Comparison of Machine Learning Algorithms and Bayesian Estimation in Predicting Tacrolimus Concentration in Tunisian Kidney Transplant Patients During the Early Post-Transplant Period.

BACKGROUND AND OBJECTIVE

Model-informed precision dosing (MIPD), based on a Bayesian approach and machine learning (ML) algorithms, is a suitable approach to personalize dosage recommendations and to improve the concentration target attainment for each patient. The objective of this study is to compare the predictive performance of two ML approaches, XGBoost and LSTM, with a previously developed Bayesian model of tacrolimus (Tac) in a cohort of Tunisian kidney transplant patients during the early post-transplant period (0-3 months) METHOD: This was a cross-sectional study conducted at the Pharmacology department in Fattouma Bourguiba's hospital in Monastir, Tunisia. We included patients who had undergone kidney transplantation in the Nephrology department of Monastir Hospital and received the Tac immunosuppressant protocol, for whom routine therapeutic drug monitoring (TDM) during the early post-transplant period (0-3 months) had been performed in our department.

RESULTS

A total of 187 Tac predose concentration (C) issued from 56 adult renal transplant patients were included in the present study. The whole population was divided into building (n = 39 patients, 119 C) and validation groups (n = 17 patients, 68 C). In the validation dataset, the RMSE was 0.76, 0.19, and 0.01, and the MAE was 0.55, 0.36, and 0.06, respectively, for the Bayesian approach, XGBoost, and LSTM.

CONCLUSION

Our study demonstrates that the LSTM approach outperforms XGBoost and Bayesian estimation in predicting tacrolimus concentration in Tunisian kidney transplant patients. Implementing TDM-based LSTM models during the first PT 3 months in clinical practice can significantly enhance patient outcomes and prevent acute kidney rejection in this population.