HS-GC/MS-Based Targeted Blood VOC Panel for Canine Cancer Screening Using an Integrated Classification Framework in...

Canine cancer screening has an important limitation in that organ-specific screening strategies used in human medicine cannot be directly applied across routine veterinary clinical practice. Blood-based volatile organic compound (VOC) profiling may provide a practical approach for pan-cancer screening; however, its application in multicenter and multinational settings requires evaluation that accounts for cohort heterogeneity and analytical variability. In this study, we evaluated the canine cancer screening performance of an integrated classification framework using a headspace gas chromatography-mass spectrometry (HS-GC/MS)-based targeted blood VOC panel in a retrospective multicohort dataset comprising dogs recruited in Korea and Thailand. Whole-blood samples from 590 dogs recruited in Korea and Thailand were analyzed, including 167 dogs in the malignant tumor group and 423 dogs in the non-cancer control group. The predefined targeted VOC panel consisted of aromatic compounds and straight-chain alkanes and was analyzed using HS-GC/MS. The Korea and Thailand cohorts were incorporated into a single integrated classification framework rather than being separated into country-specific models. To evaluate performance variability associated with data partitioning, cohort-aware stratified training-validation splitting was repeated 30 times within the integrated retrospective cohort, with both national cohorts represented in the training and validation sets in each repeat. Class imbalance correction was applied only to the training data, and validation performance was assessed using both threshold-dependent and threshold-independent metrics. Across the 30 repeated validations, the integrated framework achieved an AUROC of 0.926 ± 0.019 and an AUPRC of 0.894 ± 0.023. Balanced accuracy, accuracy, sensitivity, specificity, F1-score, MCC, and Cohen's κ were 0.858 ± 0.024, 0.895 ± 0.018, 0.772 ± 0.048, 0.943 ± 0.022, 0.805 ± 0.033, 0.736 ± 0.046, and 0.734 ± 0.045, respectively. In the sample-level prediction score analysis, the median predicted cancer probabilities for cancer and non-cancer samples were 0.954 and 0.041, respectively, indicating clear probability-based separation between the two groups. In the cohort-specific analysis, the discrimination signal was maintained in both the Korea and Thailand cohorts, although threshold-dependent performance differed between the two cohorts. These findings indicate that an HS-GC/MS-based targeted blood VOC panel provided a consistent cancer-associated classification signal under repeated internal validation within the specific integrated retrospective dataset evaluated in this study. The results demonstrate the feasibility of evaluating a common classification framework across the Korea and Thailand cohorts without constructing separate country-specific models, while not establishing independent external generalizability or platform-independent robustness. Because the Korea and Thailand cohorts were completely confounded with the Agilent and Thermo Fisher analytical platforms, respectively, the independent contributions of cohort characteristics and analytical platform cannot be determined from the present data. Further independent prospective validation in newly recruited populations and institutions, together with disease-control evaluation including clinically relevant controls, cohort-specific calibration, evaluation across analytical platforms, and assessment of clinical utility within real-world workflows, is required before the external generalizability and clinical applicability of the framework can be determined.