Ovarian cancer is one of the highest-risk gynecologic cancers, partly because it often goes undetected in its early stages, with very few visible symptoms.
As a result, many cases are diagnosed late, when treatment is more difficult, and the risk of death is higher. Existing diagnostic approaches, including serum biomarker tests, invasive procedures, and imaging methods, are expensive, invasive, and often less reliable for early-stage disease.
Many also depend heavily on expert interpretation. Even multiparameter approaches have struggled to deliver the level of accuracy needed for effective screening.
Recent methods aided by artificial intelligence have improved some aspects of diagnosis, but many still rely on established imaging tools or known biomarkers.
In answer to these issues, this study set out to create a faster, cheaper, and less invasive approach that could help detect ovarian cancer earlier.
Saving this for later? Grab a PDF here.
A Biomarker-Agnostic Alternative
By analyzing the full pattern of volatile organic compounds (VOCs), sometimes called the volatilome, produced by the body, instead of looking for a single biomarker, there may be more success.
These compounds are present in blood, breath, and urine, and can reflect metabolic changes linked to disease. In theory, that means cancer could be identified from an overall chemical pattern rather than one predefined target.
Electronic noses, or eNoses, are designed for this type of pattern detection. When combined with machine learning (ML), they can turn complex sensor signals into diagnostic predictions, offering a potentially scalable and minimally invasive route to cancer detection.
Earlier studies have shown that VOC patterns in blood plasma differ between healthy individuals and patients with ovarian cancer. But models trained to separate the disease from healthy samples have had limited success when asked to identify disease stage or handle more varied early-detection settings.
How this Study Resolved the Issue
Published in Advanced Intelligent Systems, the researchers developed a biomarker-agnostic diagnostic platform based on a 32-element metal-oxide semiconductor (MOS) eNose to analyze VOC emissions from blood plasma.
They tested plasma samples from 115 healthy controls, 134 patients with ovarian cancer, and 41 patients with endometrial cancer, providing a more varied dataset than earlier work that used the same 32-sensor platform.
The larger sampling set was intended to capture a broader range of VOC signatures and improve the model’s reliability across a more diverse population.
The team paired that sensing system with an ML framework that included sensor-exclusion algorithms, ensemble-learning strategies, and feature engineering for VOC-pattern recognition, including stage-specific analysis.
Including endometrial cancer cases addressed an important clinical question: does another gynecologic cancer interfere with ovarian cancer detection?
Misclassifying endometrial cancer as ovarian cancer could lead to incorrect follow-up tests, unnecessary patient distress, increased healthcare costs, and less appropriate treatment decisions.
Analyzing Blood Plasma
Blood plasma samples were collected from healthy volunteers aged 30 to 60 and from patients with ovarian cancer and endometrial cancer aged 18 to 90.
The plasma was divided into 1 mL aliquots and stored at -80 °C in biobank tubes until analysis. To preserve sample quality, the handling procedure was standardized, and each aliquot was thawed only once before VOC measurement at room temperature, 21 °C.
Measurements were carried out using an eNose prototype containing a 32-sensor MOS array arranged in four temperature-controlled banks.
Each 1 mL plasma sample was placed in a chamber inside the device, and sensor responses were recorded at 600 points per minute during a 10-minute protocol that included baseline recording, VOC exposure after fan activation, and post-exposure monitoring.
The researchers then used an iterative optimization process to remove sensors with the lowest discriminatory power. After identifying the best sensor setup, they ran seven comparative experiments to test the effects of different training conditions.
These included changing the train-test split from 90:10 to 75:25, varying the number of cross-validation folds from five to 10, testing hyperparameter optimization, and, in one case, excluding features before training. They also repeated the 90:10 split across 15 random cycles to test how much performance depended on the data partitioning.
Findings from the eNose Study
The researchers reported a biomarker-independent method for detecting ovarian cancer from blood plasma using an eNose and ML analysis of VOC patterns. The model used 85 features drawn from frequency-domain, statistical, and time-domain measures, along with sensor similarity coefficients.
In the main ovarian cancer-versus-healthy-control classification task, the model achieved 97 % sensitivity and 97 % specificity. A majority-vote decision algorithm then produced perfect patient-level classification in that setting.
Feature-importance analysis and Wilcoxon rank-sum testing helped identify which variables contributed most to performance.
The system was also assessed for ovarian cancer staging and for distinguishing ovarian from endometrial cancer.
Those results supported the broader potential of the approach, but they also showed its limits. Performance was weaker in more clinically realistic mixed-population and early-stage detection scenarios than in the headline comparison between ovarian cancer cases and healthy controls.
Why The Nuance Matters
That distinction matters. The strongest results came from separating ovarian cancer cases from healthy controls. When the task became more demanding, especially in early-stage and more mixed settings, sensitivity fell below the headline result.
The technology looks promising as a proof of concept, but it still needs further validation before it can be considered for broader screening use.
Even so, the study suggests that intelligent eNose systems could become a credible biomarker-agnostic route toward blood-based ovarian cancer detection, especially if future work improves performance in clinically mixed populations.
Journal Reference
Shtepliuk, I., Meng, L., Borgfeldt, C., Eriksson, J., & Puglisi, D. (2026). Biomarker-Agnostic Detection of Ovarian Cancer from Blood Plasma Using a Machine Learning-Driven Electronic Nose. Advanced Intelligent Systems, e202500838. DOI: 10.1002/aisy.202500838
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.