Evaluating the VOCORDER device for early disease detection through breath analysis: study protocol for a two-phase clinical study Authors: Flora Kontopidou, Stefan James Swift, Christos Kokkotis, Serafeim Moustakidis, Ioannis Kavouras, Anastasia Nefeli Vidaki, Sophia Brebou, Vasileios Zavvos, Angeliki Panagiotou, Christos Iatrou, Dimitrios Grosomanidis, Zoe Zacharouli, Stamatios Giannoukos Published: 2026 ABSTRACT AbstractIntroductionThe survival rate of patients with life-threatening diseases primarily depends on the speed of diagnosis. Too often, diseases are detected only after symptoms appear, which usually occurs at later stages of a disease when available treatments may be less effective. Current detection techniques primarily depend on identifying metabolites in biofluids such as blood and urine. The analysis of these fluids is typically performed in laboratories, resulting in lengthy waiting times for patients to receive their results. In severe cases, invasive biopsies and radiative methods are used to diagnose conditions such as cancer. These biopsies can cause distress for patients who are already experiencing significant emotional or physical stress, while imaging techniques involving ionising radiation may pose additional health risks. Additionally, these methods can be costly. In recent years, exhaled breath has become a biofluid matrix of interest for disease detection, allowing for the identification of volatile organic compounds (VOCs) or VOC profiles associated with specific conditions. To improve early disease detection through breath analysis, the VOCORDER project aims to develop a device that provides a fast, simple, user-friendly and cost-effective method for continuous health monitoring to identify diseases in their early stages before symptoms appear.Methods and analysesA literature review was initially conducted to identify five reference diseases of interest (lung cancer, stomach/colon cancer, breast cancer and kidney insufficiency) and previously reported VOC profiles associated with these diseases. In this trial, the project team from the MITERA Hospital will select patients, and the hospital staff will conduct personal interviews with these subjects. Each participant will also complete a questionnaire for the acquisition of demographic and medical history data, after being informed in detail about the purposes of the questionnaire and signing a consent form. The study protocol consists of two phases. Phase 1 is a baseline study designed to detect and identify breath biomarkers for the early diagnosis of the diseases mentioned above using gas chromatography-mass spectrometry (GC-MS) and secondary electrospray ionisation high-resolution mass spectrometry (SESI-HR-MS). Prescreening will select 120 healthy controls and 175 patients for the baseline phase of the clinical trial, for which breath samples will be collected in 1 L Supel-Inert Multi-Layer Foil gas sampling bags. New biomarkers and VOC profiles will be extracted from these data, and further statistical analysis will allow for artificial intelligence (AI) models to be produced and tested. For phase 2 (validation phase), 120 healthy controls and 100 patients will be selected. Breath samples will again be collected in 1 L gas sampling bags for analyses with GC-MS and SESI-HR-MS. The VOCORDER device will also be used, and its functioning with the newly developed AI models will be evaluated.Ethics and disseminationThis clinical study has been approved by the scientific council at the MITERA hospital in Athens, Greece (#513/2024). The outcomes will be disseminated through peer-reviewed journal publications and presentations at scientific conferences.Trial registration numberNCT06711939.Keywords: Clinical Trial, Methods, Biochemistry, Breast tumours Strengths and limitations of this study The development of the VOCORDER device will offer a non-invasive online breath analysis technique in the clinical setting. The development of the VOCORDER device will offer a non-invasive online breath analysis technique in the clinical setting. The instrumentation used for offline analysis during the baseline and validation phases includes gas chromatography-mass spectrometry (GC-MS) and secondary electrospray ionisation high-resolution mass spectrometry (SESI-HR-MS). The SESI-HR-MS offers high resolution with minimal fragmentation (softer than ESI) and is user-friendly. However, this technique is semi-quantitative. The GC-MS provides orthogonal separation of sample molecules based on retention time and mass-to-charge ratio (m/z), enabling compound identification using reference spectral libraries. The instrumentation used for offline analysis during the baseline and validation phases includes gas chromatography-mass spectrometry (GC-MS) and secondary electrospray ionisation high-resolution mass spectrometry (SESI-HR-MS). The SESI-HR-MS offers high resolution with minimal fragmentation (softer than ESI) and is user-friendly. However, this technique is semi-quantitative. The GC-MS provides orthogonal separation of sample molecules based on retention time and mass-to-charge ratio (m/z), enabling compound identification using reference spectral libraries. In the baseline phase, the use of reference methods (GC-MS and SESI-HR-MS) may enable the identification of novel breath biomarkers associated with the diseases under investigation. In the baseline phase, the use of reference methods (GC-MS and SESI-HR-MS) may enable the identification of novel breath biomarkers associated with the diseases under investigation. The study will not use the same cohort of patients for the baseline and validation phases due to the elapsed time between these sampling periods. Patients may either be cured, have progressed to a new stage of the disease, or may no longer be available for sampling. The study will not use the same cohort of patients for the baseline and validation phases due to the elapsed time between these sampling periods. Patients may either be cured, have progressed to a new stage of the disease, or may no longer be available for sampling. Introduction Each year, millions of premature deaths worldwide result from a lack of early disease diagnosis1 for conditions including heart disease, sepsis, Alzheimer’s, chronic obstructive pulmonary disease, diabetes and cancers.1,3 While treatments are available for these diseases to extend and improve a patient’s life, early detection is vital for increasing the survival rate.2 4 Unfortunately, symptoms often appear in the more advanced stages of a disease,3 5 when treatment success rates are significantly lower.6 7 Modern diagnostic methods are costly and time-consuming and must be performed in specialised laboratories.8,10 They cannot be repeated frequently within short time intervals, and this applies to both invasive and non-invasive tests. Moreover, a significant number of these methods, particularly in the context of preventive care, remain inaccessible to the general population.11,13 Therefore, it is essential to develop accurate, non-invasive diagnostic tools that can be routinely used in the clinical setting to improve patient survival rates. The onset of disease disrupts the body’s biochemistry and metabolic processes, leading to the production of metabolites that are excreted in biofluids such as urine, sweat and breath.114,21 These metabolic changes can occur even before symptoms appear, providing a basis for early disease detection.114 16 22,25 Volatile organic compounds (VOCs), a subset of these metabolites, pass from the bloodstream into the lungs via the alveoli and are exhaled in breath.2 26 Analysing these VOCs using gas-phase analytical techniques can reveal biomarkers linked to specific metabolic processes, offering insights into the presence and progression of disease.3 10 14 25 26 Human exhaled breath is a complex sample containing thousands of VOCs within a matrix of oxygen, nitrogen, carbon dioxide and water.727,32 In recent years, this biological fluid has become the focus of an emerging field of analytical research known as breath metabolomics.27 33 Breath sampling can be performed non-intrusively814 16 27 29 34,37 on patients and exhibits a chemical complexity that enables the potential identification of unique VOC signatures among individuals.3 14 21 31 37 Within these VOC profiles, certain VOCs have been reported as biomarkers for specific diseases, and their detection may indicate the presence of a disease in an individual.7 8 26 29 30 38 Several studies have effectively investigated breath analysis using advanced mass spectrometric techniques to assess emitted VOCs in the gas phase. These techniques, including selected ion flow tube mass spectrometry (SIFT-MS),39,46 proton transfer reaction MS,47,51 secondary electrospray ionisation high-resolution MS (SESI-HR-MS)52,59 and gas chromatography-MS (GC-MS),60,67 offer highly sensitive and accurate detection capabilities. Their potential to provide detailed molecular insights makes them valuable tools for breath analysis, with ongoing advancements (lower limits of detection and quantification, fast response times) further improving their efficiency and applicability for a variety of research and clinical purposes. Other techniques have also been exploited and tested for use in breath analysis. E-sensors,68 69 for example, can detect species such as CO, NO, ethanol and ammonia.1 7 They are made up of a series of sensors which mimic human olfactory systems1 3 37 and are size and cost-effective1 3 14 22 37 70 although they exhibit low sensitivities7 and lack physical and chemical control, resulting in reproducibility issues.3 16 Wearable sensors and optical methods have also been used in breath analysis research.7 22 29 37 While breath analysis holds significant promise for non-invasive early disease detection, the complexity of VOC profiles arising from metabolic and environmental variability presents challenges for accurate interpretation.71,75 Several factors influence an individual’s VOC profile, including both endogenous factors (such as underlying genetic conditions) and exogenous sources (such as exposure to air pollution, smoking and alcohol).7 26 27 As a result, the quality and accuracy of data for disease detection through breath analysis remain challenging. Thus, identifying VOC profiles and patterns provides a more robust approach to disease detection in breath analysis, rather than relying solely on the presence of individual biomarker species. This project aims to develop a user-friendly breath analysis tool for regular health monitoring in various settings. The goal is to create a device that seamlessly integrates into the healthcare system for routine use, demonstrating effectiveness and accuracy in clinical environments. An ideal breath analysis device would offer a quick, simple, affordable, reliable and precise method for ongoing assessments, with the capability to detect diseases at their earliest stages, even before symptoms arise, through the use of artificial intelligence (AI). Such a device could significantly improve patient survival rates, reduce emotional stress on families and lower costs for government-funded healthcare services. To achieve this vision, the European Union-funded VOCORDER project is advancing a new breath analysis technology called ‘VOCORDER’. This system employs a laser-based approach to analyse breath, simultaneously detecting small quantities of gases at specific concentrations, enabling accurate and early disease identification. The study is a single-centre clinical trial conducted in two phases. Phase 1 is a baseline study focused on detecting and identifying breath biomarkers for the early diagnosis of lung, stomach, colon and breast cancers and kidney insufficiency. This phase uses SESI-HR-MS and GC-MS to analyse breath samples from 120 healthy participants and 175 patients (at least 35 per disease group). Phase 2 aims to validate the newly developed portable breath analyser in a cohort of 120 healthy controls and 100 patients with the aforementioned diseases. Breath samples will also be collected in 1 L gas sampling bags at MITERA Hospital in Athens, Greece, and will be analysed in the breath analysis laboratory at ETH Zürich in Switzerland, using SESI-HR-MS. This manuscript outlines the clinical trial protocol for developing and validating the VOCORDER device, designed for continuous health monitoring and early disease detection. It also describes the methods used to identify new biomarker species and aims to expand the library of breath-associated biomarkers linked to the selected diseases focused on in this work. Hypothesis and objectives Hypothesis It is hypothesised that exhaled breath VOCs and their concentrations will show patterns that can be correlated to specific diseases under certain conditions compared with healthy controls. We aim to develop, test and validate an analyser that enables in situ concentration measurement of VOCs in separate breath samples from participants, featuring a user-friendly and intuitive interface. Primary objectives The primary objectives are to develop and validate a user-friendly breath analysis device (VOCORDER) for lifelong health monitoring and predictive medicine. The project aims to ensure that the device is easily integrated into daily life, minimises false positives and meets clinical standards. It includes validating the device’s accuracy in distinguishing between patient VOC profiles and those of healthy individuals, correlating breath VOCs with electronic medical records, benchmarking against established methods and exploring links between breath-detected gases and specific diseases. Methods and analysis The VOCORDER project is a single-centre, observational study aimed at deciphering correlations between VOC profiles from a human breath matrix and the presence of a specific illness, with the goal of engineering a new method for disease identification and early detection through breath analysis. The project comprises a consortium of 12 contributing partners, including Mitera Hospital, Alpes Lasers SA, Eidgenössische Materialprüfungs- und Forschungsanstalt (EMPA), UAB Metis Baltic, Neuraltech IKE, Vrije Universiteit Brussel, EULAMBIA Advanced Technologies, Cailabs, Argos Messtechnik, AiDEAS OU, Institute of Communications and Computer Systems (Athens, Greece) and ETH Zürich (Switzerland). Study design The clinical trial is divided into two phases: the baseline (October 2024–May 2025) and validation (May–December 2026) and will take place at Mitera Hospital in Athens, Greece. Patients and healthy controls will be selected and enrolled through a prescreening process that includes a questionnaire and a personal interview. The staff at Mitera will enter the information from participant questionnaires and interviews into electronic records. Breath samples will be collected and stored in 1 L Supel-Inert Multi-Layer Foil gas sampling bags. These samples will then be analysed offline using GC-MS and SESI-HR-MS during both the baseline and validation phases and with the newly developed VOCORDER device introduced in the validation phase. The baseline phase will be used to identify the optimal operating parameters of the VOCORDER device in preparation for the validation phase. Using the selected diseases as the foundation for the statistical models, we will identify correlations in the breath matrix and new biomarkers for these diseases. The selection of these diseases was based on a literature review conducted by consortium members at ETH Zürich. AI algorithms will be developed using electronic medical records and mass spectral data obtained from offline analyses. The resulting AI models will then be used in the validation phase to assess the performance of the VOCORDER device. By using the gold-standard reference methods of GC-MS and SESI-HR-MS, the baseline phase presents significant opportunities to explore and identify new breath biomarker species potentially linked to a specific disease in its early stages. It also facilitates the testing of prototype AI models in preparation for the validation phase, during which the VOCORDER device will be assessed alongside the gold-standard techniques. Figure 1 provides a general overview of the study design. Overview of the study protocol for the VOCORDER clinical trial. AI, artificial intelligence; ATH, Athens; GC-MS, gas chromatography-mass spectrometry; SESI-HR-MS, secondary electrospray ionisation high-resolution mass spectrometry; ZRH, Zurich. Regular reports on ethics risk monitoring will be provided during project meetings. The formulation and distribution of an ethics questionnaire internally to the partners involved in the technical advancement is under consideration to ensure alignment with the ethical requirements. The internal ethics committee will be responsible for addressing any questions raised by the project partners at all times and cooperating with the respective representatives of legal and ethical matters defined by each entity. VOCORDER will assess its outcomes at all phases of the project implementation against these requirements to ensure comprehensive coverage. Sampling Sampling will take place at MITERA Hospital in Athens, Greece, which has been operating for over 45 years. MITERA is a private hospital that offers general, paediatric and maternity/gynaecological clinics. It is the leading healthcare centre in Greece, specialising in paediatric medicine, although it provides care for individuals of all ages. The hospital staff and the research team will conduct sampling at MITERA Hospital. For the healthy control cohort, a minimum of 120 participants will be involved in both the baseline and validation phases. A breath sample will be collected from each participant on the sampling day at the same time. The same control group is expected to be involved in both the baseline and validation phases of the study. For the patient cohort, 175 patients will participate in the baseline phase and 100 patients will participate in the validation phase. This represents 35 and 20 patients for each disease during the respective phases of the study. A breath sample will be collected from each subject on different days, although they will be taken at the same time of day. The patients in the baseline and validation phases will predominantly consist of different individuals. This is due to the time elapsed between these two sampling stages (approximately 1 to 2 years). Patients may either be cured, progress to a different stage of the disease or become unavailable for sampling. Both patients and healthy controls will provide breath samples in 1 L gas sampling bags, filling them up to 80% capacity. GC-MS (Agilent Technology, Santa Clara, California, USA) and SESI-HR-MS (Fossil Ion Technology, SL, Q Exactive Orbitrap, Thermo Fisher Scientific) equipment represent the analytical gold standard reference methods for VOC identification and quantification available for this study and have been widely utilised in breath analysis studies.52,67 In the baseline phase, these mass spectrometric techniques will be employed to analyse the samples. During the validation phase, the samples will also be processed using the newly developed VOCORDER device. The sample and data analysts will be blinded to the participant’s group status (healthy controls or patients). Participants and eligibility criteria Healthy controls A minimum of 240 healthy control participants will be recruited for both the baseline phase (120 participants) and the validation phase (120 participants) of the study. Participants will be selected from both genders (as assigned at birth), must be over 18 years old, and must not currently suffer from any of the diseases under investigation in this study. Healthy controls must also be available to participate in both phases of the clinical trial, provided they remain healthy throughout the study. Patients 55 patients will be recruited for each disease, totalling 275 patients, across both genders (as assigned at birth) and aged 18 years or older. A total of 175 patients will be required for the baseline phase and 100 patients for the validation phase. Subjects will be selected based on the following inclusion and exclusion criteria. Inclusion criteria The criteria for inclusion for both phases are as follows: Patients must be over 18 years old. Patients must be over 18 years old. Patients should be able to communicate and understand the information given by the staff at Mitera Hospital. Patients should be able to communicate and understand the information given by the staff at Mitera Hospital. Participants must suffer from only one of the conditions selected for investigation in this study. Participants must suffer from only one of the conditions selected for investigation in this study. The disease must be detected as soon as possible; no treatment can be started for the patient yet (essential). The disease must be detected as soon as possible; no treatment can be started for the patient yet (essential). There must be no other severe acute or chronic disease that could affect the measurements (desirable). Patients will be selected in collaboration with the clinicians responsible for the treatment and/or hospitalisation of patients with the diseases under investigation. There must be no other severe acute or chronic disease that could affect the measurements (desirable). Patients will be selected in collaboration with the clinicians responsible for the treatment and/or hospitalisation of patients with the diseases under investigation. Further disease-specific requirements for enrolling patients in the study are outlined below. For patients with cancer, the type of cancer must be determined based on histopathological data and specific result indicators. The disease stage must also be established through haematological and radiological examinations. For kidney insufficiency, proof of chronic kidney disease is required, including the establishment of the disease stage as indicated by the case fatality rates, along with any evidence of haemodialysis treatment. This medical evidence must be available for successful enrolment. Exclusion criteria The exclusion criteria for both phases include: The lack of signed consent. The lack of signed consent. The inability to cooperate with project staff under any circumstances. The inability to cooperate with project staff under any circumstances. Failure to adhere to the necessary instructions and requirements of the sampling procedure. Failure to adhere to the necessary instructions and requirements of the sampling procedure. Inability to provide the necessary breath sample. Inability to provide the necessary breath sample. Currently undergoing any form of treatment or intervention (such as a specific diet or surgery) between breath sample acquisitions. Currently undergoing any form of treatment or intervention (such as a specific diet or surgery) between breath sample acquisitions. Description of the collected data The factors associated with the incidence of disease represent a complex area. An extensive evaluation of factors for various health conditions reveals an intricate interplay among genetic predisposition, environmental exposures, lifestyle choices and underlying health issues. The questionnaire developed for the VOCORDER project aims to identify potential factors that may influence the outcomes of analysing the five diseases of interest in patients. A summary of these risk factors is displayed in table 1. The risk factors may be either endogenous to the individual, such as their genetic makeup, or exogenous, concerning exposure to toxic substances. By identifying the risk factors associated with each patient, the presence of certain VOCs in breath may be clarified, enhancing the data quality assurance of the VOC profiles linked to specific diseases. Table 1Summary table of potential factors that may influence the pattern of breath samples for the five diseases studied in the VOCORDER clinical trialDiseaseExposure factorsBiological factorsLung cancer SmokingRadonAsbestosAir pollutionMetalsOccupational exposure Family history Gastric/colon cancer DietSmoking Helicobacter pylori Family history Breast cancer Hormone therapyAlcohol consumptionObesityPhysical inactivity Early menarcheLate menopauseGene mutations (eg, in the BRCA1 and BRCA2 gene)Reproductive history (breastfeeding, age at first childbirth) Kidney insufficiency Chronic use of onsteroidal Anti-Inflammatory Drugs (NSAIDs)Contrast dyes DiabetesHypertensionFamily historyAgeEthnicity NSAIDs, Nonsteroidal Anti-Inflammatory Drugs. Summary table of potential factors that may influence the pattern of breath samples for the five diseases studied in the VOCORDER clinical trial Smoking Smoking Radon Radon Asbestos Asbestos Air pollution Air pollution Metals Metals Occupational exposure Occupational exposure Family history Family history Diet Diet Smoking Smoking Helicobacter pylori Helicobacter pylori Family history Family history Hormone therapy Hormone therapy Alcohol consumption Alcohol consumption Obesity Obesity Physical inactivity Physical inactivity Early menarche Early menarche Late menopause Late menopause Gene mutations (eg, in the BRCA1 and BRCA2 gene) Gene mutations (eg, in the BRCA1 and BRCA2 gene) Reproductive history (breastfeeding, age at first childbirth) Reproductive history (breastfeeding, age at first childbirth) Chronic use of onsteroidal Anti-Inflammatory Drugs (NSAIDs) Chronic use of onsteroidal Anti-Inflammatory Drugs (NSAIDs) Contrast dyes Contrast dyes Diabetes Diabetes Hypertension Hypertension Family history Family history Age Age Ethnicity Ethnicity NSAIDs, Nonsteroidal Anti-Inflammatory Drugs. As part of the prescreening process, a comprehensive questionnaire will collect demographic and general health information from participants. This questionnaire will include inquiries about a patient’s geographical location, marital status, occupation, education level, income, race and ethnicity, gender and age. It will also cover their medical history, comorbidities, physical activity levels, diet, alcohol consumption, smoking habits, height, weight, body mass index (BMI) and mental health. Additionally, it will gather information on hospitalisations, surgeries, medication use, and the presence of infections or chronic diseases. Furthermore, questions regarding occupational exposure, stress levels, sleep patterns, genetic predispositions, and other lifestyle and environmental influences will be included. These inquiries will explore a patient’s risk factors, and the responses will be integrated into statistical models during the data analysis stage. The collected responses will also help explain and identify specific VOC patterns exhibited by a subject and assist in pinpointing which endogenous and metabolic processes lead to a particular VOC profile for a specific disease. Recruitment and screening procedure The recruitment process is as follows: The attending physician, based on predefined criteria, determines whether the patient is eligible to participate in the study and provides them with initial information. If the patient agrees to participate, the study team takes over, offering a detailed explanation and obtaining written informed consent (online supplemental appendix 1). This is followed by an interview and data collection. The healthy controls will consist of either hospital staff or patients at the Mitera Hospital who have been admitted under different circumstances. For the baseline phase, 120 healthy controls and 175 patients are needed. In the validation phase, 120 healthy controls and 100 new patients will be recruited, respectively. Both the healthy controls and the patients considered for this study will complete a comprehensive questionnaire. This questionnaire will collect information on medical history (hospitalisations, surgeries, medications, infections, chronic diseases, comorbidities, physical activity, dietary habits, alcohol consumption, smoking habits, BMI, height and weight) as well as demographic data (ethnicity, gender, age, marital status, occupation, education level). Additionally, questions regarding lifestyle, environmental exposure, genetic predispositions and psychological factors will also be included (table 2). The study aims to recruit an equal number of male and female participants. Age, sex, BMI and smoking status will be recorded for all participants via the study questionnaire and included as covariates in all statistical models. Propensity-score-based sensitivity analyses will be used to assess the robustness of VOC-disease associations to any residual demographic imbalance between groups. Table 2Type of collected data from healthy individuals and patients enrolled in the studyType of collected dataHealthy individualsPatientsDemographicsXXFactors which influence an individual’s VOC profile LifestyleXX ComorbiditiesXX DrugsXX Family medical historyXX Other substancesXX Previous hospitalisationXXCurrent hospitalisationX Screening examination tests (haematological test, X-ray, endoscopy, mammogram)XX Indicators related to the study diseasesXXVOC, volatile organic compound. Type of collected data from healthy individuals and patients enrolled in the study VOC, volatile organic compound. Outcomes Primary outcomes To evaluate the effectiveness of the VOCORDER device in the clinical setting and to encourage further research into breath analysis. To evaluate the effectiveness of the VOCORDER device in the clinical setting and to encourage further research into breath analysis. Secondary outcomes To determine and describe the operational features of the new VOCORDER technology for breath analysis and health monitoring in a clinical setting. To determine and describe the operational features of the new VOCORDER technology for breath analysis and health monitoring in a clinical setting. To evaluate the results, performance and data quality of the VOCORDER device and compare these results with comparable data from the reference methods, GC-MS and SESI-HR-MS. To evaluate the results, performance and data quality of the VOCORDER device and compare these results with comparable data from the reference methods, GC-MS and SESI-HR-MS. To investigate the correlations and statistical significance of noteworthy VOC markers or profiles in subjects with known pre-existing diseases, aiming to identify the significant biomarkers associated with the selected diseases in this study. To investigate the correlations and statistical significance of noteworthy VOC markers or profiles in subjects with known pre-existing diseases, aiming to identify the significant biomarkers associated with the selected diseases in this study. To evaluate the potential use of the VOCORDER device for early disease detection from the human breath matrix, allowing for critical early patient diagnosis of disease by identifying correlations of VOC biomarkers with important medical parameters. To evaluate the potential use of the VOCORDER device for early disease detection from the human breath matrix, allowing for critical early patient diagnosis of disease by identifying correlations of VOC biomarkers with important medical parameters. Instrumentation Secondary electrospray ionisation high-resolution mass spectrometry The SESI-HR-MS technique was specifically developed for use in breath analysis studies and has been extensively used in previous research investigating the human exhalome.52,59 Its main advantages include being non-invasive, operating in real time and producing high-resolution full-scan spectra, reaching up to approximately 1000 Da. This allows for the potential identification of new specific biomarkers of disease.76,78 In the study detailed here, the breath sample is stored in a gas sampling bag that connects to the SESI-HR-MS inlet. This setup allows the sample to flow through a heated capillary into the ionisation chamber. A high voltage is applied to a 0.1% formic acid aqueous solution, creating an ionising electrospray. The stability of this electrospray is continuously monitored using a nano-ammeter to ensure a consistent signal during the analysis. The mass spectrometer is calibrated at least weekly in both ionisation modes, which is essential for preserving the system’s high sensitivity and selectivity. Furthermore, the stability of the system is evaluated with gas standards of breath VOCs, such as acetone, and a labelled compound (butyric acid-d7), to detect any deviations on a daily basis. Specifically, our lab monitors the stability of the SESI-HR-MS system three times daily during the last 2 years using a gas standard of butyric acid-d7. In addition, a microquantity of acetone-d6 standard is introduced in the bags as an internal standard. In the ionisation chamber, ion-molecule reactions occur between electrospray droplets and sample VOC, resulting in the VOC becoming charged and detectable by an Orbitrap detector. In the positive mode, the dominant ion produced is the MH+ ion (proton transfer species), while in negative mode, the (M-H, M minus H)- ion (hydride transfer species) prevails. However, sufficient ionisation energy is necessary to produce these ionised species. However, selected ions from the Orbitrap may undergo collision-induced dissociation through tandem MS (MS2) analysis, enhancing specificity and accuracy in compound identification, including the ability to measure isomers separately. The molecular formula of a species can be determined using high mass accuracy and resolution (to five decimal places of m/z) as well as isotopic abundance ratios. The limits of detection can reach down to sub-ppt levels. The heated capillary and ionisation chamber for sample injection prevent the loss of VOC to the walls of the transfer line through condensation, particularly for species with higher molecular masses, which are common in biologically significant and disease-specific metabolites found in breath. The larger the identified metabolite, the more likely it is that the single specific compound will be unique to the prevalence of a particular disease. The SESI-HR-MS can measure up to 1000 Da (with a resolution of 140 000), potentially enabling the discovery of new metabolites related to the diseases of interest. Additionally, the heated system eliminates any delays in responses to concentrations. For the breath samples collected during the VOCORDER clinical trial described in this publication, the mass range used for spectral analysis is 50–500 m/z. The instrument has a resolution of 140 000 with an automatic gain control target of 5×106; a maximum injection time of 200 ms; a sheath gas flow rate of 15 a.u. (arbitrary unit); an auxiliary gas flow rate of 2 a.u.; an electrospray voltage of 3.5 kV; a capillary temperature of 250°C; an S-lens RF amplitude of 50.0; and measurements taken in both positive and negative modes. Gas chromatography-mass spectrometry The breath samples will also be analysed using a GC-MS system to validate and enhance the VOC analysis of the SESI-HR-MS technique. GC-MS is an orthogonal technique and a standard reference method previously employed in breath analysis research. For the offline analysis of samples from this clinical trial, a handheld pump (ACTI-VOC low-flow pump kit, Markes International Ltd, UK) will be used to transfer breath VOCs from the gas sampling bag to preconditioned hydrophobic stainless-steel tubes (Markes International, UK). These tubes are sealed and stored in the refrigerator until they are transported to the laboratory for analysis. A standard thermal desorption system will be used to extract the VOCs from the samples into the GC-MS system (Agilent GC-7890B coupled to an Agilent MSD-5977B, Agilent Technology, Santa Clara, California, USA). Calibration methods for offline techniques Calibration methods for SESI-HR-MS and GC-MS rely on external gas standards for quantification. Three calibration systems are used to generate trace VOC gas standards across a wide concentration range (ppt–ppm). The first is a commercial gas cylinder-based system that mixes VOC standards with humidified air to simulate breath, with concentrations controlled by nitrogen dilution. The second, an in-house system at ETH Zürich, injects liquid standards into evaporation chambers, applying Henry’s Law to calculate gas concentrations before mixing with humidified nitrogen. The third system modifies a commercial permeation tube unit (OVG-4, Owlstone Medical Ltd, UK) to adjust VOC concentrations via nitrogen flows. These methods offer flexibility but require careful handling to minimise artefacts from VOC interactions. The VOCORDER device The VOCORDER device is based on IR measurements using a self-mixing quantum cascade laser mid-IR array-based analyser capable of measuring VOCs at concentrations ranging from 10 to 100 ppb. It offers a resolution of approximately ±1 ppb and a limit of detection of 1 ppb. By using the National Institute of Standards and Technology mass spectral library as a reference, the device can identify and quantify VOCs. Data analysis procedure Based on a literature review, we will investigate lung, gastric, colon and breast cancer and kidney insufficiency. According to our literature search, a subject’s breath profile changes depending on the incidence of disease, which is qualitatively summarised in table 3. The selection of biomarkers was also adapted to the detection capabilities of the developed technology. Different diseases may lead to varying concentrations of the selected exhaled breath VOC and, subsequently, different profiles. Therefore, these compounds will be used to test the VOCORDER device within the clinical environment, and the selection of biomarkers to be measured was also adapted to the technical capabilities of the LASER methodology. Table 3Summary of VOC biomarker profiles from breath for five selected diseases to be investigated as part of the VOCORDER clinical trialLung cancerGastric/colon cancerBreast cancerKidney insufficiencyAcetoneXXXAmmoniaXEthanolXXEthyleneXXXIsopreneXXXXMethanolXXVOC, volatile organic compound. Summary of VOC biomarker profiles from breath for five selected diseases to be investigated as part of the VOCORDER clinical trial VOC, volatile organic compound. With this prior knowledge, a statistical analysis of the GC-MS and SESI-HR-MS breath sample spectral data will be conducted alongside clinical data obtained from questionnaires and personal interviews with patients. Utilising the diseases of interest and the previous understanding of VOCs associated with these conditions (table 3), AI techniques will differentiate between sample categories based on the participants’ health status. AI techniques, including semi-supervised, supervised and unsupervised learning methods, will draw on established machine learning approaches from the existing literature to support the exploration of statistical dependencies and correlations among breath matrices, spectral patterns, disease parameters and potential new biomarkers identified during the baseline phase. Techniques may include advanced classification methods such as decision trees and more complex models like Convolutional Neural Networks. XAI technologies will also be employed to improve model interpretability and enhance the overall accuracy of the statistical analyses. The prespecified AI training pipeline will be structured as follows. For each disease group, binary classifiers will be developed independently (disease vs healthy controls), with a secondary multi-class ensemble explored using a one-vs-rest strategy. This staged approach constrains model complexity relative to the available sample size. Feature selection will be based on literature-supported biomarker relevance combined with univariate statistical filtering (Mann-Whitney U test with Benjamini-Hochberg false discovery rate correction, q<0.05), applied exclusively to training data within each cross-validation fold to prevent data leakage. A stratified fivefold nested cross-validation scheme will be employed, with the outer loop providing an unbiased estimate of generalisation performance and the inner loop used for hyperparameter optimisation. An additional independent hold-out set (20% of each cohort) will be withheld prior to any analysis and used solely as the final performance benchmark. AI training and validation safeguards are built into the pipeline at multiple levels. First, nested cross-validation (five outer folds, three inner folds) ensures strict separation between model selection and performance estimation, preventing data leakage. Second, the adversarial deconfounding architecture structurally penalises retention of demographic information in the learnt representation, providing a formal guard against confound-driven bias. Third, all trained models, hyperparameter configurations and random seeds will be archived in a versioned GitHub repository (DOI-registered via Zenodo) to ensure full reproducibility. Fourth, SHAP (SHapley Additive exPlanations) values will be computed for each prediction to provide clinician-interpretable feature attribution and support regulatory auditability. All preprocessing steps—including imputation, scaling and feature selection—will be fitted exclusively on training folds and applied to test folds without exception. Class imbalance between disease patients and healthy controls will be addressed using class-weighted loss functions and stratified sampling during training. Model performance will be reported using the full confusion matrix, AUC-ROC, balanced accuracy, sensitivity, specificity, positive predictive value, negative predictive value and calibration curves. SHAP values will be computed for all models to identify the VOC features driving classification decisions and to provide clinical interpretability. Sample size calculation The sample size calculation was based on the objective of diagnosing each disease using binary machine learning classifiers (disease vs healthy controls). A specificity and sensitivity of over 90% are required to differentiate between healthy controls and each specific disease of interest. Using a one-proportion z-test, it was determined that at least 35 subjects would be necessary per disease group in the baseline phase and 20 in the validation phase, resulting in a 95% confidence level (±10%). The 10% error margin exceeds the standard 5% typically used in clinical trials but was maintained within the VOCORDER clinical trial due to the exploratory, first-in-device nature of the study, consistent with precedent in breath biomarker protocol studies of comparable cohort size. The analytical strategy proceeds from binary disease-specific classifiers to a secondary multiclass ensemble, constraining model complexity relative to sample size. To mitigate overfitting risk, nested cross-validation, an independent hold-out set, regularisation and SHAP-based dimensionality reduction will be employed, as described in the Data Analysis Procedure. Multiple hypothesis testing across VOC features will be controlled using the Benjamini-Hochberg false discovery rate correction. A total of 240 healthy participants are required across both phases. The primary power calculation pertains to binary (disease vs healthy control) comparisons, which represent the fundamental inferential unit of this study. Multiclass discrimination across all five disease categories will be conducted as a secondary, exploratory analysis. To control the familywise error rate across multiple disease comparisons, Benjamini-Hochberg false discovery rate correction will be applied. Each pairwise classifier retains the statistical power established by the binary calculation, since the one-vs-rest decomposition preserves the effective sample size per comparison. The healthy participant cohort needed to be significantly larger than the patient subjects to account for natural discrepancies observed in some diseases, such as the higher incidence of breast cancer among women compared with men. Consequently, 240 healthy participants were required. This consideration is crucial for developing unbiased machine learning AI techniques. Ethics and dissemination Participation in the VOCORDER clinical trial is entirely voluntary. A comprehensive information sheet outlining the procedures, risks and objectives of the project, along with a trained physician providing a verbal explanation to participants in Greek (a language spoken by all participants), will be available. Individual participants must provide consent, and a consent form will be signed. For those who are illiterate, non-written consent will be obtained and independently witnessed. The data collected from the questionnaires and personal interviews will be compiled by the MITERA research team and uploaded onto a secure network, with access restricted solely to the VOCORDER project research team. Patient data will also be anonymised. Data scrubbing may be permitted once the request has been evaluated and there is no hindrance to achieving the project aims. Patient and public involvement Patients were not involved in the design of this clinical study. However, they will contribute to dissemination activities concerning the acceptance of the VOCORDER device and study outcomes, particularly through communication materials tailored to patients and the general public (questionnaires). The project results will be published in open-access journals, presented at national and international conferences and communicated to the public and other relevant stakeholders via the communication channels of all investigators and project partners. All results derived from the study will be accessible, in line with the European Union and the State Secretariat for Education, Research and Innovation open access policy. If an external scientist wishes to access the GC-MS and SESI-HR-MS breath data, there are straightforward options to do so (for example, setting up a guest account at ETH that provides VPN access). Such access will be granted on reasonable request. Data statement The data will be collected and managed by the MITEPA project team and stored in a computerised working folder within the hospital. This folder is only accessible to the MITERA project team. Patient-related data will then be anonymised and entered into a specially developed database shared with authorised partners. The database will also include results from both methodologies. The protection of patients’ personal data will be ensured in accordance with Greek legislation, the relevant European General Data Protection Regulation and MITERA hospital regulations. Data availability The original data generated by the study will be accessible in a curated data archive at ETH Zurich (https://www.research-collection.ethz.ch), the so-called ‘green way’ for open access. Discussion The accuracy of VOCORDER technology is fundamentally determined by the spectral information obtained during the baseline and validation phases of the clinical trials analysing breath samples. The study’s advantages include (a) it is a prospective study and (b) the diagnosis of the diseases under investigation occurs before the initiation of treatment. Regarding the limitations, the small sample size is a factor; however, all reference studies also involve a small number of patients. Samples will be collected in 1 L Supel-Inert Multi-Layer Foil gas sampling bags, based on previous studies79,81 as well as preliminary tests conducted in-house observing the change in VOC concentrations over 1 week, which showed reduced occurrence of positive or negative artefacts (often problematic in offline sampling techniques). Despite the effectiveness of the sampling bags for storing breath samples and the fact that offline breath samples are analysed as soon as possible, the analyses are nonetheless still conducted offline, and the formation of potential artefacts must be taken into account. One potential source of artefacts that may be concerning is the pressure changes that occur during the transportation of samples. The 1 L sampling bags are transported from Greece to Zürich (approximately a 3-hour flight). Future quality assurance efforts should focus on determining whether the mix of VOCs in a breath sample could be affected by these pressure changes over the course of a flight. One potential method to investigate this is to directly measure a participant’s breath online in addition to collecting a breath sample in a 1 L gas sampling bag. However, the analytical instruments would have needed to be available at the clinical trial site, which they were not. This test would however help demonstrate how the breath matrix and the presence of specific biological metabolites of interest could change in response to external stimuli. Furthermore, while gas sampling bags are a strong option for storing breath samples and conducting offline chemical analyses, the loss of specific VOC species is unavoidable, particularly for the larger, more specific biomarker metabolite molecules. To avoid potential artefacts, the future directions of this work will consider online analyses. Online analyses have been used in several prior breath studies,82,89 although the logistics of the VOCORDER clinical trial made it impractical on this occasion. To further enhance the study, incorporating an additional analysis method that enables online real-time measurements of VOCs in the gas phase and serves as a quantitative technique would be beneficial. SIFT-MS has been used in prior breath analysis studies,39,46 sampling both online and offline specimens. A notable advantage of SIFT-MS is its ability to separately quantify isomers by switching between different reagent ions, which exhibit unique ion-molecule reactions.90 For both online and offline measurement techniques used for breath analysis, another potential source of error may arise from environmental artefacts.2627 91,93 VOC concentrations in the breath can be altered in situ, depending on a participant’s lifestyle or exposure to external sources of VOCs (eg, air pollution).10 26 27 93 Breath VOC concentrations may also shift based on a subject’s medication, use of cosmetics, level of physical exercise and circadian rhythm.94,97 Consequently, accurately measuring a VOC profile and correlating it with disease incidence becomes increasingly challenging. The spectra and data obtained from the GC-MS and SESI-HR-MS, along with the medical data from questionnaires and personal interviews, will be input into statistical AI models. Five reference diseases (based on a literature review conducted by our group) have been used as a foundation for the investigation (table 1) into the correlations between VOC matrix profiles of patients with specific disease incidences. These models developed during the baseline phase will be used for the newly developed VOCORDER device to enable early disease detection using the discovered VOC profile patterns. The use of unsupervised AI models for statistical analysis of breath samples has previously been investigated as a method for detecting cancer from breath, employing principal component analysis. Maiti et al22 used a laser-free broadband mid-infrared Fourier-transform spectroscopy technique to analyse breath samples for metabolite VOCs from patients with prostate, bladder and kidney cancers. However, they report that further work is required. Aslam et al98 also explored unsupervised methods to assess the possibility of diagnosing gastric cancer in the early stages of disease, demonstrating an overall accuracy of 97.3% in early diagnosis. Another study that exploited unsupervised AI models is that of Alkhalifah et al.99 Similar to our work, the authors acquired GC-MS data from participants’ breath samples. These authors employed a data clustering technique to identify similarities in the spectra from the samples. Semi-supervised and supervised AI machine learning approaches have previously been used for the detection of SARS-CoV-2 (COVID-19).100 Additionally, work by Shi et al101 successfully developed a deep transfer learning method to identify lung cancer incidence by detecting nodules in patients’ lungs. Supervised AI machine learning techniques were applied in a study to detect leukaemia biomarkers from breath102 and for lung cancer detection.103 In all these prior works, the potential of AI methods for early disease detection has been demonstrated. However, the research community has identified a need for further study before these techniques can become standard in medical practice. Supplementary material Acknowledgements The authors extend their sincere thanks to the entire project consortium: Mitera Hospital, Alpes Lasers SA, EMPA, UAB Metis Baltic, Neuraltech IKE, Vrije Universiteit Brussel, EULAMBIA Advanced Technologies, Cailabs, Argos Messtechnik, AiDEAS OU, ICCS and ETH Zürich. Special acknowledgments to the personnel of Mitera Hospital for their support and facilitation of breath sample collection. Footnotes Funding: The research leading to this review paper has received funding from the VOCORDER project, which is co-funded by the European Union under grant agreement 101115442 and the State Secretariat for Education, Research and Innovation (SERI), REF-1131-52304/SBFI-Nr.23.00369. Prepublication history and additional supplemental material for this paper are available online. To view these files, please visit the journal online (https://doi.org/10.1136/bmjopen-2025-108562). Provenance and peer review: Not commissioned; externally peer reviewed. Patient consent for publication: Not applicable. Patient and public involvement: Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research. References Supplementary Materials Data Availability Statement The original data generated by the study will be accessible in a curated data archive at ETH Zurich (https://www.research-collection.ethz.ch), the so-called ‘green way’ for open access.