Journals | Policy | Permission

Journal of Clinical Medicine Research

Review

Volume 15, Number 8-9, September 2023, pages 391-398

Beyond Human Limits: Harnessing Artificial Intelligence to Optimize Immunosuppression in Kidney Transplantation

In the dynamic field of medicine, artificial intelligence (AI) has become a prominent topic of discussion, extending its influence on various disciplines, including transplantation. By equipping machines with cognitive abilities, AI has brought about a revolution in healthcare practices. AI involves the development of computer systems that possess human-like cognitive abilities such as reasoning, problem-solving, and learning. By creating software and systems that mimic human intelligence, AI enables machines to exhibit intelligent behavior. Machine learning (ML), a subset of AI, focuses on developing algorithms and models that allow computers to learn from data without explicit programming [1]. ML algorithms are trained on extensive datasets, enabling them to recognize patterns, make predictions, and continuously enhance their performance. Three broad categories shape the landscape of ML algorithms: supervised learning, unsupervised learning, and reinforcement learning. These categories facilitate the decoding of relationships between input variables and known outputs, the discovery of hidden patterns in unlabeled data, and the iterative refinement of prediction models, respectively.

In the context of solid organ transplantation, numerous challenges persist throughout the transplant process. Allocating organs to suitable recipients, considering factors such as patient demographics, comorbidities, genetics, and graft quality, remains a significant challenge due to the limited supply of donor organs. Additionally, the growing complexity of transplant candidates, including advanced age and associated risk factors, necessitates personalized treatment strategies that optimize immunosuppressive therapy while mitigating the risks of infections, malignancies, and medication-induced side effects [2]. The role of AI on kidney transplant has been extensively reviewed recently, with most of the discussion being on various aspects of AI in kidney transplantation, such as waitlist prioritization, donor-recipient matching, rejection prediction, and post-transplant outcomes [1, 3]. This review article aimed to specifically delve into the optimization of immunosuppression - an essential component of successful kidney transplantation.

Achieving the delicate balance between underimmunosuppression and overimmunosuppression is crucial in kidney transplantation, considering the risks of rejection, infections, and medication-related complications. By integrating AI-powered algorithms, personalized and data-driven immunosuppression protocols can be developed, enabling clinicians to make informed decisions and improve patient care.

Clinicians face challenges in accurately dosing immunosuppressive medications due to multiple drug interactions and narrow therapeutic windows. Predicting the response to therapy can be complicated, impacting graft survival, adverse events, and length of hospital stay. To address these challenges, researchers have employed ML modeling to improve the accuracy of tacrolimus dosing and other immunosuppressive drug dosing after renal transplant compared to clinicians’ decisions. ML models provide a more precise approach to dosing immunosuppressive medications, aiding in personalized treatment for transplant patients.

We performed a comprehensive systematic review of articles focusing on the use of AI or ML on immunosuppression management in kidney transplantation within the last decade, limiting our search to English language publications. We conducted our search across two major electronic databases, PubMed MEDLINE and EMBASE, up to July 1, 2023. Employing a combination of relevant key terms and synonyms, including “kidney transplant”, “artificial intelligence”, “machine learning”, and “immunosuppression”, and aliases, we meticulously screened retrieved articles. This process began with the removal of duplicates, followed by independent abstract assessments by two authors. Discrepancies in article selection were resolved through consensus. Subsequently, selected articles underwent full-text examination to make the final inclusion decisions. Reference lists of included studies were also reviewed for relevant articles. In total, our systematic search identified nine studies meeting our criteria, where ML protocols were employed to adjust conventional immunosuppressive agents in kidney transplant patients. These studies are detailed further, categorized by the specific immunosuppressant under investigation.

In a prospective study with 80 kidney transplant patients at the Oslo University Hospital, Storset et al [4] investigated the use of computerized dosing of tacrolimus in de novo renal transplant recipients. They compared the effectiveness of computerized dosing with conventional dosing by experienced transplant physicians. The results demonstrated that computerized dose individualization significantly increased the proportion of tacrolimus concentrations within the target range as compared to conventional dosing. The time to achieve target levels was also shorter in high-risk patients. Additionally, computerized dosing showed benefits in terms of glucose metabolism and renal function.

One of the earliest studies in this context was probably by McMichael et al [5], who published a paper that presents an evaluation of an innovative dosing system aimed at optimizing FK 506 and prednisone, called the “intelligent” dosing system (IDS), which was developed to standardize patient management and improve patient care. The algorithm utilized stochastic open loop control theory to optimize drug dosing and has been shown to accurately predict plasma levels of FK 506. The IDS was designed to be simple and accurate, with an easy-to-use interface that required no previous computer experience. The system used a knowledge base containing facts and rules to determine the best course of action for dosing. The paper presented dosing examples and observed versus predicted plasma levels to demonstrate the effectiveness of the IDS. A prospective validation study demonstrated that the model achieved a 95% accuracy rate in describing the correlation between FK 506 dosage and FK 506 plasma level. Furthermore, the study found no biases in the dosing predictions. Importantly, the study also confirmed that the dosing predictions made by the model were unbiased, indicating that the model was reliable in providing accurate estimations. The authors conclude that the IDS is a simple and accurate automated drug dosing program that can improve patient outcomes.

Using a dataset from the Vienna General Hospital, Seeling et al [6] developed a knowledge-based system for guiding tacrolimus therapy in kidney transplant patients. The goal of the study was to identify adaptation rules for tacrolimus therapy based on a clinical dataset and integrate them into a clinical decision support system. The authors utilized patient data from 1995 to 2008 collected from the Department of Nephrology and Dialysis of the Vienna General Hospital. The dataset included patient demographics, laboratory parameters, time since kidney transplantation, and other immunosuppressive drugs administered. The researchers used a regression tree to create homogeneous groups of data and developed semi-automated models for these groups to predict the drug concentration for the next ward round. The models were used to create a knowledge base that was integrated into a clinical decision support system for tacrolimus therapy planning to guide nephrologists. The paper also highlights the importance of creating separate knowledge bases for each hospital due to differences in medication methods and views.

In a study with a large Chinese cohort of renal transplant recipients, Tang et al [7] compared the performance of multiple linear regression (MLR) and various ML techniques in predicting the stable dose of tacrolimus. A total of 1,045 renal transplant patients were included in the study, with 80% randomly selected as the derivation cohort and the remaining 20% as the validation cohort. ML models including artificial neural network (ANN), regression tree (RT), multivariate adaptive regression splines (MARS), boosted regression tree (BRT), support vector regression (SVR), random forest regression (RFR), lasso regression (LAR), and Bayesian additive regression trees (BART) were compared with MLR. Among the ML models, the RT model performed the best in both the derivation and validation cohorts, showing higher prediction accuracy compared to MLR. The RT model demonstrated a prediction accuracy of 0.71 in the derivation cohort and 0.73 in the validation cohort. The study highlights the potential of ML models, particularly the regression tree model, in predicting the stable dose of tacrolimus in renal transplant recipients.

Thishya et al [8] explored the use of ANN and logistic regression (LR) models to predict the bioavailability of tacrolimus and the risk for post-transplant diabetes in patients with renal transplantation. The study investigated the role of genetic polymorphisms in ABCB1 and CYP3A5 in predicting the bioavailability of tacrolimus. The ANN model, with five-fold cross-validation, demonstrated a good correlation with the experimental data of tacrolimus bioavailability. Factors such as younger age, male gender, and optimal body mass index were associated with lower bioavailability of tacrolimus. Genetic polymorphisms, specifically ABCB1 1236 C>T, 2677G>T/A, and CYP3A53, were found to be inversely or positively associated with the bioavailability of tacrolimus. Gender bias was observed in association with the ABCB1 3435 C>T polymorphism. Additionally, synergistic interactions between CYP3A53 and ABCB1 2677 G>T/A were identified as determinants of the risk for post-transplant diabetes. The LR model revealed an independent association of ABCB1 2677 G>T/A with post-transplant diabetes. The study employed multifactor dimensionality reduction analysis (MDR) to identify synergistic interactions between CYP3A5*3 and ABCB1 2677 G>T/A as important factors contributing to the risk of post-transplant diabetes. Overall, the ANN and MDR models utilized in this study provide insights into the individual and synergistic effects of variables in modulating the bioavailability of tacrolimus and the risk for post-transplant diabetes in patients with renal transplantation.

In their study, Camps-Valls et al [9] investigated the use of neural networks to personalize the dosage of cyclosporine A (CyA) in kidney transplant patients. They employed different types of neural networks, including multilayer perceptron (MLP), finite impulse response (FIR), and Elman recurrent networks. The researchers created a two-model scheme where the blood concentration predicted by the first model served as input for the dosage prediction model. They trained the models using data from 22 patients for training and tested them on data from 10 patients. The ensemble of FIR and Elman networks demonstrated the best performance, achieving an r value of 0.977 in the validation set. The authors highlighted that neural models are suitable for this task due to their accuracy, precision, and robustness.

In another study, Goren et al [10] discusses the use of the adaptive-network-based fuzzy inference system (ANFIS) to predict CyA blood levels in renal transplantation patients. The model was implemented using therapeutic drug monitoring (TDM) data collected from 138 patients, with 20 input parameters including concurrent use of drugs, blood levels, sampling time, age, gender, and dosing intervals The results of the study showed that the ANFIS model produced eight rules and exhibited a root mean square error (RMSE) of 0.045 with respect to the training data and an error of 0.057 with respect to the checking data in the MATLAB environment. This indicates that the ANFIS model was able to effectively predict CyA concentration in blood samples. The authors concluded that their model could effectively assist physicians in choosing the best therapeutic drug dose in the clinical setting.

Although initially marketed as a fixed-dose drug, mycophenolate mofetil (MMF) faces challenges due to its pharmacokinetic variability, which results in different processing and elimination rates among individuals. Another challenge is the weak relationship between the dose of MMF and the exposure of the body to its active form, MPA. A recent consensus of the international association of therapeutic drug monitoring and clinical toxicology [11], recommended MPA therapeutic drug monitoring by estimating the MPA area under the curve (AUC) to optimize treatment and improve patient outcomes. Woillard et al [12] developed a machine-learning model to accurately estimate the concentration of MMF in transplant patients. The models for estimating the concentration of MMF in patients who have received kidney or heart transplants were developed using extreme gradient boosting (Xgboost R package) ML models. The models were trained on a total of 12,877 MPA AUC from 0 to 12 h (AUC0 - 12 h) requests from 6,884 patients sent to the Immunosuppressant Bayesian Dose Adjustment expert system for AUC estimation and dose recommendation based on MPA concentrations measured at least at three sampling times (about 20 min, 1 and 3 h after dosing). The data were split into a training set (75%) and a test set (25%), and the Xgboost models in the training set with the lowest root mean squared error (RMSE) in a 10-fold cross-validation experiment was evaluated in the test set and in four independent full-pharmacokinetic (PK) datasets from renal or heart transplant recipients. The models were based on two or three concentrations, differences between these concentrations, relative deviations from theoretical times of sampling, presence of a delayed absorption peak, and five covariates (dose, type of transplantation, associated immunosuppressant, age, and time between transplantation and sampling). The authors showed that the model allowed for accurate estimation of the AUC of MPA over a 12-h period. These models can be utilized for routine exposure estimation and dose adjustment of MPA. Furthermore, the researchers plan to implement these ML models in a dedicated web interface for convenient use.

Operative tolerance, a state of long-term allograft acceptance without continuous immunosuppression, is an important tenet for the success of solid organ transplantation that can help minimize exposure of immunosuppressive treatments. In a recent study, Fu et al [13], investigated the identification of potential biomarkers for allograft tolerance in kidney transplantation using ML techniques. The study utilized three publicly available gene expression datasets from peripheral blood lymphocytes of 63 tolerant patients. The researchers compared 14 different ML models to predict spontaneous kidney graft tolerance, and the best subset selection (BSS) regression approach emerged as the most powerful model. It exhibited a sensitivity of 91.7% and a specificity of 93.8% in the test group, as well as a specificity of 86.1% and a sensitivity of 80% in the validation group. Using the BSS model, a feature set comprising five genes (HLA-DOA, TCL1A, EBF1, CD79B, and PNOC) was identified as predictive of allograft tolerance. Furthermore, the downregulation of EBF1 was identified as an independent predictor of graft rejection and graft loss. By employing a two-gene signature (EBF1 and HLA-DOA) as input to their classifier, the researchers achieved an AUC value of 84.4%.

Overall, this study highlights the potential of ML in uncovering gene sets that could influence tolerance to renal allografts. The identified genes, particularly EBF1, hold promise as novel biological targets and may guide patient selection for immunosuppressant withdrawal in clinical practice.

AI and ML algorithms are promising new tools for improving immunosuppressive drug adjustments in kidney transplantation. AI-powered algorithms can consider individual patient characteristics, such as age, gender, weight, genetic polymorphisms, and concomitant medications, to develop personalized dosing regimens. This can help to optimize drug efficacy and minimize side effects. AI algorithms can also analyze large datasets of patient data to identify patterns and relationships that may be difficult to detect by human experts. This can lead to more accurate and efficient drug dosing. Additionally, AI algorithms can help to reduce the risk of human error in drug dosing, which is especially important for complex dosing regimens or for patients taking multiple medications. The integration of AL/ML techniques in the realm of kidney transplantation holds the promise of not only predicting immunosuppressive drug levels but also synergizing with existing quality improvement initiatives [14]. Finally, AI algorithms can be used to monitor patient drug levels and adjust dosing regimens in real-time.

Limitations of most of the studies include data quality, small sample sizes, and inconsistency in the number of cases used for model training, which can affect generalization. Inconsistent data collection and classification may lead to the use of incorrect features and introduce bias. Prospective and external validation of AI models is lacking, and their improvement over traditional methods may be marginal in certain transplantation aspects. ML algorithms can be computationally complex and time-intensive. Interpretability of ML models is a challenge. Ensuring rigorous validation, considering computational resources, and assessing the nature of the data and clinical questions are important. Fairness evaluation and integration of non-clinical variables are necessary for equitable systems. None of the studies conducted so far have been validated on geographically distant cohorts, and they have primarily focused on acute tacrolimus dosing. It remains unclear whether these approaches would be beneficial for longer-term dosing decisions. Further research is needed to determine the effectiveness of these methods in different populations and for extended treatment durations.

In this comprehensive review, we have examined the studies that specifically investigate the use of AI-based algorithms to predict immunosuppressive drug levels (Table 1) [4, 5, 7-10, 12]. However, it is important to acknowledge that tailoring immunosuppression involves considering various factors beyond drug levels. Numerous studies have explored the application of ML techniques in optimizing donor-recipient matching, facilitating the identification and availability of potential donors, predicting allograft function and rejection, and assessing post-transplant survival outcomes. Incorporating all this valuable information will contribute to the personalized approach to immunosuppression in kidney transplant patients. By leveraging AI, we can enhance outcomes, provide more accurate personalized care, and shape the future of kidney transplantation. Further advancements and integration of AI technology are necessary to realize its full potential in optimizing immunosuppression strategies for kidney transplant patients.

Click to view

Table 1. Description of the AI Models Used in the Reviewed Studies

Acknowledgments

ChatGPT, a generative AI tool, was used to assist in the writing process in a few places only to enhance the readability and language of the paper. It was not used for any selection, analysis, or interpretation of the reviewed studies. The authors carefully reviewed and take full responsibility for the article.

Financial Disclosure

None to declare.

Conflict of Interest

The author has no conflict of interest to declare.

Author Contributions

Both authors actively contributed to the selection, reviewing, interpreting the studies, and preparing the manuscript, engaging in writing, and revising the article before its submission.

Data Availability

The authors declare that data supporting the findings of this study are available within the article.

1. Gotlieb N, Azhie A, Sharma D, Spann A, Suo NJ, Tran J, Orchanian-Cheff A, et al. The promise of machine learning applications in solid organ transplantation. NPJ Digit Med. 2022;5(1):89.
   doi pubmed pmc
2. Seyda M, Quante M, Uehara H, Slegtenhorst BR, Elkhal A, Tullius SG. Immunosenescence in renal transplantation: a changing balance of innate and adaptive immunity. Curr Opin Organ Transplant. 2015;20(4):417-423.
   doi pubmed pmc
3. Seyahi N, Ozcan SG. Artificial intelligence and kidney transplantation. World J Transplant. 2021;11(7):277-289.
   doi pubmed pmc
4. Storset E, Asberg A, Skauby M, Neely M, Bergan S, Bremer S, Midtvedt K. Improved tacrolimus target concentration achievement using computerized dosing in renal transplant recipients—a prospective, randomized study. Transplantation. 2015;99(10):2158-2166.
   doi pubmed pmc
5. McMichael J, Irish W, McCauley J, Shapiro R, Gordon R, Van Thiel DH, Lieberman R, et al. Evaluation of a novel "intelligent" dosing system for optimizing FK 506 therapy. Transplant Proc. 1991;23(6):2780-2782.
   pubmed pmc
6. Seeling W, Plischke M, Schuh C. Knowledge-based tacrolimus therapy for kidney transplant patients. Stud Health Technol Inform. 2012;180:310-314.
   pubmed
7. Tang J, Liu R, Zhang YL, Liu MZ, Hu YF, Shao MJ, Zhu LJ, et al. Application of machine-learning models to predict tacrolimus stable dose in renal transplant recipients. Sci Rep. 2017;7:42192.
   doi pubmed pmc
8. Thishya K, Vattam KK, Naushad SM, Raju SB, Kutala VK. Artificial neural network model for predicting the bioavailability of tacrolimus in patients with renal transplantation. PLoS One. 2018;13(4):e0191921.
   doi pubmed pmc
9. Camps-Valls G, Porta-Oltra B, Soria-Olivas E, Martin-Guerrero JD, Serrano-Lopez AJ, Perez-Ruixo JJ, Jimenez-Torres NV. Prediction of cyclosporine dosage in patients after kidney transplantation using neural networks. IEEE Trans Biomed Eng. 2003;50(4):442-448.
   doi pubmed
10. Goren S, Karahoca A, Onat FY, Goren MZ. Prediction of cyclosporine A blood levels: an application of the adaptive-network-based fuzzy inference system (ANFIS) in assisting drug therapy. Eur J Clin Pharmacol. 2008;64(8):807-814.
    doi pubmed
11. Bergan S, Brunet M, Hesselink DA, Johnson-Davis KL, Kunicki PK, Lemaitre F, Marquet P, et al. Personalized therapy for mycophenolate: consensus report by the international association of therapeutic drug monitoring and clinical toxicology. Ther Drug Monit. 2021;43(2):150-200.
    doi pubmed
12. Woillard JB, Labriffe M, Debord J, Marquet P. Mycophenolic acid exposure prediction using machine learning. Clin Pharmacol Ther. 2021;110(2):370-379.
    doi pubmed
13. Fu Q, Agarwal D, Deng K, Matheson R, Yang H, Wei L, Ran Q, et al. An unbiased machine learning exploration reveals gene sets predictive of allograft tolerance after kidney transplantation. Front Immunol. 2021;12:695806.
    doi pubmed pmc
14. Varnell CD, Jr., Margolis P, Goebel J, Hooper DK. The learning health system for pediatric nephrology: building better systems to improve health. Pediatr Nephrol. 2023;38(1):35-46.
    doi pubmed pmc

This article is distributed under the terms of the Creative Commons Attribution Non-Commercial 4.0 International License, which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Journal of Clinical Medicine Research is published by Elmer Press Inc.