters could be displayed graphically. PCA accomplished non-parametrically a significant dimension reduction by minimising minor UTI data variations so that information could be depicted on a few two-dimensional principal component score plots. There was a complete discrimination between all UTI classes (Figure 1a). DFA identified a group of parameters that best discriminated between the 4 urine groups. Two conditions were satisfied: (a) the distance between the UTI classes was as far as possible and (b) distances within UTI clusters were as close as possible. Eventually a new axis Z was identified such that it could achieve maximum discrimination. A clear 4group separation was achieved and cross-validation reclassified correctly 7 "unknown" UTI cases (Figure 1b). Furthermore by extracting all "genetically" selected sensor parameters that had been previously used as input neurones it was possible to reveal hidden non-linear patterns characteristic of each UTI group. A combination of good sensorial sensitivity, and reproducibility was also observed (Figure 2). Figure 2: Extraction of 35 "genetically" selected sensor parameters reveals 4 different groups of non-linear patterns, (a): 9 E.coli infection, (b): 6 mixed infection, (c): 5 Proteus mirabilis infection and (d): 5 normal urine. On X-axis, (divn) stands for Divergence which describes the peak sensor maximum response. Absorption (ab.) measures the rate of resistance change during volatile absorption on the conducting polymer surface. Additionally (dsn) represents the change of resistance during desorption of volatiles and area (an) that calculates the area under the actual sensor-gram. The Y-Axis measures resistance response 2"d UTI recognition problem Two parallel evolutionary algorithms selected 2 NN solutions. The first was a 3-layer (28-124) back-propagation NN that used an adaptive learning rate, a momentum of 0.42, an input pattern noise of 0.03 and achieved a 98% prediction rate. Thirteen out of 14 "unknown" UTI samples were identified correctly with a prediction output confidence ranging from 0.75 to 1.01. The intelligent system failed to characterise only one urine sample previously diagnosed with E. coli infection. However, this single pattern confusion was limited to the case of distinguishing between E. coli infection and normal urine. Both their prediction confidence outputs were very close-0.37 for Ecoli and 0.43 for normal urine- but below a 0.5 test tolerance limit. Twenty-eight "genetically" selected parameters performed PCA and DFA, which displayed two graphical cluster separations between Proteus sp., Staphylococcus Sp. UTI and normal samples. Cross-validation reclassified correctly 6 "unknown" patient samples (Figures 3a & 4a). Figure 3. Extraction of "genetically" selected sensor parameters and two-dimensional representations of PCA clustering between: a. normal urine (n), Proteus sp. (p) and Staphylococcus sp. (st) and b: E.coli (e), Proteus sp. (p) and Staphylococcus sp. (st). (Inner and outer circles divide most closely linearly discriminated patterns from the most drifted ones, respectively). Furthermore the second 3-layer NN (22-15-4) achieved a 95% prediction rate and recognised 18 out 19 "unknown" UTI cases. Only one norinal patient sample had been mistaken for E coli infection. A two-dimensional discrimination plot between 3 of the tested UTI groups (e, st, p) was produced by PCA. DFA also separated patient samples infected with E coli, Staphylococcus sp. and non-nal urine samples. Cross validation recognised 7 "unknown" UTI cases (Figures 3b & 4b).

Figure 4. DFA and 3-group separation between: a. normal urine (n), Proteus sp. (p) and Staphylococcus sp. (st) and b. normal urine, E.coli (e) and Staphylococcus sp. Cross-validation reclassified 6 (a) and 7 (b) "unknown" UTI cases.

Discussion It has long been recognised that analysis of urine may provide valuable clinical and physiological information. In this study we have demonstrated that the application of a novel diagnostic technique combining sensor technology with artificial intelligence may lead to rapid and acurate discrimination between different infective organisms in fresh samples of urine, based on the patterns of volatile compounds produced.

In the past three decades several investigators have used gas chromatography (GQ to 13 study urine volatile compounds, postulating a potential diagnostic role in infectious and 14 metabolic diseases. Other investigators had introduced the use of biochemical nutrient precursors in volatile stimulation, such as methionine and lactose, which could trigger the production of dimethyl sulphide by Proteus sp. and ethanol by Ecoli, respectively"" within 4 bours. The deterministic nature of these studies was focused on quantitative analysis of individual volatile markers by using conventional statistical methods. However highly complex biological fluids such as urine are difficult to analyse during an active infection by conventional quantitative techniques and linear methods. It has also been shown in the past that different regions of the MHC complex can contribute to the unique individual mammalian urine volatile composition and increase its complexity 16. Lack of advanced computing methods, userfriendly techniques and detection of consistent hidden diagnostic patterns prohibited those early attempts from reaching their diagnostic goal.

We are currently investigating the hypothetical existence of 2 non-linear dynamic 17 systems, in a move towards a novel diagnostic protocol: "a" complex metabolic changes during active UTI and their expression as volatile pulses and "b" the interaction of chemosensory surfaces and urinary volatile groups and their complex pattern recognition. To understand the complexity of in vivo UTI metabolic patterns and according to our hypothesis we developed a model, which has been applied successfully in this study. Four interdependent factors appeared to be key in our model: 1. Biochemical stimulation of UTI patterns, where the use of complex media managed to drive each pathogen into different catabolic pathways within 4-5 hours and release a number of non-linear volatile patterns (to achieve this there was a direct input from Biotechnology 18-20 and specially adjusted physiological parameters of VGK such as water activity and nutrient concentration increased the stimulatory volatile effect); 2. a FIA system for rapid volatile delivery; 3. a group of sensitive conducting polymers, which sensed the UTI pulses during metabolic pathogen-VGK reactions and transformed them to non- linear patterns; and 4. a hybrid intelligent system of multiple parallel GA-NN algorithms that evolved the best diagnostic solutions and recognised "unknown" UTI cases. The above intelligent system introduces a unique biochemical "dialogue" with UTI pathogens and the disease itself de profundis. It sensed their active metabolic pulses and recognised complex profiles of such a dynamic physiological phenomenon.

Further work is currently being undertaken towards the development of a new generation of diagnostic instruments based on sensors and artificial intelligence capable of providing rapid detection of infectious agents in vitro and in vivo, with enormous implications for future clinical practice.

0

Claims

1. Apparatus for use in rapidly diagnosing and / or monitoring urinary tract infection comprising:

(a) A vapour or gas generating system which by interaction with a urine sample taken from the patient results in the production of volatile chemicals that are characteristic of the infection; (b) A volatile chemical delivery system which delivers the product of the vapour or gas generating system to the detection system in a precise and reproducible way; (c) a detection system comprising an array of sensors each having a different pattern of sensitivities to potential components of the generated vapour or gas and being adapted to provide an electrical output signal in response to one or more of said components; (d) a data processing system arranged to receive said electrical output signals, said data processing system being adapted to analyse the output signals to detect patterns indicative of the presence of predetermined infections and / or stages of predetermined infections.

2. Apparatus according to claim I 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 said pattern.

3. Apparatus according to claim I or claim 2 wherein the detection system comprises an array of partially specific conducting polymer sensors.

4. Apparatus according to any of claim I to 3 wherein the vapour or gas delivery system comprises a flow injection system.

5. Apparatus according to any of claim I to 4 wherein the vapour or gas generating system is adapted to contain a sample of urine taken from a patient, said vapour or gas generating system including a mixture of chemicals which interact with the sample, a vessel defining a sample receiving volume and a headspace, and a means for withdrawing a gas sample from the headspace.

6. A method of rapidly diagnosing and / or monitoring urinary tract infection comprising:

(a) generating volatile chemicals that are characteristic of an infection from a urine sample taken from the patient; (b) passing said volatile chemicals to a detection system in a precise and reproducible way said detection system comprising an array of sensors each having a different pattern of sensitivities to potential components of the generated vapour or gas and being adapted to provide an electrical output signal in response to one or more of said components; (c) passing said electrical output signals to a data processing system said data processing system being adapted to analyse the output it signals to detect patterns indicative of the presence of predetermined infections and / or stages of predetermined infections.

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

8. A method of claim 6 or 7 wherein the vapour or gas is generated from the sample by incubation at a controlled temperature with a mixture of chemicals.

9. A method according to claim 8 wherein the mixture of chemicals is capable of supporting significant growth of micro-organisms.

10. A method according to claim 8 or claim 9 wherein said mixture of chemicals includes one or more amino acids.

11. A method according to claim 10 wherein said amino acids comprise one or more of L-serine, L-methionine, L-leucine, L-proline, L-valine, Lhistidine and L-tryptophan.

i