Research Square · Nov 5, 2026
A seven-state latent progression model of lung cancer: calibration against SEER incidence, state-resolved readouts, and intervention projections
By wei q, ma y, zhu g
Background Lung cancer incidence is the observable end of a multistage process, yet most natural-history models reduce that process to three to five states. We tested the capacity of a seven-state latent progression model to reproduce population incidence, resolve state-specific readouts, and project interventions. Methods We constructed a nine-state model (H, L1-L7, D) with transitions dependent on age, sex, smoking, and chronic kidney disease. The model was calibrated against SEER 2012–2016 age-specific incidence by sex at ages 40–84, with published readouts and precursor-lesion progression as priors. We compared one-, three-, and seven-state models using AIC, BIC, and leave-one-out cross-validation, profiled parameter identifiability, and simulated screening and smoking cessation. Results The seven-state model reproduced SEER incidence (leave-one-out log-RMSE 0.0775; AIC 34.3) and outperformed the one-state (AIC 84.8) and three-state (AIC 46.0) models. The L6-to-L7 rate was 0.143 per year (95% CI 0.081–0.250). State occupancy moved from L1 to L6 with age and smoking; at age 60, current smokers had 3.3% in L6 and 1.36% in L7, compared with 0.10% and 0.03% in never smokers. Incidence data did not identify the individual upstream rates. A static 14-readout panel did not resolve L5; two challenge-recovery readouts raised L5 recall from 10.2% to 87.3%. A mitochondrial-efficiency checkpoint, E(a) = 1/(1 + exp((a − 85)/4)), reproduced the 85 + turnover (329.0 versus 308.2 per 100,000) while the base L6-to-L7 rate continued to rise with age. A four-ring product approximation fitted the age curve worse (AIC 113.1, leave-one-out CV 0.383), showing that the product is an individual-level variable and that population aggregation loses the threshold distribution. A first-four-ring composite (albumin, prealbumin, Treg, MDSC, IL-6, MCC, MUC5AC, LDH, LOXL2, skin potential, PX) achieved AUC 0.923 for L5 + and 0.944 for L6+; the full 18-readout panel achieved 0.991 and 0.990. Screening did not reproduce the 20% NLST mortality reduction below 41.5% downstaging. Cessation did not reproduce the age gradient of the UK targets. Conclusions Single static readouts do not define any of the nine states. The nine-state negative results do not falsify the seven-link chain. The four-ring product is not identified by population incidence. The first-four-ring composite does not separate L5 individuals. The first-four-ring composite does not replace methylation. Clone detection does not replace first-tier triage. Individual causality is not proven.