1216P A spectroscopic liquid biopsy for the earlier detection of multiple cancer types

Abstract

Background
Employing a rapid liquid biopsy platform that can support clinicians in the diagnosis of different cancers, particularly for patients who develop cancers not targeted in current screening programs, would cause a paradigm shift in cancer diagnostics. Current liquid biopsies focus on single tumor derived biomarkers, such as circulating tumor DNA (ctDNA), which limits test sensitivity, especially for early-stage cancers that do not shed enough genetic material.

Methods
The Dxcover® Cancer Liquid Biopsy has been assessed upon its ability to predict individual cancers in organ-specific classifications: brain, breast, colorectal, kidney, lung, ovarian, pancreatic, and prostate cancer. The test uses Fourier transform infrared (FTIR) spectroscopy to build spectral profiles of serum samples, and machine learning algorithms to predict disease status. We also made a further exploratory evaluation of the ability to differentiate the signature from any one of the 8 cancers from non-cancer patient samples. We assessed the test performance when the cancer samples were grouped together to mimic patients with non-specific symptoms where the cancer site was uncertain. Additionally, we have examined non-generative data augmentation methods to improve machine learning performance.

Results
Area under the receiver operating characteristic curve (AUROC) values were calculated for 8 cancer types v symptomatic non-cancer controls: most classifiers achieved AUROC values above 0.85. The cancer v asymptomatic non-cancer classification detected 64% of stage I cancers when specificity was 99% (overall sensitivity 57%). When tuned for higher sensitivity, this model identified 99% of stage I cancers (with specificity 59%). For the colorectal cancer dataset, data augmentation using a WGAN led to an increase in AUROC from 0.91 to 0.96, demonstrating the impact data augmentation can have on deep learning performance, which could be useful when the amount of real data available for model training is limited.