The urinary-metabolite-based lung cancer index (uLCI): an interpretable machine-learning risk model for early-stage disease
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Five-year survival from lung cancer exceeds 60% at stage I–II but falls below 10% once metastasis occurs. Low-dose CT (LDCT) screening reduces mortality in heavy smokers but carries a false-positive rate of approximately 29% and is restricted to smoking-based eligibility, leaving most cases undetected. We aimed to develop and independently validate an interpretable machine-learning urinary metabolite risk index (uLCI) for non-invasive lung cancer detection.
Methods
Four urinary metabolites—creatine riboside (CR), N-acetylneuraminic acid (NANA), 27-nor-5β-cholestane-3α,7α,12α,24,25-pentol (CP), and cortisol sulfate (CS)—and three clinical variables (age, race, smoking) were integrated by Lasso-regularised logistic regression into a uLCI score. The model was developed under 10-fold cross-validation in the NCI-Maryland (NCI-MD) cohort (n=845; 470 controls, 375 cases, stages I–IV) and applied without refitting to the independent Colorado Lung Cancer Cohort (n=488; 211 controls, 277 cases). Analyses were prespecified; reporting followed TRIPOD+AI.
Findings
uLCI achieved an area under the curve (AUC) of 0·906 (95% CI 0·887–0·926) in NCI-MD and 0·748 (0·701–0·793) in the independent Colorado cohort. Scores rose monotonically across stages in both cohorts (Spearman ρ=0·69 and 0·45; both p<0·0001). Stage-specific discrimination was preserved from stage I to IV (NCI-MD 0·900–0·927; Colorado 0·722–0·843). Net reclassification improvement over clinical variables was 1·24 (1·14–1·36) and 0·74 (0·56–0·90). uLCI tertiles stratified post-resection survival in stage I–II disease (adjusted hazard ratio 2·03, 1·26–3·27).
Interpretation
uLCI is an independently validated, interpretable urinary risk index that detects lung cancer across all stages, with monotonic stage progression and post-resection prognostic value. Its false-positive rate compares favourably with published estimates for LDCT and cell-free-DNA assays, supporting prospective head-to-head evaluation as a non-invasive triage tool, including in screening-ineligible populations.
Funding
Intramural Research Program, Center for Cancer Research, National Cancer Institute, US National Institutes of Health.
Research in context
Evidence before this study
We searched PubMed, Embase, and Web of Science from Jan 1, 2000, to Jan 31, 2026, without language restriction, for biomarker-based diagnostic or risk models for lung cancer, using “lung cancer”, “early detection”, “biomarker”, “urine”, “metabolite”, “machine learning”, “TRIPOD”, and “validation in an independent cohort”. Five-year survival exceeds 60% at stage I–II but falls below 10% after metastasis. Low-dose CT reduces mortality in heavy smokers but carries an approximately 29% false-positive rate and excludes never-smokers, who account for 10–25% of US lung cancers and up to 40% globally; blood-based cell-free DNA and methylation assays report rates near 27%. Most prior urinary-metabolite work, including our 2024 creatine-riboside and N-acetylneuraminic-acid report, used case–control designs without a locked, integrated model, and—even where two cohorts were analyzed—lacked prespecified independent validation or TRIPOD+AI-compliant reporting; an interpretable urinary index integrating an expanded metabolite panel with clinical variables under these standards had not been described.
Added value of this study
Across two cohorts (1333 individuals), we developed and independently validated uLCI, an interpretable Lasso-regularised logistic index combining four urinary metabolites (CR, NANA, CP, CS) with age, race, and smoking, reported to TRIPOD+AI standards. The prespecified locked model achieved an AUC of 0·906 (95% CI 0·887–0·926) in NCI-MD development and 0·748 (0·701–0·793) in the independent Colorado cohort without refitting, with discrimination preserved from stage I to IV (0·900–0·927; 0·722–0·843). uLCI rose monotonically with stage in both cohorts (Spearman ρ=0·693 and 0·449; both p<0·0001), improved reclassification over clinical variables (net reclassification improvement 1·24 and 0·74), and independently stratified post-resection survival in stage I–II disease (adjusted hazard ratio 2·03 in NCI-MD; 3·81 in Colorado). On indirect benchmarking, the 17–19% false-positive rate was below published estimates for low-dose CT (∼29%) and DELFI (∼27%), and discrimination was preserved in never-smokers — whom low-dose CT excludes — and across racial subgroups in the diverse development cohort.
Implications of all the available evidence
A locked, independently validated, non-invasive urinary index that is interpretable, accurate from stage I, stage-responsive, and prognostic after resection addresses a defined detection gap, with a plausible role as a low-cost triage test that raises pre-test probability before imaging and extends risk assessment to screening-ineligible never-smokers, complementary to low-dose CT. Prospective screening-cohort evaluation (planned in the NCI PLCO and Southern Community Cohort biobanks), head-to-head comparison with blood-based assays, recalibration to screening prevalence, and replication in diverse cohorts are the warranted next steps.
