Validation of online clearance monitoring and machine learning-based prediction of dialysis adequacy in Vietnamese hemodialysis patients: a cross-sectional study.

BACKGROUND: Online clearance monitoring (OCM) offers non-invasive real-time dialysis adequacy assessment, but validation data from Southeast Asian populations are limited. We evaluated OCM performance and developed machine learning models for adequacy prediction in Vietnamese hemodialysis patients. METHODS: This cross-sectional study included 97 maintenance hemodialysis patients at a regional dialysis center in Ho Chi Minh City, Vietnam. Kt/V was measured using OCM (ionic dialysance, Fresenius 4008 S) and calculated using the second-generation Daugirdas formula (reference standard). Machine learning models were developed using OCM-derived Kt/V, age, and body mass index to predict adequacy per Kidney Disease Outcomes Quality Initiative (KDOQI) guidelines (Kt/V ≥ 1.2 or urea reduction ratio ≥ 65%). RESULTS: The median Kt/V was 1.29 (interquartile range [IQR]: 1.18-1.35) by the Daugirdas formula and 1.23 (IQR: 1.12-1.30) by OCM, with 67.0% of patients achieving adequacy targets. OCM underestimated Kt/V by a mean of 0.055 units but correlated excellently with the Daugirdas formula (Spearman ρ = 0.971, p < 0.001). Random Forest regression achieved R² = 0.877 for continuous Kt/V prediction. Logistic regression classification demonstrated perfect test-set discrimination (area under the receiver operating characteristic curve [AUC] = 1.000, sensitivity = 100%), with cross-validation AUC of 0.989 ± 0.010, outperforming direct OCM thresholding (sensitivity 84.6%). CONCLUSIONS: OCM demonstrates excellent correlation with the Daugirdas formula despite systematic underestimation, validating its use for real-time monitoring with appropriate calibration. Machine learning models incorporating patient-specific factors may provide calibrated adjustment to OCM-derived Kt/V. A freely accessible web-based calculator was developed for clinical application, though external validation is needed before widespread implementation.