Confounding factors in olfactory aroma data, such as high humidity levels, substantially affect sensor outputs, masking subtle volatile organic compound (VOC) patterns and hindering generalizable machine learning models. Traditional representation learning methods often require large datasets to mitigate confounder-induced variance, a resource unavailable in specialized sensor applications with limited data. This study presents Confounder-Invariant Representation Learning (CIRL), a method designed to mitigate confounding influences in data-scarce settings by leveraging explicit confounder information, such as relative humidity. CIRL enhances learned representations by reducing confounder effects, improving data purity and model robustness. Applied to three breath aroma datasets—acetone, ketosis, and peppermint-oil breath, all affected by high humidity—CIRL was integrated with standard autoencoder models. Evaluated within the same framework, CIRL improved generalization performance by 10–15% in classification accuracy across all three datasets. These results demonstrate CIRL’s potential to advance reliable artificial olfaction for applications like breath-based diagnostics in challenging real-world conditions.

Confounding factors in olfactory aroma data, such as high humidity levels, substantially affect sensor outputs, masking subtle volatile organic compound (VOC) patterns and hindering generalizable machine learning models. Traditional representation learning methods often require large datasets to mitigate confounder-induced variance, a resource unavailable in specialized sensor applications with limited data. This study presents Confounder-Invariant Representation Learning (CIRL), a method designed to mitigate confounding influences in data-scarce settings by leveraging explicit confounder information, such as relative humidity. CIRL enhances learned representations by reducing confounder effects, improving data purity and model robustness. Applied to three breath aroma datasets—acetone, ketosis, and peppermint-oil breath, all affected by high humidity—CIRL was integrated with standard autoencoder models. Evaluated within the same framework, CIRL improved generalization performance by 10–15% in classification accuracy across all three datasets. These results demonstrate CIRL’s potential to advance reliable artificial olfaction for applications like breath-based diagnostics in challenging real-world conditions.

Background/Objectives: Breast cancer is the most common malignancy among women, and early detection is critical for improving outcomes. The Breast Imaging Reporting and Data System (BI-RADS) standardizes reporting, but the BI-RADS 4 category presents a major challenge, with malignancy risk ranging from 2% to 95%. Consequently, most women in this category undergo biopsies that ultimately prove unnecessary. This study evaluated whether exhaled breath analysis could distinguish malignant from benign findings in BI-RADS 4 patients. Methods: Participants referred to the McGill University Health Centre Breast Center with BI-RADS 3–5 findings provided multiple breath specimens. Breathprints were captured using an electronic nose (eNose) powered breathalyzer, and diagnoses were confirmed by imaging and pathology. An autoencoder-based model fused the breath data with BI-RADS scores to predict malignancy. Model performance was assessed using repeated cross-validation with ensemble voting, prioritizing sensitivity to minimize false negatives. Results: The breath specimens of eighty-five participants, including sixty-eight patients with biopsy-confirmed benign lesions and seventeen patients with biopsy-confirmed breast cancer within the BI-RADS 4 cohort were analyzed. The model achieved a mean sensitivity of 88%, specificity of 75%, and a negative predictive value (NPV) of 97%. Results were consistent across BI-RADS 4 subcategories, with particularly strong sensitivity in higher-risk groups. Conclusions: This proof-of-concept study shows that exhaled breath analysis can reliably differentiate malignant from benign findings in BI-RADS 4 patients. With its high negative predictive value, this approach may serve as a non-invasive rule-out tool to reduce unnecessary biopsies, lessen patient burden, and improve diagnostic decision-making. Larger, multi-center studies are warranted.

Background/Objectives: Breast cancer is the most common malignancy among women, and early detection is critical for improving outcomes. The Breast Imaging Reporting and Data System (BI-RADS) standardizes reporting, but the BI-RADS 4 category presents a major challenge, with malignancy risk ranging from 2% to 95%. Consequently, most women in this category undergo biopsies that ultimately prove unnecessary. This study evaluated whether exhaled breath analysis could distinguish malignant from benign findings in BI-RADS 4 patients. Methods: Participants referred to the McGill University Health Centre Breast Center with BI-RADS 3–5 findings provided multiple breath specimens. Breathprints were captured using an eNose powered breathalyzer , and diagnoses were confirmed by imaging and pathology. An autoencoder-based model fused the breath data with BI-RADS scores to predict malignancy. Model performance was assessed using repeated cross-validation with ensemble voting, prioritizing sensitivity to minimize false negatives. Results: The breath specimens of eighty-five participants, including sixty-eight patients with biopsy-confirmed benign lesions and seventeen patients with biopsy-confirmed breast cancer within the BI-RADS 4 cohort are analyzed. The model achieved a mean sensitivity of 88%, specificity of 75%, and a negative predictive value of 97%. Results were consistent across BI-RADS 4 subcategories, with particularly strong sensitivity in higher-risk groups. Conclusions: This proof-of-concept study shows that exhaled breath analysis can reliably differentiate malignant from benign findings in BI-RADS 4 patients. With its high negative predictive value, this approach may serve as a non-invasive rule-out tool to reduce unnecessary biopsies, lessen patient burden, and improve diagnostic decision-making. Larger, multi-center studies are warranted.