Point-of-care cancer screening using saliva transmission spectroscopy and machine learning.

Abstract

e22526 Background: Tier-1 cancer screening would benefit from point-of-care (POC) testing that is rapid, inexpensive, and highly specific to minimize unnecessary downstream workup. We evaluated saliva transmission hyperspectral spectroscopy combined with machine learning algorithms as a POC screening approach. Methods: Saliva was collected from outpatient consented participants in a lung cancer clinic (n = 99) and a high-risk clinic (n = 152) on an IRB-approved protocol (PREDICT NCT05802069). Two drops of saliva were scanned in ~3 seconds on a ProSpectral hyperspectral spectrophotometric device to acquire transmission spectra. Cancer status and stage were abstracted from the electronic health record and categorized as cancer-positive, hereditary cancer-predisposition but without diagnosis, no evidence of disease (NED) while on treatment, and NED at least 2 months after treatment completion. Spectra and labels were used to train and validate a Pattern Discovery Engine (PDE) classifier, which yields a human-readable symbolic equation in the spectral domain; operating thresholds were calibrated to tune specificity for low false-positive tier-1 operation. Results: Across clinically relevant labeling i.e., cancer-positive (n = 99 patients), hereditary cancer-predisposition but without diagnosis (n = 86), no evidence of disease (NED) while on treatment (n = 16), and NED at least 2 months after treatment completion (n = 50), models achieved high median balanced accuracy and maintained discrimination under very high specificity settings suitable for tier-1 screening (very few false positives). At comparable cohort scale, discrimination exceeded that reported for sequencing-based cfDNA methylation blood assays, while reducing time-to-signal from centralized laboratory workflows (~2-week turnaround) to seconds at the POC and avoiding reagent costs. Saturation analysis indicates that Specificity > 98% is likely attainable with fewer than 500 samples, providing a path to a clinically actionable screening protocol in subsequent phases of work. Conclusions: Saliva transmission hyperspectral spectroscopy with the PDE supports a practical, near-real-time POC paradigm for early cancer screening and triage. Ongoing work will attribute discriminative spectral regions and identify contributing analytes and host-response biomarkers using orthogonal assays (e.g., fractionation and targeted mass spectrometry), improving interpretability and enabling prospective validation. Samples Balanced Accuracy Sensitivity Specificity F1-Score 251 61% 44% 91% 0.55