Machine learning integration of multi-dimensional fragmentomic features of cell-free DNA for early detection of head and neck squamous cell carcinoma

Head and neck squamous cell carcinoma (HNSCC) is often diagnosed at advanced stages, highlighting the need for accurate, non-invasive detection strategies. In the DECIPHER-HNSC study, we developed an ensemble machine learning model integrating multi-dimensional cell-free DNA (cfDNA) fragmentomic features, including focal copy number variation, fragment size profile, and genome-wide methylation. The model was trained on 144 HNSCC patients and 148 healthy controls, and its performance was validated in two independent cohorts (internal, n = 196; external, n = 193). Through validation, it achieved an area under the curve (AUC) of 0.962 (95% CI: 0.937–0.987) in the internal cohort and 0.966 (95% CI: 0.943–0.989) in the external cohort. At 95% specificity, it yielded sensitivities of 78.0% and 81.1%, respectively, and demonstrated consistent performance across clinical subgroups. Fragmentomic profiles were associated with tumor aggressiveness, and higher prediction scores independently predicted shorter overall survival after adjustment for key clinical covariates, whether modeled using the diagnostic cutoff or as a continuous variable. These findings establish artificial intelligence (AI)-driven cfDNA fragmentomics as a promising approach for early detection and risk stratification in HNSCC, offering potential to advance precision medicine.