Author(s) :

Blossom Patel, Hangpeng Li.

Volume/Issue :

Abstract :

Ovarian cancer is one of the deadliest gynecological cancers; survival over five years drops sharply, from about 90% when found early to under 30% if detected late. Although ovarian cancer detection based on blood tests that analyze cell-free DNA (cfDNA) could provide a less invasive option, finding consistent signals in the limited data found in blood samples is difficult. To address the high-dimensional nature of this data, the critical challenge of data leakage in machine learning pipelines was investigated. Taken together, combining sophisticated AI methods with highly sensitive methylation tests appears to offer the best chance for developing practical early-detection screening tools. A synthetic dataset was generated that mirrors plasma cfDNA methylation fragments, and a simulation was performed to determine the number of cancerous versus noncancerous differentially methylated regions (DMRs). First, a standard “global” selection model was simulated, which displayed the artificial inflation (overfitting) of accuracy that occurs due to data leakage. Second, a nested-CV elastic-net model was tested to isolate the true biological signal from cfDNA methylation. After modeling this procedure, the leakage-safe model could successfully distinguish early ovarian cancer from non-tumor samples (AUROC 0.938; AUPRC 0.914). This project lays the foundation for future exploration of theory-driven, AI-powered liquid biopsy models.