Predicting Cytomegalovirus Infection in Seropositive Lung Transplant Recipients: How to Improve the Predictive Performance?

To the Editor: We read with great interest the article by Gardiner et al1 in the May 2025 issue of Transplantation. The study highlights the utility of donor serostatus, absolute lymphocyte count (ALC), and the mitogen component of the QuantiFERON-cytomegalovirus (CMV) assay for predicting postprophylaxis CMV infection in seropositive lung transplant recipients. We congratulate the authors on this excellent study! However, some key aspects deserve further discussion. The authors reported that ALC and mitogen were independently associated with CMV infection, but their combined application did not improve model performance. This limited incremental value can be explained by a complex relationship between the biomarkers. Although the authors mentioned that variance inflation factors (VIFs) were used and concluded no collinearity between the 2 variables in discussion, no VIF values of ALC and mitogen were reported. If the VIF value is indeed low, this unsatisfactory result may arise from the simple linear model’s failure to effectively capture the relationship between variables. If the VIF value is relatively high (even < 10 or 5), the combined indicators may not bring additional useful information because of potential multicollinearity. In addition, the mitogen assesses the activation and proliferation capacity of immune cells by detecting the response of T lymphocytes to mitogen stimulation in biological mechanisms.2 To a certain extent, the level of mitogen is closely related to the quantity and functional status of ALC. To put it simply, both factors reflect immune status and potentially lead to overlapping predictive information.3 The complex relationship between the 2 biomarkers may have led to only a slight improvement in the Cox regression model.4 Compared with the Cox regression model, machine learning algorithms can capture complex nonlinear patterns and interaction effects inherent in the data. This powerful capability can achieve outstanding performance in model fitting based on complex relational data. Our previous study showed that the random survival forests model, a well-known machine learning algorithm, outperformed the Cox regression model by 0.20–0.23 in integrated time-dependent area under the curve for posttransplant survival prediction.5 Furthermore, some advanced machine learning algorithms, such as extreme gradient boosting and gradient boosting decision tree, have shown good performance in many prediction tasks. Thus, machine learning algorithms have the potential to improve the C-statistic of CMV prediction. In conclusion, we congratulate the authors on their significant efforts again and advocate for the further development of tools to enhance predictive performance to personalize CMV management in seropositive lung transplant recipients.