Integrating Circulating Tumor DNA Dynamics and Molecular Biomarkers for Longitudinal Precision Oncology Monitoring Through Data Driven Cancer Surveillance Models

Abstract

Longitudinal precision oncology requires surveillance methods capable of distinguishing clinically meaningful molecular changes from biological variability, sequencing noise, and transient fluctuations in circulating tumor DNA. Conventional circulating tumor DNA surveillance commonly relies on fixed variant allele frequency thresholds, isolated molecular measurements, or predictive models that inadequately represent temporal dependencies among tumor-derived signals. This study develops a data-driven cancer surveillance framework integrating longitudinal circulating tumor DNA dynamics with molecular biomarkers for early detection of disease progression, treatment response, molecular residual disease, and recurrence. A novel Longitudinal Multimodal Cancer Surveillance Network (LMCS-Net) is proposed to jointly model circulating tumor DNA variant allele frequency, mutant molecules per milliliter, circulating tumor DNA fraction, fragmentomic characteristics, copy-number alterations, methylation signatures, somatic mutation burden, selected protein biomarkers, treatment exposure, and sequential sampling intervals. LMCS-Net combines adaptive molecular-noise suppression, temporal attention, gated biomarker fusion, patient-specific baseline normalization, Bayesian uncertainty calibration, and dynamic risk-state estimation. For patient (i) at monitoring time (t), the multimodal observation vector (\mathbf{x}{i,t}) is transformed into a latent molecular state (\mathbf{h}{i,t}), from which a continuous cancer-surveillance risk score (R_{i,t}\in[0,1]) and molecular progression probability are estimated. Temporal change detection further evaluates (\Delta R_{i,t}=R_{i,t}-R_{i,t-1}), enabling clinically significant trajectory changes to be distinguished from isolated biomarker excursions.
The proposed model is benchmarked against fixed-threshold ctDNA classification, logistic regression, random forest, XGBoost, long short-term memory networks, gated recurrent units, conventional transformers, and Bayesian change-point detection. Performance evaluation incorporates area under the receiver operating characteristic curve, area under the precision-recall curve, sensitivity, specificity, F1-score, Brier score, calibration error, concordance index, false-alert rate, and median lead time for detecting molecular progression before conventional clinical confirmation. Comparative analyses are designed to determine whether multimodal temporal fusion improves sensitivity for low-abundance ctDNA while maintaining robust specificity and calibration. Results are visualized through receiver operating characteristic and precision-recall curves, longitudinal ctDNA-risk trajectories, calibration plots, lead-time distributions, biomarker contribution graphs, patient-level molecular state maps, and comparative algorithm-performance graphs. Ablation and sensitivity analyses quantify the contributions of temporal modeling, fragmentomic features, molecular biomarker fusion, uncertainty calibration, and individualized baselines. The proposed framework establishes an interpretable computational architecture for converting repeated molecular measurements into dynamic cancer-surveillance states and provides a foundation for personalized, datadriven oncology monitoring across treatment and post-treatment follow-up.