A device that essentially “smells” cancer in a drop of blood has demonstrated near-perfect accuracy in detecting ovarian cancer, according to new research published in Advanced Intelligent Systems.

The electronic nose — a 32-sensor array that detects volatile organic compounds (VOCs) emitted by blood plasma — correctly identified ovarian cancer patients with 97 per cent sensitivity and 97 per cent specificity. When results were aggregated at the patient level using a majority-vote algorithm, diagnostic accuracy reached 100 per cent.

A “smell test” for the silent killer

Ovarian cancer is among the deadliest gynecologic cancers, largely because it rarely produces clear symptoms until it has spread. Patients diagnosed at stage I have a five-year survival rate of 93 per cent, but that figure collapses to 31 per cent for those diagnosed at stage IV. The World Cancer Research Fund recorded 324,603 new cases and nearly 207,000 deaths globally in 2022, with projections suggesting those numbers will climb to over 500,000 cases and 351,000 deaths by 2050.

Current screening tools — blood tests for biomarkers like CA-125, transvaginal ultrasound, CT scans, and biopsies — are either too imprecise, too expensive, too invasive, or too dependent on specialist interpretation to work as population-level screening. Even composite approaches like the Risk of Ovarian Malignancy Algorithm (ROMA) fall short of what would be needed for routine use.

How the electronic nose works

The device, developed by VOC Diagnostics AB in Linköping, Sweden, contains 32 metal-oxide semiconductor gas sensors arranged in four temperature-controlled banks. A one-millilitre blood plasma sample is placed in a chamber, and a fan draws the VOCs released by the sample across the sensor array over a 10-minute measurement cycle. Each sensor produces a distinct electrical signal in response to the chemical cocktail, creating a multi-dimensional “smell fingerprint.”

Rather than hunting for any single biomarker, the system takes a biomarker-agnostic approach — it reads the entire volatile chemical profile of the sample and lets a machine learning algorithm sort out the patterns. The researchers extracted 85 features from each sensor signal, including time-domain, frequency-domain, and statistical measures, and fed them into an ensemble boosting model that was systematically optimized for the task.

Tested on nearly 300 patients

The study analysed blood plasma from 134 ovarian cancer patients, 41 endometrial cancer patients, and 115 healthy volunteers — a substantially larger and more diverse dataset than previous electronic nose studies, which typically involved fewer than 100 cancer cases. Samples were collected from three Swedish university hospitals between 2015 and 2022.

After evaluating 43 different machine learning architectures, the team settled on an Optimizable Ensemble model using the GentleBoost method — essentially a chain of 495 weak decision-tree classifiers that iteratively correct each other’s errors. A systematic sensor evaluation identified two of the 32 sensors as contributing minimal discriminatory value, and excluding them actually improved performance.

In rigorous testing with 90-10 train-test splits and five-fold cross-validation repeated across 15 independent runs, the optimized model delivered a mean test accuracy of 97.2 per cent, with sensitivity of 97.1 per cent and specificity of 97.3 per cent. The narrow confidence intervals confirmed the results were stable and reproducible rather than artifacts of a lucky data split.

Beyond detection: staging and cancer-type classification

The researchers went further than simply distinguishing cancer from healthy. Using the same feature extraction and sensor evaluation framework, they built a second classifier that could differentiate stage I ovarian cancer from more advanced stages (II through IV), achieving a mean test accuracy of about 93 per cent. This is clinically significant because early-stage detection is precisely where current methods fail most.

A third model tackled an equally important problem: distinguishing ovarian cancer from endometrial cancer, a neighbouring malignancy that can produce overlapping biomarker signatures in blood. That classifier achieved a mean test accuracy of 97.1 per cent, with perfect patient-level classification after majority voting.

The team also outlined a sequential diagnostic pipeline in which the system would first detect the presence of cancer, then determine whether it was ovarian or endometrial, and finally stage the ovarian cancer — a layered approach designed to mimic the kind of step-by-step clinical reasoning that could translate directly into screening workflows.

Benchmarking against the field

Compared to 34 recent studies using a range of diagnostic methods — including liquid biopsy, imaging, and biomarker panels — the electronic nose system achieved among the highest combined sensitivity and specificity, and did so with a moderate cohort size and a relatively simple, low-cost testing procedure.

Previous electronic nose studies for ovarian cancer detection have achieved sensitivities ranging from 79 to 98 per cent and specificities from 86 to 100 per cent, but generally used smaller sample sizes, fewer sensors, and did not attempt staging or multi-cancer classification.

Practical implications

The procedure requires only one millilitre of blood plasma, takes 10 minutes per measurement, requires no specialist interpretation, and uses commercially available sensors. These characteristics make it a candidate for high-throughput population screening — the kind of affordable, scalable early detection tool that ovarian cancer outcomes have long needed.

The researchers note that further validation in larger, independent cohorts will be necessary before clinical deployment, and that sensitivity for early-stage detection in mixed populations (where early-stage ovarian cancer must be identified among patients with other conditions) still needs improvement. But the results represent what they call “a significant advance toward reliable, accessible, and minimally invasive VOC-based oncology diagnostics.”

The study was published in Advanced Intelligent Systems on January 6, 2026.

Source: Shtepliuk, I., Meng, Q., 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, 2500838. DOI: 10.1002/aisy.202500838