Di Gilio Alessia, Palmisani Jolanda, Nisi Marirosa, Varesano Niccolò, Catino Annamaria, Galetta Domenico, de Gennaro Gianluigi
Department of Biosciences, Biotechnologies and Environment, University of Bari Aldo Moro, Bari, Italy
Apulian Regional Centre for the Breath Analysis, Istituto Tumori ‘Giovanni Paolo II’, Bari, Italy
Thoracic Oncology Unit, Istituto Tumori ‘Giovanni Paolo II’, Bari, Italy
Introduction
Malignant Pleural Mesothelioma (MPM) is a rare neoplasm cancer with still a poor prognosis and mainly caused by previous asbestos occupational and environmental exposure. The management of MPM is challenging due to the long latency period between exposure and diagnosis and due to symptoms appearing only at an advanced stage [1]. Thus, there is an ever-increasing need to introduce reliable and smart diagnostic tools into large-scale screening programs. Recently, the chemical characterization of Volatile Organic Compounds (VOCs) in human breath and the identification of disease-related metabolites pattern has been recognized as non- invasive and promising approach for the early detection of neoplastic diseases [2]. The aim of this cross-sectional observational study was to identify a MPM-related VOCs pattern in human breath in order to develop a rapid, reliable and non-invasive method for early diagnosis of MPM.
Methods
An overall number of 133 individuals were enrolled and stratified into three groups: 64 patients with confirmed MPM, 8 subjects MPM in follow-up (MPM-FU), and 61 healthy controls (HC). The enrollment of volunteers in the clinical trial fulfilled specific criteria, after approval by the Ethical Committee. End-tidal breath fraction (EXP) of all the enrolled subjects was collected directly onto two-beds adsorbent cartridges (Biomonitoring steel tubes, Markes International) by means of the automated sampler Mistral (Predict srl). Ambient air samples (AA) were simultaneously collected at each sampling session. Breath and AA samples were thermally desorbed (Unity Ultra-xr Markes) and analyzed by Gas Chromatography/Mass Spectrometry (GC Agilent 7890/MS Agilent 5975). Experimental data were statistically processed by a machine learning approach based on a Random Forest algorithm. The model was trained using a k-fold cross-validation framework (with k = 10) to ensure robust performance evaluation. Feature selection for the RF classifier was carried out using an embedded method based on the Gini index. Starting from this ranking, models were iteratively built using progressively larger subsets of variables in order to identify the optimal number of compounds to achieve the best classification performance.
Results and Discussions
Statistical treatment of experimental data, e.g., VOCs abundances in MPM and HC samples, resulted in a high classification accuracy, achieving a sensitivity equal to 92% and diagnostic accuracy equal to 86% (area under the curve AUC: 0.861). Feature selection revealed a subset of 15 VOCs (including alkanes, alkenes, aldehydes, ketones and alcohols) contributing most significantly to the classification and discrimination between MPM and HC. The explorative application of a Random Forest classification model to blinded follow-up samples yielded probability scores consistent with clinical expectations, demonstrating the feasibility of breath-based surveillance in longitudinal monitoring of MPM.
Conclusions
Despite the promising results obtained in this study and their coherence with previous literature, the size and the homogeneity of the sample population deserves further investigation to validate breath analysis as a helpful tool in the screening and clinical management of MPM. More specifically, validation in larger and independent cohorts is essential to confirm the method specificity and robustness in the follow-up of MPM patients. However, breathomics, when combined with machine learning-based classification and rigorous methodological controls, could be a promising tool for the early diagnosis and surveillance of MPM patients.
