Frontiers in Oncology · Oct 30, 2026
Integrating reaction-diffusion tumor models with longitudinal CA-125 trajectories for earlier pre-diagnostic detection of ovarian cancer
By Yun Cao, Na Lu, Jing Shi et al.
Each year more than 200,000 women die from ovarian cancer, and roughly 70% present at advanced stages (III–IV) with five-year survival below 50%. Standard surveillance monitors serum CA-125, yet existing change-detection algorithms discriminate well only near diagnosis (AUC ≈0.95) and lose more than 30% of that performance one to two years earlier, a pattern shared by algorithms that treat CA-125 as a statistical time series rather than as a biophysical process. We present RD-NODE, a physics-informed framework that couples reaction-diffusion tumor-growth equations (Fisher-KPP) with a continuous-time Neural ODE to model CA-125 trajectories. The intended clinical application is pre-diagnostic dynamic risk assessment for incident ovarian cancer among women already undergoing longitudinal CA-125 monitoring, and the primary endpoint is discrimination one year before histopathologic diagnosis. In a de-identified retrospective clinical cohort of 300 histologically confirmed epithelial ovarian carcinomas and 300 matched surveillance controls, evaluated by patient-level cross-validation and transferred without fine-tuning to the independent PLCO screening cohort, RD-NODE achieved an AUC of 0.84 for this endpoint versus 0.80 for the best statistical baseline and extended median detection lead time by 2.4 months. The advantage was largest in the clear cell and mucinous histotypes, in which CA-125 tracks tumor burden least faithfully and conventional surveillance fails most often. Learned proliferation, invasion and CA-125 secretion parameters agreed within 20% of published histotype-specific kinetics, and conformal, physics-tightened intervals supplied calibrated uncertainty alongside every estimate. These retrospective results support prospective evaluation of whether earlier algorithmic alerts improve clinical outcomes.