Study design. In this multicentric study, absorbent skin patches for capturing the VOCs were developed at Technion, IIT, Israel. These patches were sent to the 2 collaborating centers (All India Institute of Medical Sciences (AIIMS), New Delhi, India, and Groote Schuur Hospital, Cape Town, South Africa.) for VOCs sampling from the study groups. Sample collection took place in Groote Schuur Hospital in Cape Town from August 2015 until November 2016. During April 2016 and June 2017, samples were collected at AIIMS hospital in New Delhi. All participants signed informed consent forms. The clinical trial received ethical approval by the Ethical Committee of the respective hospitals: AIIMS, New Delhi: IEC/NP-103/13.03.2015, RP-39/2015 and University of Cape Town: 307/2014. The study design included 3 groups at each site with 105 participants per group: confirmed pulmonary active TB cases, healthy volunteers and confirmed non-TB cases. The clinical classification referred to 2 gold standards: sputum culture on liquid medium [BACTEC Mycobacteria Growth Indicator Tube MGIT 960 System (MGIT 960)] and GexeXpert MTB/RIF. Moreover, all the participants were screened for HIV and QuantiFERON-TB Gold In-Tube (QFT-TB) tests for further evaluation on the effect of potential confounding factors. The participants were aged between 18-85 years and the following inclusion criteria were applied. Volunteers with a skin disease that was precluded at the sampling area were excluded from the studies. In addition, smoking within half an hour prior to testing was an additional exclusion criterion. The inclusion criteria for confirmed pulmonary active TB patients included: 1) clinical symptoms 2) positive microbiology (either a positive GeneXpert MTB/RIF or/and MGIT culture for M. tb; 3) newly diagnosed patients. For the non-TB patients, the inclusion criteria included 1) clinical symptoms; 2) negative culture result (for HIV infected and uninfected) or Negative GeneXpert test result (HIV uninfected only); 3) chest x-rays not supporting the diagnosis of active TB; 4) no clinical symptoms at follow-up at 8 weeks. The inclusion criterion for healthy controls were: 1) no clinical symptoms in the past 12 months. For the Indian site, all the samples were collected at a single location with the same staff. For the SA site, the sampling was done in three clinics within Cape Town city with the same staff. The study cohort was designed initially for sampling with PDMS in both body locations and Tenax in the chest area only. The decision to include Tenax sampling in the anterior arm area was made after the beginning of the sampling in the two clinical sites. Therefore, the number of Tenax-based samples from anterior arm area was lower than other sampling procedures. In addition, before statistical analysis, samples were excluded due to technical reasons such as broken vials during shipment, or during GCMS failures. For the exploratory study for the wearable device, similar inclusion and exclusion criteria were applied Riga, Latvia (Nr.12-A/19). The study included 18 confirmed pulmonary active TB patients and 29 healthy controls.
Preparation of poly(2,6-diphenylphenylene oxide)-based pouch as off-line sampling tool. 20/35 meshed poly(2,6-diphenylphenylene oxide) was used as an absorbing material. Poly(2,6-diphenylphenylene oxide) (134 mg) from Buchem BV, The Netherlands, was thermal conditioned at 300°C in a constant flow of pure nitrogen for 180 min. Poly(2,6-diphenylphenylene oxide) polymer was conditioned in a glass tube with conditioned glass wool at 240°C for 48 h. After conditioning it was stored in polyester (SAATI, PE AM 120.31 PW) meshed pouches (40.3 mm X 65.11 mm, mesh opening: 47µ), which had been cleaned with a solution of 5% Decon 90 (Decon Laboratories Limited) decontaminant in distilled water (18.2 MΩ) and later stored in a vacuum oven at 100°C for >15 h. The absorbing materials were stored in vials, closed, and wrapped with Parafilm at average temperature of 4ºC.
Sampling procedure. Each participant had to wear 2 poly(2,6-diphenylphenylene oxide) envelopes on an anterior arm area, which had to be analyzed by both GC/MS and a nanomaterial-based sensor array. In addition, room samples were collected for each participant to monitor the exogenous VOCs during skin sampling. The absorbent materials were placed on the skin after cleaning with an alcohol pad (saturated with >70% isopropyl alcohol) for 10 min before sampling. The absorbent materials were covered with aluminum foil and sealed with medical adhesive tape to avoid any VOCs absorption from the surrounding environment. No shaving procedure was done in order to avoid injure to the skin and change the VOC pattern. Sampling was done over 1h. A questionnaire was filled in for every participant and the absorbent material vial numbers were also recorded. The questionnaire included data regarding the main content of the last food and drink taken prior to sampling, hygiene, vaccinations, genetic diseases, chronic diseases, infectious diseases, family TB history, smoking and drinking habits, allergies, medications and vitamins, among other details. For the room samples, the poly(2,6-diphenylphenylene oxide) patch was placed on the table near the participant for 1h to be exposed to the room air. During the whole sampling process, participants wore face masks. After sampling, the absorbing materials were stored in vials, closed and wrapped with Parafilm. These samples were stored in a refrigerator in an average temperature of 4ºC up for maximum period time of 8 months. The air transportation of the samples was with the same conditions. Opening of the vails was done at biological hood in BSL2+ laboratory with the needed protection equipment. The disinfecting material was Oosafe® Surface Disinfectant (SparMED, Denmark) that do not contain alcohol and Bactericide (confirmed for M. tuberculosis). The manufacturer claim that these Disinfectants do not release VOCs. Prior the instrumental runs of the samples, the polyester pouches were cut, and the poly(2,6-diphenylphenylene oxide) power was transferred into a glass tube containing glass wool stopper. After the transfer, the second opening was closed manually with the glass wool as well.
Sample analysis with the Gas Chromatography−Mass Spectrometry (GC-MS). An analytical evaluation of the compounds absorbed on the poly(2,6-diphenylphenylene oxide) was done with a GC/MS-QP2010 instrument (Shimadzu Corporation). It was equipped with a SLB-5ms capillary column (with 5% phenyl methyl siloxane; 30 meters in length; 0.25 mm internal diameter; 0.5 mm thicknesses; purchased from Sigma-Aldrich), and was combined with a thermal desorption (TD) system (TD20; Shimadzu Corporation). Samples were analyzed by the GC-system in split mode (20%) at 30 cm/sec constant linear speed and in a 0.70 ml/min column flow. The following oven temperature profile was set: (a) 6 min at 40ºC; (b) 13ºC/min ramp up until 170ºC; (c) a hold-time 2 min; (d) 6ºC/min ramp up until 300ºC; and (e) 15 min at 300ºC. The run duration was 55 min in total. A mixture of alkane standard solution C8-C20 (Sigma-Aldrich) in hexane solvent was used as an external standard for GC/MS system calibration and normalization of the retention times and abundance changes as a result of a column aging. Compounds present in >80%54 of skin samples until 30 minutes of retention time due to compound release from Tenax and glass wool components. Sample chromatograms were further analyzed using an open source program OpenChrom® Community Edition, version 1.1, and custom codes using MATLAB version 9.5.0.944444 (R2018b). The chromatograms were converted into txt files with the following batch processing: 1. Denoising filter (M.Z. 73, 75, 28, 147, 207, 221, 281, 295, 335, 429); 2. Savitzky-Golay filter; 3. smoothed TIC baseline detector; 4. Peak detector first derivative (MSD); and 5. Combined integrator trapezoid. These analyses steps were programed in order to overcome retention time shifts due to a prolong study run and changes in the detector sensitivity. The statistical evaluation was based on adjusted p-value for multiple peaks evaluation using a non-parametric Kruskal-Wallis test and a Steel method in comparison to the TB group as a post hoc testing. Non-parametric tests as well as the ROC curve derived Youden’s cut-off point were run with JMP, version 14.0.0 (SAS Institute Inc., Cary, NC, USA, 1989−2005). For ethyl-cyclopropane analysis, 60 samples were excluded due to a prior close by saturated peak of IPA, leading to the following tested groups: (i) 85 confirmed pulmonary active TB patients; (ii) 182 non-TB patients with healthy controls; and (iii) 148 room samples.
Calibration curves of VOCs for GCMS. Identification and quantification of the VOCs that were found, involved the creation of a calibration curve for each candidate. The VOCs at different concentrations were generated using a commercial permeation/diffusion tube dilution (PDTD) system (Umwelttechnik MCZ, Germany). The system allows controlling the concentration of the VOCs. Purified dry nitrogen (99.9999%) from a commercial nitrogen generator (N-30, On Site Gas Systems, USA) equipped with a nitrogen purifier was used as a carrier gas. Samples were actively absorbed on poly(2,6-diphenylphenylene oxide) tubes at the same weight (134 mg) as used for the skin sampling from the PDTD system by pumping for 2.5 min at a flow-rate of 0.2 L/min. 3-5 repetitions were done per each concentration. The following concentration were generated: Toluene: 60,100,212,300,400,583,744 ppb; Acetic acid: 700, 900, 1100, 1300 ppb and for 2-ethyl-1hexanol: 6, 20, 40, 60, 80 ppb. The samples were analyzed by the same GC/MS methods, and a calibration curve was generated and compared to the abundance range of the clinical and room samples, using a weighted linear regression with errors in abundance.
Sample analysis with the nanomaterial-based sensor array. A stainless-steel cell for exposure contained an array of 40 nanomaterial-based sensors mounted on a customized polytetrafluoroethene circuit. The sensors included gold-nanoparticles (organically-stabilized spherical Au nanoparticles (core diameter: 3-4 nm), 2D random networks of single-walled carbon nanotubes (RN-SWCNTs), and polymers capped with different organic layers. For the ANN modeling, the following sensors proved to be key: (i) Au nanoparticles covered with octadecanethiol, decanethiol, tert-dodecanethiol, butanethiol, 2-ethylhexanethiol, dibutyl disulfide, 2-nitro-4-(trifluoromethyl) benzenethiol, benzylmercaptan, 4-chlorobenzenemethanethiol, 3-ethoxythiophenol and octadecylamine. (ii) Random networks (RNs) of carbon nanotubes (CNTs) with crystal hexa-perihexabenzocoronene (HBC) with C12 chemiresistor (HBC-C12). (iii) Polymer composites black carbon with poly(propylene-urethaneureaphenyl-disulfide) PPUU-2S chemiresistor and a composite of black carbon with poly(propylene-urethaneureaphenyl-disulfide) PPUU-2S mixed with poly(urethane-carboxyphenyl-disulfide) PUC-2S chemiresistor. Details regarding the fabrication and modification of the abovementioned sensors can be found in the literature55–58. The poly(2,6-diphenylphenylene oxide) samples were transferred prior to analysis into an empty thermal desorption (TD) tubes (Sigma-Aldrich), containing a glass wool stoppers, compatible with the TD system. The samples were thermally desorbed at 270°C for 10 min in an auto-sampler desorption system (TD20; Shimadzu Corporation, Japan). The sample was injected into the GC-system (Shimadzu Corporation, Japan) in a direct (splitless) mode at a constant 3 mL/min total flow and the desorbed sample was temporarily stored in a stainless steel column (150°C). The samples from the TD were then delivered by a 6-way Valco valve, equipped with 10 mL stainless steel loop (VICI, Valco Instruments Company Inc., USA) into a stainless-steel chamber containing the sensors with a volume of 330 cm3. When a one-way valve connecting the chamber to the column was open, the sample was sucked into the chamber, while the remaining volume was filled with N2 until reaching atmospheric pressure in the chamber. A Keithley 2701 DMM data acquisition/data-logging system was used to measure the resistance of all the sensors simultaneously as a function of time. The sensors’ baseline responses were recorded for 5 min in vacuum (~30 mtorr), 5 min under pure nitrogen (99.999%), 5 min in vacuum, and 5 min under sample exposure, followed by a further 3 min under vacuum conditions. To supervise the sensor’s functionality during the experiment, and to overcome possible sensor response drift, a fixed calibration gas mixture containing 11.5 ppm isopropyl alcohol, 2.8 ppm trimethylbenzene and 0.6 ppm 2-ethylhexanol was exposed to the sensors daily. This calibration gas was generated PDTD system. The calibration mixture was absorbed on a clean tube for 2 min. Several features are extracted from each of the sensor's signals upon exposure, including area under the curve, delta R peak, delta R middle and delta R end. The last 3 features are based on the difference between the baseline resistance, usually during vacuum, and the resistance during the response towards the exposure: peak point, middle part and the end part of the signal.
Artificial neural networks. To validate a multilayer perceptron (MLP) using k-fold cross-validation, the verification dataset was used to test performance. This dataset is not involved in the weight modification process, which enables this validation method. To carry it out, the global database containing samples from India and South Africa was divided randomly into k parts equal in size. The k selected was 6, and, therefore the MLP was evaluated 6 times by swapping the verification dataset for a new one in each test. In essence, for every test ~83% of training samples and ~17% of verification samples were used, and the final statistical performance of the model was evaluated by averaging the results from all k tests, which covered every sample in the initial database. When a model provides accurate results during this validation method, it typically signifies that it can generalize well; and therefore, it is reliable for data that is different from the one used in the training or verification datasets59,60. The model utilized 20 sensors with a total of 46 features. Feature selection and MLP-related calculations relied on MATLAB version 9.3.0.713579 (R2017b).
Discriminant function analysis (DFA) DFA is a statistical method for data analysis when the groups to be discriminated are defined (labeled) before analysis61. The input variables are the features extracted from sensors’ responses towards the skin samples. The decision on either linear or quadratic model was based on homogeneity of the variance-covariance matrices of the tested groups according to statistical tests, e.g. Bartlett's62,63. During this study equal prior probabilities were set to confirmed pulmonary active TB and non-TB with healthy volunteers, respectively. For the off-line approach, the model was evaluated by randomly splitting the original database into 70% training set and 30% test set. For the on-line device system, a linear DFA model was applied based on three sensor features with leave-one-out validation.
Wearable device. The device was connected to the PC and the developed software was launched. Initially, the device samples room air for 10 min. Then the patch was attached to the participant. The patch was attached for 60 min. There was no direct contact between the skin and the sensors. The device included eight sensors as described in previous method sections. On exposure, several features, such as the Area under curve, Delta R peak, Delta R middle and Delta R end, are extracted from the sensor signal. The last three characteristics are based on the difference between the resistance of the baseline, typically during the room air exposure, and the resistance in the exposure response: peak point, middle and end part, respectively. The obtained a linear DFA model was based on three features from three sensors.
Data availability
The authors declare that the data supporting the findings of this study are available within the paper and its supplementary information file. Any additional data if needed will be provided upon reasonable request.
Code availability
Custom codes are available on request.