Machine learning-driven analysis of CMV-DNA levels and associated factors

Cytomegalovirus (CMV) infection is a major cause of morbidity and mortality in immunocompromised populations. Accurate interpretation of CMV-DNA levels is important for early detection, monitoring disease progression, and optimizing antiviral therapy. Machine learning (ML) may facilitate integration of clinical and laboratory data for diagnostic assessment and risk stratification.

To identify determinants of CMV-DNA levels and develop an automated machine learning (AutoML)-based predictive framework integrating clinical and laboratory parameters for CMV risk stratification.

A retrospective single-center cohort of 1,000 patients was analyzed using an AutoML-based framework including demographic, clinical, laboratory, and infection-related variables. Patients were classified into three mutually exclusive categories: CMV-DNA not detected, CMV-DNA detected below the quantification limit, and CMV-DNA quantifiable. Performance was assessed using receiver operating characteristic area under the curve (ROC-AUC), precision-recall area under the curve (PR-AUC), precision, recall, F1-score, and log-loss.

The model demonstrated moderate overall discrimination (ROC-AUC: 0.773). The highest predictive performance was observed for quantifiable CMV-DNA (ROC-AUC: 0.847; PR-AUC: 0.772). Favorable performance was also achieved for CMV-DNA detected below the quantification limit (ROC-AUC: 0.852). Performance was limited for CMV-DNA not detected (ROC-AUC: 0.667; PR-AUC: 0.375), with low recall. Clinical department, primary disease, and age were key determinants.

ML-based models integrating CMV-DNA with clinical parameters may serve as adjunctive decision-support tools for CMV risk stratification. However, limited performance in the CMV-DNA not detected category warrants cautious interpretation and external validation. These models should remain supportive rather than standalone diagnostic systems.