(95CI 76-98%) and a specificity of 71% (95CI 55-84%). Diagnostic properties were somewhat lower when microbiological culture results were used as reference standard;

(1)

(2)

(3)

ABSTRACT

Objectives: The aim of this study was to explore whether an electronic nose, Aetholab, is able to discriminate between infected versus non-infected wounds, based on headspace analyses from wound swabs.

Methods: A total of 77 patients participated in this pilot study. Each wound was assessed for infection based on clinical judgment. Additionally, two wound swabs were taken; one for microbiological culture and one for measurement with Aetholab. Diagnostic properties with 95%Confidence Interval (95CI) of Aetholab were calculated with clinical judgment and microbiological culture results as reference standards.

Results: With clinical judgment as reference standard, Aetholab had a sensitivity of 91%

(95CI 76-98%) and a specificity of 71% (95CI 55-84%). Diagnostic properties were somewhat lower when microbiological culture results were used as reference standard;

sensitivity 81% (95CI 64-91%), specificity 63% (95CI 46-77%).

Conclusions: Aetholab seems a promising diagnostic tool for wound infection given the

diagnostic properties presented in this pilot study. A larger study is needed to confirm

our results.

(4)

INTRODUCTION

In healthcare, odours are frequently used as a warning sign of disease. For example, ammoniacal odour from urine is associated with bladder infection, acetone smelling breath is usually related to Diabetes Mellitus and people with diphtheria often have a sweet sweat odour

¹

. To identify certain odours, our nose first has to pick up a certain smell, e.g. mixture of volatile organic compounds (VOCs), while our brain then tries to match it to a certain odour it has recognized before. The ability to translate mixtures of VOCs into identifiable odours is limited and differs between humans. That is one of the reasons for the development of electronic noses. Like the human nose, an electronic nose analyses a mixture of VOCs and matches it to previously observed patterns of VOCs

2

. The difference between an electronic and human nose is that electronic noses are able to analyse a broader variety of VOC mixtures in an objective manner.

Aetholab (figure 8.1), an electronic nose developed by The eNose Company (Zutphen, the Netherlands), was developed to measure VOCs in the headspace (i.e. the gaseous part of a sample container) of solid or liquid samples

²

. VOCs that are able to oxidize can interact with the Aetholab sensors; redox reactions take place upon contact with the sensors, which results into conductivity changes that are measured within the device.

The pattern of these conductivity changes can be used to differentiate between different odours

³

. Like the human nose and brain, the electronic nose needs to be trained to recognize odours. Therefore, during the calibration phase, a series of measurements with an electronic nose is labelled as ‘disease’ or ‘no disease’ based on a reference standard (e.g., based on clinical judgment). Next, an algorithm matches new measurements to the previously labelled ones

⁴

. An added value of this electronic nose is that algorithms can be transferred to other electronic noses to ensure that each electronic nose provides the same results

⁵

.

Over the past few years, several studies have demonstrated the ability of electronic noses to discriminate between presence versus no presence of several diseases (e.g.

lung cancer, colorectal cancer, head & neck cancer, tuberculosis)

^6-8

. This potential could

also show added value in detecting infection in complex wounds. These are wounds that

experience a delay or failure in the healing process, usually caused by underlying

comorbidities like vascular insufficiency or diabetes mellitus

^9,10

. Such longer existing

open wounds are prone to infection, which further delays wound healing, and can also

lead to hospitalization and sepsis. To prevent complications, it is important to timely and

accurately detect wound infection. Moreover, an accurate test correctly ruling out

infection could prevent unnecessary use of antibiotic treatment

¹¹

.

(5)

Figure 8.1. Aetholab used for analyses in this study. Closed device (left), open device (middle) with a built- in eNose in the lid for each sample container (four in total), and an example of a sample container (right).

Despite many efforts, currently no diagnostic aid is available to meet these needs.

Clinical judgment of wound infection is often used as diagnostic criterium. However, clinical signs and symptoms of wound infection are usually masked in patients with complex wounds

¹²

. Therefore, microbiological culture results are regularly used to support detection of infection. However, the presence of specific microorganisms in a wound is not necessarily related to wound infection

^13,14

. In addition, microbiological culture results are only available after a few days. Two earlier studies

^15,16

have explored the potential of electronic noses to fulfill the need for accurate and fast detection of wound infection in clinical practice. They demonstrated that electronic noses are able to discriminate between specific microorganisms present in a wound swab. To our knowledge, up till now, no study was presented on the ability of an electronic nose to detect actual presence of wound infection. Therefore, we conducted a pilot study to explore whether Aetholab is able to discriminate between patterns from infected versus non-infected wounds, based on headspace analyses from wound swabs.

METHODS

To determine the discriminative ability of Aetholab for wound infection, we included

patients with open wounds during a regular appointment at the wound care department

of Medisch Spectrum Twente (MST), location Oldenzaal, the Netherlands. Patients had

to be able to provide informed consent, and their wound should be suitable for wound

swabbing (e.g.; no completely dry wounds or wounds with a diameter that was smaller

than a standard wound swab). Given the exploratory character of this study, we aimed

(6)

to include a maximum of 80 patients, with an approximate ratio of 1:1 for infected and non-infected wounds to have a sufficient number of ‘events’ for the development of Aetholab algorithms. Patients were included consecutively, but were excluded when the targeted number of patients with (non)infected wounds was already reached (figure 8.2).

The decision to exclude these subsequent patients was based on initial clinical judgment of the wound status. Before the start of the study, approval was obtained from the ethical committee.

Figure 8.2. In- and exclusion of eligible patients in the study sample.

(7)

After providing informed consent, clinical information about the wound and relevant comorbidities were registered in a case report file. In addition, the wound was clinically assessed for infection by an expert clinician (> 20 years’ experience). The Clinical Signs and Symptoms Checklist

¹⁷

was used to structure the identification of signs and symptoms of wound infection. Subsequently the wound was cleaned with a gauze and a sterile saline solution. A wound swab (Copan ESwab

^TM

) for microbiological culture was taken according to the Levine technique, which includes rotating a swab with a small amount of pressure on a 1 cm

²

area of viable tissue in the wound. Immediately afterwards, a second wound swab was taken for analysis with Aetholab. The swab was taken from the same location and in the same manner as the swab taken for microbiological culture, and it was stored in a sterile Aetholab sample container. Both swabs were stored in a monitored refrigerator (4°C) until transportation. The first swab was sent to the microbiological laboratory for standard culture, while the second swab was sent to MST, location Enschede, for analysis with Aetholab.

Measurement with Aetholab

In the Aetholab, four (disposable) sample containers can be analysed simultaneously by four built-in eNoses. Each eNose consists of three different metal-oxide sensors (AMS AG, Austria) that measure conductivity changes when there is a reaction between the sensor and VOC’s from the sample. This results in an odour signature. Measurement time for one sample with Aetholab is 12.5 minutes. More detailed information about Aetholab can be found in the online supplement. During data analysis, odour signatures from infected and non-infected wounds are separated using artificial neural network (ANN) techniques. Data were retrieved on a laptop and processed further for data analysis.

Microbiological culture

In the microbiological laboratory, the wound swab sample was inoculated onto media for the detection of aerobic and anaerobic bacteria. Isolated pathogens were identified using MALDI-TOF (Bruker). Culture reports were provided semi-quantitatively (negative, + to +++) for each isolated pathogen.

Reference standard for wound infection

As there is currently is no definitive method to determine wound infection status, we

calculated diagnostic properties of Aetholab for two commonly used methods in clinical

practice: 1) clinical judgment 2) microbiological culture results. Clinical judgment of

wound infection was based on extensive clinical experience including the use of clinical

signs and symptoms of infection and other relevant wound information.

(8)

The second method consisted of judgment of microbiological culture results. Clinical information about the wound (e.g. wound diagnosis, clinical signs and symptoms of infection) was provided to enable interpretation of microbiological culture results.

Statistical analyses

Descriptive analyses (IBM SPSS Statistics, version 24) were used to describe the study population. Data obtained by Aetholab were analysed using Aethena, a proprietary software package developed by the eNose Company. In Aethena, the full process of data analysis is conducted, starting from data pre-processing to data compression, ANN training, cross validation, and reporting. During data compression, data are rearranged to optimally represent differences between groups, while at the same time reducing dimensions. Several mathematical methods can be used. In this case, a Tucker3-like algorithm is applied

¹⁸

. During ANN training, compressed data samples are fed to the ANN together with the wound infection status (based on a reference standard) in order to tune the ANN parameters. ANN training was performed for each reference test separately. ‘Leave-10%-out’-cross validation was used to make sure training was not based on artefacts. Further details about Aethena are described by Kort et al.

⁴

. Predicted scores based on Aethena, ranging from -1 (not infected wound) to +1 (infected wound), were plotted for each patient. Cut-off values were used to translate predicted scores into a dichotomous outcome (infected vs. non-infected); e.g. predicted scores above the cut-off value were classified as ‘infected’ and below the cut-off value as ‘non- infected’. Cut-off values were chosen at such a level that false positive/negative results were minimized. Sensitivity, specificity, positive and negative predictive value were calculated in comparison to each reference test. Receiver Operating Characteristic (ROC) curves were constructed and appropriate Areas Under the Curve (AUC’s) were calculated with 95% confidence intervals (95% CI).

RESULTS

Eighty patients were included in this study. Three patients had to be excluded from Aetholab measurements as their wound samples were not stored at 4°C until

transportation and measurement. The characteristics of the remaining 77 patients are

presented in table 8.1.

(9)

Table 8.1. Demographics study population.

Frequency (%)

Median (range)

Sex Male 41 (53.2)

Female 36 (46.8)

Age in years 77

(36 – 95)

Wound type Amputation wound 15 (19.5)

Diabetic foot ulcer 9 (11.7)

Venous/arterial leg ulcer 24 (31.2)

Postoperative wound 2 (2.6)

Pressure ulcer 18 (23.4)

Traumatic wound 8 (10.4)

Sinus pilonidalis 1 (1.3)

Wound existence (in weeks)

11.4 (0.4 – 138.6)

Use of antibiotics 9 (11.7)

Diabetes Mellitus 29 (37.7)

Thirty-five wounds (45%) were infected according to clinical judgment, while 37 wounds were assessed as infected based on microbiological culture results. Figure 8.3 depicts the predicted scores based on the data collected by Aetholab for each patient, with either clinical judgment (3a) or microbiological culture results (3b) as reference standard.

In both figures, patients with an infected wound (in red) tend to have higher predicted scores than patients with non-infected wounds (in green).

Diagnostic properties of Aetholab are presented in table 8.2. Sensitivity was fairly high;

91% (76-98%) and 81% (64-91%) respectively for clinical judgment and microbiological

culture as reference standard. The number of false positive results varied between 12

(16%; 95CI 8-26%) for clinical judgment and 15 (20%, 95CI 11-30%) for microbiological

culture, which resulted in a slightly lower specificity. Overall, diagnostic properties were

better when clinical judgment was used as reference standard.

(10)

Figure 8.3a. Scatterplot with predicted scores from Aetholab for each patient and a cut-off value of - 0.26. In red; patients with an infected wound, green; patients with a non-infected wound (according to clinical judgment).

Figure 8.3b. Scatterplot with predicted scores from Aetholab for each patient and a cut-off value of - 0.20. In red; patients with an infected wound, green; patients with a non-infected wound (according to microbiological culture results).

(11)

Table 8.2. Diagnostic properties of Aetholab, with clinical judgment and microbiological culture results as reference standard. AUC 0.82 0.76 NPV (95% CI)) 0.91 (0.75-0.98) 0.78 (0.60-0.90)

PPV (95% CI)) 0.73 (0.57-0.85) 0.67 (0.51-0.80)

Spec (95% CI)) 0.71 (0.55-0.84) 0.63 (0.46-0.77)

Sens (95% CI)) 0.91 (0.76-0.98) 0.81 (0.64-0.91)

Cut-off value -0.26 -0.20

Clinical judgment No infection 12 30 Microbiology No infection 15 25 Infection 32 3 Infection 30 7

Positive Negative Positive Negative

Aetholab Aetholab

(12)

DISCUSSION

Sensitivity of Aetholab for the detection of wound infection was high; 81% when microbiological culture was used as reference standard and 91% when the reference standard was based on clinical judgment. This means that 81% of the patients with an infection based on culture results was also classified as infected based on Aetholab; and in 91% of the patients with an infected wound based on expert clinical judgment Aetholab classified the wound as infected too. Specificity and positive predictive value were somewhat lower (63% and 67% respectively for microbiological culture, 71% and 73% respectively for clinical judgment). Thus, of all wounds that were classified as infected by Aetholab, 67% were also infected according to microbiological culture results. In 73% of the patients with an infected wound according to Aetholab, the clinician indicated an infection too. These first results indicate that Aetholab might be a promising tool for the detection of wound infection in clinical practice.

Despite the clinical need for accurate and timely detection of infection, only a few promising diagnostic aids have been developed that can be used in clinical practice.

Lammers et al. constructed a neural network-derived decision model to predict infection in uncomplicated traumatic wounds and used development of infection during follow- up as reference standard. They were able to predict infection with a sensitivity of 70%, specificity of 76% and positive and negative predictive values of 18% and 97%

respectively

¹⁹

. Blokhuis-Arkes et al. demonstrated fair diagnostic properties for a diagnostic tool based on inflammatory enzyme assays. Sensitivity varied between 55–

90%, specificity between 58-80% and positive and negative predictive values between 57-63% and 74-90% respectively

²⁰

. As demonstrated in the current pilot study, the diagnostic properties of Aetholab are non-inferior to these diagnostic tools.

Limitations

The robustness of our results is influenced by several factors. At first, this study was

designed as a pilot study and therefore included no more than 77 patients. A small

sample size does not only influence the generalizability and robustness of predictive

models based on Aetholab measurements, but it also allows substantial variation around

point estimates (e.g. diagnostic properties). Moreover, our study sample comprised of

approximately equal numbers of infected and non-infected wounds, which is not

representative for our, and probably other, wound clinics. A lower prevalence of wound

infection, as is expected in clinical practice, would have resulted in a lower positive

predictive value of Aetholab. For example, with a prevalence of 25% (as expected in our

wound care clinic), positive predictive value of Aetholab would have been 50% (33-67%)

(13)

with a negative predictive value of 95% (83-99%) when compared to clinical judgment as reference standard. The primary aim of this study was, however, to investigate the feasibility of Aetholab in the differentiation between infected and non-infected wounds and use the results to decide on conducting a full diagnostic study. Secondly, our results strongly depend on the reference test used to determine presence of wound infection.

Since a gold standard for wound infection is currently missing, we used two regularly used diagnostic reference standards for wound infection in clinical practice; clinical judgment and microbiological culture results. However, the reference tests are known to be troublesome and may not have represented the true infection status of the wound.

This has influenced both the training phase of Aetholab and the final calculation of diagnostic properties. Nevertheless, in our opinion, we have used the most appropriate reference tests available for this study.

Although there are some limitations to this study, diagnostic properties of Aetholab are promising. Moreover, measurement time of Aetholab is 12.5 minutes for each sample, which enables clinicians to obtain results during a patient’s wound care appointment.

Given the promising results, we aim to design and conduct a larger scale study to determine the diagnostic performance of Aetholab more accurately. This study will include a larger sample size, originating from multiple study centres to increase generalizability of study results. Secondary aims could be differentiating between (species of) microorganisms. Bruins et al. have demonstrated that electronic noses are able to identify pathogens and that the used algorithms for differentiation are transferrable between electronic noses

²¹

. Saviauk et al. performed a proof-of-concept study in which they found that headspace analyses using an electronic nose can differentiate between several microorganisms that are frequently seen in infected wounds

¹⁶

. Therefore, it seems promising to explore whether Aetholab is able to identify pathogens from wound samples. The combination of wound infection detection and microbiological discovery into one tool may support appropriate use of initial antibiotic treatment, since Aetholab is able to provide results within 15 minutes instead of 3-5 days for culture results. Microbiological culture results, however, remain necessary to determine antibiotic susceptibility.

TRANSPARENCY DECLARATION

M. Haalboom has nothing to disclose; Dr. J. van der Palen reports grants from The eNose

Company, outside the submitted work; Dr. Gerritsen reports personal fees as employee

from The eNose Company.

(14)

REFERENCES

1. Pavlou AK, Turner AP. Sniffing out the truth: clinical diagnosis using the electronic nose. Clinical chemistry and laboratory medicine. 2000;38(2):99-112.

2. Bruins MG. Transferable Odor Differentiation Models for Infectious Disease Diagnostics, Erasmus University Rotterdam; 2014.

3. Bruins M, Gerritsen JW, van de Sande WWJ, van Belkum A, Bos A. Enabling a transferable calibration model for metal-oxide type electronic noses. Sensors and Actuators B: Chemical. 2013;188:1187-1195.

4. Kort S, Brusse-Keizer M, Gerritsen JW, van der Palen J. Data analysis of electronic nose technology in lung cancer: generating prediction models by means of Aethena. Journal of breath research. 2017;11(2):026006.

5. Bruins M. Transferable Odor Differentiation Models for Infectious Disease Diagnostics [PhD Thesis]. Enschede, Erasmus Universiteit Rotterdam; 2014.

6. Fitzgerald J, Fenniri H. Cutting Edge Methods for Non-Invasive Disease Diagnosis Using E-Tongue and E-Nose Devices. Biosensors. 2017;7(4).

7. Kort S, Tiggeloven MM, Brusse-Keizer M, et al. Multi-centre prospective study on diagnosing subtypes of lung cancer by exhaled-breath analysis. Lung cancer (Amsterdam, Netherlands). 2018;125:223-229.

8. Wilson AD. Advances in electronic-nose technologies for the detection of volatile biomarker metabolites in the human breath. Metabolites.

2015;5(1):140-163.

9. Demidova-Rice TN, Hamblin MR, Herman IM. Acute and Impaired Wound Healing: Pathophysiology and Current Methods for Drug Delivery, Part 1:

Normal and Chronic Wounds: Biology, Causes, and Approaches to Care.

Advances in skin & wound care. 2012;25(7):304-314.

10. Haalboom M. Chronic wounds: Innovations in diagnostics and therapeutics.

Current medicinal chemistry. 2017.

11. Lipsky BA, Dryden M, Gottrup F, Nathwani D, Seaton RA, Stryja J. Antimicrobial stewardship in wound care: a Position Paper from the British Society for Antimicrobial Chemotherapy and European Wound Management Association.

The Journal of antimicrobial chemotherapy. 2016;71(11):3026-3035.

12. Reddy M, Gill SS, Wu W, Kalkar SR, Rochon PA. Does this patient have an infection of a chronic wound? Jama. 2012;307(6):605-611.

13. Bowler PG. The 10(5) bacterial growth guideline: reassessing its clinical

relevance in wound healing. Ostomy/wound management. 2003;49(1):44-53.

(15)

14. Bowler PG, Duerden BI, Armstrong DG. Wound microbiology and associated approaches to wound management. Clinical microbiology reviews.

2001;14(2):244-269.

15. Pisanelli AM, Persaud KC, Bailey A, Stuczen M, Duncan R, Dunn K.

Development of a diagnostic aid for bacterial infection in wounds. AIP Conference Proceedings. 2009;1137(1):133-135.

16. Saviauk T, Kiiski JP, Nieminen MK, et al. Electronic Nose in the Detection of Wound Infection Bacteria from Bacterial Cultures: A Proof-of-Principle Study.

Eur Surg Res. 2018;59(1-2):1-11.

17. Gardner SE, Frantz RA, Doebbeling BN. The validity of the clinical signs and symptoms used to identify localized chronic wound infection. Wound repair and regeneration : official publication of the Wound Healing Society [and] the European Tissue Repair Society. 2001;9(3):178-186.

18. Kroonenberg PM. Applied Multiway Data Analysis. Wiley; 2008.

19. Lammers RL, Hudson DL, Seaman ME. Prediction of traumatic wound infection with a neural network-derived decision model. The American journal of emergency medicine. 2003;21(1):1-7.

20. Blokhuis-Arkes MH, Haalboom M, van der Palen J, et al. Rapid enzyme analysis as a diagnostic tool for wound infection: Comparison between clinical judgment, microbiological analysis, and enzyme analysis. Wound repair and regeneration : official publication of the Wound Healing Society [and] the European Tissue Repair Society. 2015;23(3):345-352.

21. Bruins M, Bos A, Petit PL, et al. Device-independent, real-time identification of

bacterial pathogens with a metal oxide-based olfactory sensor. Eur J Clin

Microbiol Infect Dis. 2009;28(7):775-780.

(16)

ONLINE SUPPLEMENTARY TEXT

Detailed description of Aetholab measurements

In the Aetholab, four (disposable) sample containers can be analysed simultaneously by four built-in eNoses. Each eNose consists of three different metal-oxide sensors (AMS AG, Austria) that are guided through a temperature profile. VOC’s originating from the sample containers that react at the sensor surfaces cause conductivity changes. During this process, sensor conductivities are measured for 32 times during each cycle. The cap of the sample containers contains two septa that are perforated with a needle-shaped tube for extracting the headspace gases.