A Clustering-Based Grey Modeling Approach for Tacrolimus Concentration Prediction Under Sparse Therapeutic Drug Monitoring

BACKGROUND: Tacrolimus therapeutic drug monitoring after liver transplantation is characterized by significant interindividual variability and sparse, irregularly sampled concentration data, which limits the applicability of conventional pharmacokinetic and data-intensive modelling approaches. METHODS: We developed a hierarchical prediction framework integrating density-based spatial clustering of applications with noise-derived patient stratification with self-memory algorithm-based nonlinear grey Bernoulli model (SA-NGBM) using retrospective data from 129 liver transplant recipients. Patients were stratified into homogeneous subgroups using routinely available clinical indicators, and cluster-specific SA-NGBM models calibrated on representative patients were used to predict subsequent tacrolimus trough concentrations. Performance was further evaluated in an independent same-center validation cohort of 60 patients. RESULTS: In the development cohort, the overall mean absolute relative prediction error for the next tacrolimus concentration decreased from 41.4% with the nonclustered baseline to 21.2% with the clustered SA-NGBM framework. In the independent validation cohort, consistent performance gains were observed, with the mean absolute relative prediction error decreasing from 56.8% to 27.3% in the largest patient subgroup. Full longitudinal concentration profiles were required for only 4 representative patients. CONCLUSIONS: Overall, this clustered SA-NGBM framework reduces prediction error under sparse and irregular therapeutic drug monitoring conditions and provides a data-efficient stratified modelling strategy. However, further refinement and prospective validation are required before clinical implementation.