A compact aluminum sample well collects a Raman spectrum from 60 microlitres of blood plasma in under two minutes. In a 372-sample biobank study, the ROWS method paired that optical readout with machine learning to distinguish several lung conditions, including early- and late-stage lung cancer.
Raman spectroscopy reads the small shifts in laser light scattered by molecules. Those shifts provide a composite fingerprint of proteins, lipids and other components without adding chemical reagents. Liquid samples are difficult because spontaneous Raman signals are weak and can be obscured by the container. ROWS, short for Raman of Well-based Samples, uses a custom metal well and a single-point spectrometer to collect a usable spectrum from a plasma droplet.
The study, published in Scientific Reports on September 18, used plasma stored by a national biobank. The final dataset included 92 controls, 99 people with stage I–II lung cancer, 46 with stage III–IV disease, 36 with benign lung lesions and 99 with other lung conditions. The last group included disorders such as asthma and chronic obstructive pulmonary disease, which can produce symptoms that overlap with lung cancer.
Different lung conditions, different accuracy
The strongest reported model separated lung cancer from non-tumoral lung conditions. In five-fold cross-validation within the biobank cohort, it reached 94% sensitivity, 90% specificity and 93% accuracy, with an area under the receiver operating characteristic curve of 0.97. A staging model that separated stage I–II from stage III–IV cancer reached 88% accuracy and an area under the curve of 0.94.
The broader test was harder. When all 227 samples without cancer, including healthy controls, benign lesions and other lung conditions, were compared with all 145 cancer samples, accuracy was 78%, with 78% sensitivity and 78% specificity. That gap matters because the 94% sensitivity figure belongs to a narrower question in a selected cohort, not to general screening for lung cancer.
The models used support vector machines, a method that draws a boundary between groups from selected spectral features. The researchers first reduced roughly 900 spectral bins to bands around 21 recurring Raman peaks, then selected features and tuned the classifier across five folds. Several protein-associated peaks became relatively stronger in cancer samples, while lipid- and carotenoid-associated signals often weakened. Because the spectra were normalized and a Raman peak can represent more than one molecule, the paper treats those biochemical assignments as tentative rather than direct concentration measurements.
Testing stored samples rather than new patients
ROWS addresses an instrumentation problem: obtaining a reagent-free spectrum quickly from a small liquid volume. It does not yet establish a diagnostic workflow. The samples were frozen biobank plasma collected from veins, while finger-prick collection is proposed as a compatible future route. The study did not prospectively test fresh samples, lock the model before evaluation or validate it at an independent hospital.
Five-fold cross-validation shows how the models behaved when different parts of this cohort were held out, but it cannot measure changes caused by another site's patients, handling process or instrument. The reported sensitivity and specificity also depend on the classification threshold chosen from the study data.
The paper was funded through a Natural Sciences and Engineering Research Council of Canada Alliance grant in collaboration with Reveal Life Science. It also discloses that three authors are Reveal shareholders and three are employees.
Sources
- Scientific Reports article and accepted-version metadata, published September 18, 2026
- Accepted article PDF with cohort definitions, model comparisons and methods
- Earlier medRxiv version with the five-fold classification procedure
- Lead image: representative Raman instrument, not the ROWS prototype, photographed by Ijon and reused under CC BY-SA 4.0 via Wikimedia Commons.