More Accurate Early Tacrolimus Trough Predictions with Machine Learning

Predictive Model DevelopmentPredictive External Validation

Authors:Hui Yu; Zihan Qin; Logan S. Smith, et al.

Journal:British Journal of Clinical Pharmacology

PK-anchored ML models for early tacrolimus TDM beat a Bayesian PopPK approach on internal and external data (external MAE 1.69–1.71 ng/mL vs 2.22; R2 ≈0.40 vs 0.21), signaling more reliable forecasts. For transplant pharmacists, this auditable CDS could standardize early dose adjustments and prioritize reviews—pending local validation and supervised pilots.

January 13, 2026

Quick Take

Mechanism-informed ML models (GRU and XGBoost) outperformed a Bayesian PopPK model for predicting early post-transplant tacrolimus troughs on both internal and external testing: external MAE ≈ 1.69–1.71 ng/mL vs PopPK 2.22 ng/mL; external R2 ≈ 0.40 vs 0.21 (models applied to MIMIC‑IV without retraining).

When clinicians’ administered doses fell inside the GRU-recommended range, 51.3% of subsequent troughs were 8–12 ng/mL versus 37.0% overall (p < 0.01). The models are intended as auditable TDM decision support to standardize early adjustments — not to automate prescribing — and require local validation and supervised pilots before clinical use.

Why it Matters

Early post-kidney transplant tacrolimus dosing is high-stakes and unstable: troughs are often drawn before steady state while physiologic variables (haematocrit, albumin, renal/hepatic function, fluid status) change rapidly and no universal dose‑adjustment protocol exists.

Under current practice fewer than 40% of patients reach target troughs at first steady state and it can take 2–3 weeks to reach target exposure, exposing patients to rejection risk when underexposed and to nephrotoxicity, neurotoxicity or infections when overexposed.

For pharmacy services this variability increases workload and inconsistent decisions; a principled, auditable CDS that fits TDM workflows could help standardize early adjustments, support immunosuppressant stewardship, and scale where staffing is limited.

What They Did

Retrospective single‑center cohort (University of Michigan) of 1,624 adult kidney transplant recipients with 4,311 tacrolimus trough concentrations collected during the first 7 days post-transplant.

Extracted time‑stamped EHR doses, LC–MS/MS troughs, demographics, dynamic labs (haematocrit, albumin, creatinine) and concomitant medications; CYP3A genotype data were largely unavailable (≈23% available) and excluded from the final models.

Developed two mechanism‑informed ML models: a GRU that outputs interpretable PK parameters (α and ke) constrained to physiologic ranges, and an XGBoost model using PK‑engineered features; both grounded in a linear one‑compartment, first‑order PK approximation and producing rounded twice‑daily dose suggestions.

Compared models with patient‑level five‑fold cross‑validation (nested tuning), a Bayesian PopPK forecasting approach and purely data‑driven ML variants; performed external validation on an independent MIMIC‑IV cohort (N = 164, 611 troughs). A prototype web tool and code were published for research use.

What They Found

Mechanism‑informed GRU and XGBoost outperformed the Bayesian PopPK model in internal testing (GRU MAE 2.02 ng/mL, R2 0.537 vs PopPK MAE 2.32 ng/mL, R2 0.450) and retained superiority on external validation without retraining (GRU MAE 1.71 ng/mL, R2 0.397; XGBoost MAE 1.69 ng/mL, R2 0.411 vs PopPK MAE 2.22 ng/mL, R2 0.211).

Mechanistic PK constraints helped preserve an approximately linear dose–concentration relationship and improved robustness to distributional shifts: purely data‑driven ML variants showed larger external performance drops (pure GRU R2 0.322; pure XGBoost R2 0.279) compared with the PK‑anchored models.

Dose‑matching (retrospective) analysis: when the clinician’s administered dose fell within the GRU‑recommended range, 51.3% of subsequent troughs were within the 8–12 ng/mL target versus 37.0% overall (p < 0.01); administered doses outside the GRU range hit target only 23.4%. These are associations from retrospective data, not results of prospective dosing by the algorithm.

External cohort characteristics differed (lower doses and troughs, different demographics), yet the mechanism‑informed models maintained better error metrics than PopPK, suggesting improved generalizability in this test — still, local calibration/validation is required before deployment.

Feature‑importance analyses identified prior troughs and dose history, demographic covariates (race, age, BSA) and PK‑derived features (empirical α, remaining drug amount) as key contributors to predictions.

Takeaways

Mechanism‑informed ML could serve as pharmacist‑supervised TDM decision support for early tacrolimus adjustments — a tool to prioritize and standardize dose changes, not to replace clinician judgement or clinical oversight.

Anchoring ML to PK principles preserves a monotonic, interpretable dose–response that pharmacists can sanity‑check (e.g., review estimated elimination rates, confirm proportionality tests, and inspect examples where dose changes produce expected concentration shifts).

Scope is limited: models were trained on trough‑only data for immediate‑release twice‑daily tacrolimus during the first 7 inpatient days. Confirm fit with local trough timing, lab availability and patient mix; monitor subgroup performance (genotype was not included).

External validation results are encouraging but not definitive — treat them as an early signal. Proceed with local validation (silent run / pilot), light‑touch monitoring for bias or unsafe recommendations, and pharmacist oversight in any clinical implementation.

Strengths and Limitations

Strengths:

Rigorous evaluation: patient‑level five‑fold cross‑validation with nested hyperparameter tuning and independent external validation on MIMIC‑IV.; PK‑anchored design that outputs interpretable parameters (α, ke), published code and a prototype web tool improve interpretability, auditability and facilitate evaluation and integration — but do not by themselves make the tool ready for unmonitored clinical use.

Limitations:

Retrospective, single‑center development cohort with sparse, trough‑only sampling (mean ≈ 2.7 measurements per patient in the first week) and exclusion of CYP3A genotype, limiting ability to capture some sources of variability.; Scope limited to adult immediate‑release twice‑daily tacrolimus during the first 7 inpatient days; pediatric patients, once‑daily formulations, IV/other formulations, outpatient titration and prospective clinical outcomes (rejection, nephrotoxicity) were not evaluated.; External validation used MIMIC‑IV (ICU population) which differs from the development cohort; while results were robust in that test, broader multi‑center prospective validation and local calibration are needed before routine clinical use.; The web tool and code are provided for research/illustration only and the study does not address operational maintenance (model monitoring, retraining, EHR integration, governance) that are required for safe clinical deployment.

Bottom Line

Mechanism‑informed ML produced more accurate trough predictions and was retrospectively associated with higher target attainment than a Bayesian PopPK model in internal and external testing; it is a promising, auditable CDS for early tacrolimus adjustments that warrants local validation and pharmacist‑supervised piloting rather than immediate autonomous prescribing.