Image description

We've published work in Analytical Chemistry on a serum-based blood test for lung cancer that reads the full Raman spectrum of a patient's blood rather than chasing a single biomarker. The appeal of this "data-first" approach is that it needs no predetermined target and only 5 microliters of serum, but is limited by the high similarity in chemical composition between healthy and cancer serum, which leaves the disease-relevant differences buried in noise and in the natural biological variation from one patient to the next.

Our method, CWT-SerumCODE, tackles this by converting each 1D Raman spectrum into a 2D scalogram using a continuous wavelet transform, which naturally separates sharp molecular signals from the broad fluorescence background, and then classifying those scalograms with a convolutional neural network. To reflect real clinical use, we split the data at the patient level, so no spectra from a given person appear in both training and testing. This "leakage-safe" validation mirrors how spectra would actually be classified in the clinic, rather than measuring accuracy at the single-spectrum level. The method is also interpretable through Grad-CAM plus inverse wavelet reconstruction which ties predictions to specific Raman shifts.


On an independent validation cohort, the model reached 90.5% accuracy (91.7% sensitivity, 88.9% specificity), well above the conventional machine-learning baselines we tested. The features driving its decisions map onto phenylalanine, lipids, nucleotides, and tryptophan, all molecules with established roles in cancer metabolism, which suggests the classifier is activating on real biochemistry rather than acquisition artifacts. We consider this a proof of concept given the modest validation set (n = 21), and larger multicenter cohorts would be needed before any clinical use, but it establishes that comprehensive spectral analysis can deliver both accuracy and mechanistic transparency.


Full Text at Analytical Chemistry