General characteristics of the studies
The cited methodology’s systematic search resulted in nine studies included in the final sample, which used the eNose to identify bacteria present in infected wounds. An overview of the studies is presented in Table 2, whose temporal distribution shows a higher frequency of publications in the years 2017 (n: 03) and 2019 (n: 03). 66.7% (n: 06) of the experimental studies were conducted in China, while the rest were conducted in Spain, Finland, and the Netherlands.
As one way to verify the reliability of these studies, it was found that 77.8% (n: 07) of the research was published in journals/conferences classified as Qualis A1 (highest level) by the Coordination for the Improvement of Higher Level Personnel (CAPES) in Brazil. At the global level, the Journal Citation Reports (JCR) 2021 impact factor of these journals were analyzed, presenting a range of 2.351 to 8.236 (minimum and maximum scores respectively), demonstrating that research whose objective is the presentation/validation of devices such as eNose in the diagnosis of infected wounds are published in journals or conferences of great relevance worldwide.
Table 2
General data of the studies included in the sample.
ID | Authors/Year | Objectives | Country | Periodical/ Qualis (JCR) | Type of study | Database |
01 | H. Sun et al /2017 | Identification of Escherichia coli, Staphylococcus aureus, Pseudomonas aeruginosa bacteria in wound infection using an eNose system with 34 sensors | China | IEEE Transactions on Industrial Electronics/ A1-8,236 | Experimental Study | IEEE |
02 | P. He et al /2017 | Present a type of self-directed learning based on a sparse autoencoder for eNose in wound infection detection. | China | Sensors (Basel)/ A1-4,019 | Experimental Study | Pubmed |
03 | Z. Liang et al /2017 | To present a new interference suppression method based on removing correlated information in eNose for bacterial detection. | China | Analytica chemistry acta / A1-5,977 | Experimental Study | Pubmed |
04 | T. Saviauk et al /2018 | Identify the most common bacteria species causing skin infections using eNose. | Finland | European Surgical Research / B1-2,351 | Experimental Study | Pubmed |
05 | Z. Yuan et al /2018 | Report the design of an embedded system using sensors, a wearable device, and AI models to notify patients with open wounds caused by diabetes about their healing status. | China | IEEE Asia Pacific Conference on Postgraduate Research in Microelectronics and Electronics | Experimental Study | IEEE |
06 | M. Haalboom et al /2019 | Explore whether an eNose, Aetholab, is able to discriminate between infected and uninfected wounds based on headspace analysis of wound swabs. | Nederland | Clinical microbiology and infection / A1-7,117 | Pilot Study | Pubmed |
07 | Z. Liang et al /2019 | Report an eNose system designed and used for the detection of bacteria in wound infection. | China | Sensors (Basel)/ A1-4,019 | Experimental Study | Scopus |
08 | C.S. Alvarez et al / 2019 | Experiment with the use of low-cost gas sensors to detect bacteria in wounds using a non-intrusive technique. | Spain | Sensors (Basel)/ A1-4,019 | Experimental Study | Scopus |
09 | T. Sun et al /2020 | Integrate two odor detection technologies - eNose and FAIMS (field ion mobility spectrometry) - to discriminate between three common bacterial wound infections and non-infectious by directly smelling mouse wound samples. | China | Sensors and Actuators B: Chemical/ A1-7,100 | Experimental Study | Scopus |
Characteristics Of Enoses And Protocol Of Experiments
Characteristics of eNoses and protocol of experiments
The eNose technology applied in diagnosing infected wounds in the studies included in the sample has their characteristics, where only two were presented as to their type in the publications analyzed (ChemPro 100 and Aetholab), identified in studies ID04 and ID06, respectively [24, 25]. The other researches presented only the characteristics of the eNoses used, their conformation composed of sampling, detection, and control unit.
The basic protocol for collecting data from samples of microorganisms (bacteria) for analysis in an eNose consists initially of growing these bacteria in Petri dishes inside an oven with a defined average temperature. Data is then extracted from the Petri dish with the infectious agent. The aspiration of the VOCs released by these bacteria occurs in a few seconds; the aspirated air is injected into an eNose air chamber with the sensors connected, remaining in this compartment for seconds for the molecules to interact with the sensor surfaces. In the next step, the air content of the chamber is purged through suction and injection of filtered air without VOCs. Throughout the described protocol, data is collected, which are signals from the different sensors inside the device.
The research protocols identified in the studies reviewed, can serve as a basis for performing the proposed methods for analyzing the VOCs released by bacteria more accurately. Each sample used in the identified eNoses devices was subjected to a measurement ranging from 10 minutes (ID04) [24] to 28 minutes (ID02) [26], the most common being a total cycle time of approximately 14 minutes (ID01, ID03, ID07, ID09) [27, 28, 28, 29] (Table 3). The experiments involved a period for baseline collection, whose time ranged from 3 to 5 minutes, a period for sample collection, which ranged from 3 to 4 minutes, and a period for purging the system, which ranged from 4 to 8 minutes. The number of samples ranged from 48 (ID09) [29] to 2664 (ID02) [26].
Type Of Bacteria And Wounds Analyzed
The bacteria samples used in the analyzed studies represented or simulated infection of seven types of wounds: Infected skin wound (ID01, ID02, ID03, ID04, ID07, ID09) [27, 26, 28, 24, 30, 29], diabetic foot ulcer (ID05, ID06) [31, 25], pressure injury (ID06, ID08) [25, 32], postoperative wound (ID06) [25], traumatic wound (ID06) [25], venous/arterial ulcer (ID06) [25] and Sinus pilonidalis (ID06) [25]. The wide diversity of scenarios makes it possible to evaluate the application of eNose in various real-world wound contexts.
This also applies to the microorganisms analyzed by eNose, the representation of which can be found in Table 3. 15 species of bacteria had their VOC analyzed by eNose, of which the most used were: P. aeruginosa (n: 08), E. coli (n: 07) and Methicillin Susceptible Staphylococcus aureus - MSSA (n: 07). Five studies obtained the bacterial cultures from in vivo experiments (ID02, ID03, ID06, ID07, ID09) [26, 28, 28, 30, 29]. However, only two collected the samples from patients with infected wounds (ID03, ID06) [28, 25].
Artificial Intelligence techniques used to classify models and their effectiveness
Regarding the Machine Learning algorithms used to classify the trained eNose models, eight classifiers were used, highlighting Support Vector Machine (SVM), used in four studies (ID01, ID03, ID05, ID07) [27, 28, 31, 30]. Different parameter optimization methods were tested in the included studies, mainly Principal Component Analysis (PCA).
Research ID02 [26] presented a new approach to train models using self-directed learning, a new machine learning framework. Other studies (ID03, ID07) [28, 30] were concerned with solving some deficiencies that make eNoses face some interference problems, since situations like this can affect system performance and interfere with the experiment.
Some algorithms such as Extreme Learning Machine (ELM), PCA, Orthogonal Signal Correction (OSC), Domain Regularized Component Analysis (DRCA), Domain Correction and Adaptive Extreme Learning Machines (DC-AELM) and Domain Adaptation Extreme Learning Machines (DAELM) were compared. with a new algorithm - Subspace Alignment-Based Interference Suppression (SAIS) [30]. The method is motivated by the different distribution of data collected on different bacterial agents and by the idea of subspatial alignment. The proposed method eliminates the need to add a regularization term and adjust the corresponding parameters [30]. Another new method was used to suppress interference in eNose: Interference Suppression Based on Correlated Information Removal (CIRIS) in another experimental study. It is the “removal of correlated information” rather than removal of uncorrelated information, in order to suppress interference [28].
The efficacy tests presented in the studies were heterogeneous as to whether there was a standard of evaluation. The summary of test values can be identified in Table 3. Sensitivity, specificity, and accuracy values were not presented in all studies, demonstrating a weakness in reporting the results. Thus, only the precision values were considered for analysis purposes.
The average accuracy of eNoses in identifying bacteria present in infected wounds was 95.13% for the training set, with a minimum value of 78.3% from the Partial Least Squares Discriminant Analysis (PLSDA) classifier [4] and a maximum of 100% with the Radial Basis Function (RBF) [27]. On the test set, the average accuracy value of the eNoses was 91.5%, with values of 70.19% (SVM - PCA) [27] and 100% (ELM; DC-AELM; DAELM; SAIS; ELM - PCA) [30] representing the minimum and maximum values, respectively. These results cannot be generalized due to the different classification methods used and the different species of bacteria used.
In the ID01 study [24], the fact of verifying the recognition rate of classification without optimization of the sensor array using only SVM stands out, presenting relatively low accuracy values (96.32% - training; 86.56% - test), in contrast with the application of optimization methods (Table 3), whose LDA was the best method under the circumstances performed (98.16% - training; 95.19% - test) and PCA the worst method (91.18% - training; 70.19% - test).
We highlight the study of Saviauk [24], whose eNose was able to differentiate MSSA and MRSA with an accuracy of 78%, demonstrating that the differentiation between bacteria of the same genus but sensitive or resistant to antibiotics can also be performed through Artificial Intelligence.
Table 3
Main results of the included studies.
ID | Wound type | Analyzed Bacteria | Sample origin | N (sample)/ Total cycle time cycle | AI techniques | Efficacy Test Results (Accuracy) |
| Training Phase | Test (validation) phase |
01 | Infected cutaneous wound | E. coli, MSSA*, P. aeruginosa | In vitro | 480/ 14 minutes | Classifier: Support Vector Machine (SVM); Linear Discriminant Analysis (LDA) Optimization methods: Principal Component Analysis (PCA); Genetic Algorithm (GA); Wilks’ Statistics; Mahalanobis distance | SVM: 96.32%. Sensor optimization: LDA: 98.16% Wilks’ statistic: 97.43% Mahalanobis distance: 97,43% GA: 97.43% PCA: 91.18 | SVM: 86.54%. Sensor optimization: LDA: 95.19% Wilks’ statistic: 87.50% Mahalanobis distance: 93,27% GA: 90.38% PCA: 70.19%. |
02 | Infected cutaneous wound | E. coli, MSSA*, P. aeruginosa | In vitro (mouse) | 2664/ 28 minutes | Classifier: Radial Basis Function (RBF); Partial Least Squares Discriminant Analysis (PLSDA). Optimization method: Enhanced Quantum-behaved Particle Swarm Optimization (EQPSO) | RBF: 100% PLSDA: 78,3% | RBF: 90% PLSDA: 75% |
03 | Infected cutaneous wound | E. coli, MSSA*, P. aeruginosa | In vitro (human) | 599/ 14 minutes | Classifier: Support Vector Machine (SVM) Optimization method: Principal Component Analysis (PCA); Independent Component Analysis (ICA); Orthogonal Signal Correction (OSC) | SVM: 100% | SVM: 97,85% |
04 | Infected cutaneous wound | MSSA*, MRSA**, S. pyogenes, E. coli, P. aeruginosa, C. perfringens | In vitro | 138/- | Classifier: Linear Discriminant Analysis (LDA) | -⦽ | LDA: 78% |
05 | Diabetic foot ulcer | MSSA*, S. epidermidis, P. aeruginosa | - | 2000/ | Classifier: Support Vector Machine (SVM) | Precision of basic parameters: Temperature: error − 0.9%; Humidity: error − 2.3%; SpO2: error − 5.8%; Pressure: error − 1.5% Gas sensors: Ethanol concentration: error − 8.7%; Aldehyde concentration − 12.8%; Sulfur compound concentration − 3.4% Prediction status: 2.8%. |
06 | Diabetic foot ulcer; Venous/arterial ulcer; Postoperative wound; Pressure injury; Traumatic wound; Sinus pilonidalis | - | In vitro (human) | 77/ 10 minutes | Classifier: Artificial Neural Networks (ANN) | -⦽ | _⦽ Sensitivity: 81% (95% CI 76–98); Specificity: 63% (95% CI 55–84) |
07 | Infected cutaneous wound | A. baumannii, E. coli, MSSA*, P. aeruginosa, E. cloacae | In vitro (mouse) | 804/ 14 minutes | Classifier: Support Vector Machine (SVM); Extreme Learning Machine (ELM); Domain correction and adaptive extreme learning machines (DC - AELM); Domain Adaptation Extreme Learning Machines (DAELM). Optimization method: Principal Component Analysis (PCA); Orthogonal Signal Correction (OSC); Domain regularized component analysis (DRCA); Subspace Alignment-Based Interference Suppression (SAIS) | -⦽ | ELM: 100% DC - AELM: 100% DAELM: 100% SVM: 99,01% Optimization: SAIS: 100% PCA (ELM): 100% PCA (SVM): 94,04% OSC (ELM): 96,97% DRCA: 95,46% |
08 | Pressure Injury | E. coli, P. aeruginosa, A. hydrophila, M. luteus, E. faecalis | In vitro | 100/ | - | +The standard deviation for each sensor used: Ammonia and amine (3.7 ppm); CO2 (0.46 ppm); Alcohol (15.1 g / L); Acetone (1.34 ppm). The distance at which the sensor loses its sensitivity under pure substances: Ammonia and amine (4 cm); CO2 (4 cm); Alcohol (7.6 cm); Acetone (4 cm). |
09 | Infected cutaneous wound | E. coli, P. aeruginosa, MSSA* | In vitro (mouse) | 48/ 14 minutes | Classifier: Least-squares support-vector machine (LS-SVM) | -⦽ | LS-SVM: 24h: 89,38% 48h: 83,66% |
*Methicillin Susceptible Staphylococcus aureus; **Methicillin-resistant Staphylococcus aureus; +Not reported results for the training and test set. ⦽Not reported precision values |
Main primary gas sensors used in eNoses applied to the diagnosis of infected wounds
By analyzing the most commonly used bacteria for the testing of eNoses in the sample studied, the main volatile products emitted by these microorganisms and the primary gas sensors to be used were verified for criteria for formulating new study protocols as presented in Table 4. Primary gas sensors are sensitive to typical metabolites of common bacteria in wound infections.
Table 4
Main gas sensors are sensitive to typical metabolites of common bacteria in wound infections.
Pathogens | Volatiles released by the pathogen | Main gas sensors suggested for pathogens |
| Isobutane, isovaleric acid, 2-methylbutanal, butanol, acetic acid | |
Staphylococcus aureus | TGS822 (alcohol and organic solvent), TGS823 (ethanol), TGS826 (ammonia and amines), TGS2602 (ammonia and hydrogen sulfide), TGS2600 (air pollutants), MQ135 (CO2), MQ138 (acetone) MQ3 (alcohol) |
Staphylococcus epidermidis | Acid and alcohol |
| Pyruvate, hydrogen cyanide, 1-undecene, 2-butanone, 4-methylquinazoline |
Pseudomonas aeruginosa |
Escherichia coli | Methanol, pentanol, ethyl acetate, propanol, butanol, acetone, propionic acid |
Assessment of the quality level of the studies included in the review
The methodological quality assessments of the included studies are summarized in Fig. 3, with a minimum score of 5 and a maximum score of 9 on a scale of 0 to 10. Three studies (ID05, ID07, and ID08) [31, 30, 32] were classified as level B, for scoring less than 6; these studies did not present a clear description of the results obtained in the eNose efficacy tests, such as sensitivity, specificity, and accuracy, and did not clearly explain the methods proposed for analyzing the databases developed.
In contrast, the other studies (ID01, ID02, ID03, ID04, ID06, and ID09) [27, 26, 28, 24, 25, 29] scored above 6 and were classified as level A, demonstrating the care taken by the researchers in the context of the study, as well as in defining the protocols of the experiments and presenting the results as clearly as possible.