Using Artificial Intelligence Resources in Dialysis and Kidney Transplant Patients: A Literature Review

Background The purpose of this review is to depict current research and impact of artificial intelligence/machine learning (AI/ML) algorithms on dialysis and kidney transplantation. Published studies were presented from two points of view: What medical aspects were covered? What AI/ML algorithms have been used? Methods We searched four electronic databases or studies that used AI/ML in hemodialysis (HD), peritoneal dialysis (PD), and kidney transplantation (KT). Sixty-nine studies were split into three categories: AI/ML and HD, PD, and KT, respectively. We identified 43 trials in the first group, 8 in the second, and 18 in the third. Then, studies were classified according to the type of algorithm. Results AI and HD trials covered: (a) dialysis service management, (b) dialysis procedure, (c) anemia management, (d) hormonal/dietary issues, and (e) arteriovenous fistula assessment. PD studies were divided into (a) peritoneal technique issues, (b) infections, and (c) cardiovascular event prediction. AI in transplantation studies were allocated into (a) management systems (ML used as pretransplant organ-matching tools), (b) predicting graft rejection, (c) tacrolimus therapy modulation, and (d) dietary issues. Conclusions Although guidelines are reluctant to recommend AI implementation in daily practice, there is plenty of evidence that AI/ML algorithms can predict better than nephrologists: volumes, Kt/V, and hypotension or cardiovascular events during dialysis. Altogether, these trials report a robust impact of AI/ML on quality of life and survival in G5D/T patients. In the coming years, one would probably witness the emergence of AI/ML devices that facilitate the management of dialysis patients, thus increasing the quality of life and survival.


Author
Year Title of study Type of AI/ML algorithm used Type of study & No HD patients Results Nordio, M. 1994 Projection and simulation results of an adaptive fuzzy control module for blood pressure and blood volume during HD Integrative discretetime fuzzy control module for blood pressure and blood volume Observational study 10 HD patients The adaptive control system was simulated on an IBM PC, and rules and terms were expressed by linguistic judgments such as: IF "situation", THEN "action". A pre-processor converted the rules into the numerical values rendered in the tables. The simulation results thus obtained were satisfactory, while the introduction of Na control allowed the achievement of the target dry weight of the patient with a stable blood pressure. Nordio, M. 1995 A new approach to blood pressure and blood volume modulation during HD: an adaptive fuzzy control module Discrete-time fuzzy control module Observational study 10 HD patients A smooth function of volemia acted on the second control variable, Na concentration in the dialysate. The adaptive control system was simulated on an IBM-PC, rules and terms were expressed by linguistic judgments such as: IF "situation", THEN "action". A preprocessor converted the rules into the numerical values rendered in the reference tables. The obtained simulation results were satisfactory, the introduction of the Na control allowed reaching the target dry weight of the patient with a stable blood pressure. Guh, J. Y. 1998 Prediction of equilibrated postdialysis BUN by an ANN in high-efficiency HD ANN was used to predict the equilibrated BUN (Ceq) and equilibrated Kt/V (eKt/V60) by using both pre-dialysis, postdialysis and low-flow post-dialysis BUN.
Observational study 74 patients on high-efficiency or high-flux HD In patients with a high urea rebound (>30%), although Smye formula lost its accuracy, low-flow ANN remained accurate.
In the prediction of eKt/V60, both Daugirdas' formula and lowflow ANN were equally accurate, although the Smye formula was less so.
In patients with a high urea rebound, although both Smye and Daugirdas' formulas lost their accuracy, low-flow ANN remained accurate.
Low-flow ANN can accurately predict both Ceq and eKt/V60 regardless of the degree of urea rebound. Akl, A. I. 2001 AI: a new approach for prescription and monitoring of HD therapy AI using neural networks (NNs) studied and predicted concentrations of urea during a HD session Observational study 15 chronic HD patients Comparing results of the NN model with the DDQ (direct dialysate quantification) model, the prediction error was 10.9%, with a nonsignificant difference between predicted total urea nitrogen (UN) removal and measured UN removal by DDQ. NN model predictions of time showed a non-significant difference with actual intervals needed to reach the same SRI (solute removal index) level at the same patient conditions, except for the prediction of SRI at the first 30-minute interval, which showed a significant difference (p= 0.001). Goldfarb-Rumyantzev, A. 2003 Prediction of single-pool Kt/v based on clinical and HD variables using multilinear regression, treebased modeling, and ANNs Multilinear regression (LM), tree-based modeling (TBM), and ANNs to predict actual spKt/V.
Observational study 602 HD patients Prediction algorithms were developed from a "training" dataset and were validated on a separate "testing" dataset.
Correlation coefficients of predicted spKt/V with measured spKt/V with and without nPNA (total nitrogen appearance) respectively were 0.745 and 0.679 for LM, 0.6 and 0.512 for TBM, and 0.634 for ANN, which performed better without using nPNA. Martin Guerrero, J. D. 2003 Use of ANNs for dosage individualisation of erythropoietin in patients with secondary anemia to chronic renal failure Neural models for dosage individualisation of erythropoietin in CKD patients with secondary anemia Observational study 110 HD patients Neural models carried out an individualised prediction of the erythropoietin dosage to be administered to patients with secondary anemia due to chronic renal failure undergoing periodic HD Since the results obtained were excellent, an easy-to-use decisionaid computer application was implemented. Martin-Guerrero, J. D.

2003
Dosage individualization of erythropoietin using a profile-dependent support vector regression The support vector regressor (SVR) was benchmarked with the classical multilayer perceptron (MLP) and the Autoregressive Conditional Heteroskedasticity (ARCH) model for an individualized prediction of the EPO dosage to be administered to chronic renal failure in periodic HD Observational study

HD patients
The so-called profile-dependent SVR (PD-SVR) displayed improved results of the standard SVR method and the MLP.
The sensitivity analysis on the MLP and inspection the distribution of the support vectors in the input and feature spaces were useful for gaining knowledge about the issue. ANN was seen in the ability to detect a follow-up PCR <1.00 g/kg/day expressed as a percentage of correct predictions, sensitivity, specificity, and predictability.
The inter-institutional performance of the ANN was positively influenced by the size and the variability of the population used to build the mathematical model. For both cases (URR and Kt/V), the minimum doses required to achieve the lowest FPR (false positive rate) and ER (error rate) for the standard methods (stdURR and Kt/Vsp) were higher than those reported by the DOQI guidelines, being 70% for stdURR and 1.35 for Kt/Vsp, whereas for those methods using the double-pool Kt/V or equilibrated URR, the dose targets were close to those recommended by DOQI and ERA.
The method for target dose selection was easy to understand, and it took into account both accuracy and confidence of the adequacy tool.
ANN method was identified to be superior to the Smye method for estimation of equilibrated urea, the results suggesting that ANN methods could be useful tools in the analysis of nephrology data. For a specificity of 50%, the sensitivity of ANNs compared to linear regressions in predicting the erythropoietin dose to reach the hemoglobin target was 78 vs. 44% (p < 0.001).
The ANN built to predict the monthly adaptations in erythropoietin dose, compared with the nephrologists' opinions, allowed to detect 48 vs. 25% (p < 0.05) of the patients treated with an insufficient dose with a specificity of 92 vs. 83% (p < 0.05).
In predicting the erythropoietin dose required for an individual patient and the monthly dose adjustments, ANNs were superior to the nephrologists' opinions. ANN may be a useful and promising tool that could be implemented in clinical wards to help nephrologists in prescribing erythropoietin. Tangri A fuzzy logic (FL) control ran in the system, using instantaneous BP as the input variable governing the ultrafiltration rate (UFR) according to the BP trend.
The system was user-friendly and just required the input of two data: critical BP (individually defined as the possible level of DH risk) and the highest UFR applicable (percentage of the mean UFR).
Sessions with (treatment A) and without (treatment B) ABPS were alternated for 30 dialysis sessions per patient (674 with ABPS vs 698 without).
Mild DH fell non-significantly (-12.3%). There was a similar percentage of sessions in which the planned body weight loss was not achieved and dialysis time was prolonged.
FL may be suited for interpreting and controlling the trend of a determined multi-variable parameter like BP. Gaweda The achieved Hb levels were 12.3 +/-0.6 g/dL for AMP and 11.6 +/-0.4 g/dL for MPC (p < 0.001), mean SDs were 0.75 +/-0.30 g/dL for AMP and 0.60 +/-0.21 g/dL for MPC (p < 0.01), and mean absolute differences from target were 0.8 +/-0.6 g/dL for AMP and 0.3 +/-0.3 g/dL for MPC (p < 0.001). Black race, female sex, winter season, and hypoalbuminemia (serum albumin ≤3.1 g/dl) were the strongest predictors of vitamin D deficiency.

MPC of ESAs
In the validation set, the presence of hypoalbuminemia and winter season increased the likelihood of vitamin D deficiency in black women (from 90% to 100%), black men (from 85% to 100%), white women (from 82% to 94%), and white men (from 66% to 92% These results suggested that data provided to the USRDS can allow for predictive models which have a high degree of accuracy years following the initiation of dialysis. Azar, A. T. 2011 ANN for prediction of equilibrated dialysis dose without intradialytic sample ANNs to predict the equilibrated urea (C eq) at 60 min after the end of HD Observational study

HD patients
The mean urea rebound observed from the ANN was 18.6 ± 13.9%, while the means were 24.8 ± 14.1% and 21.3 ± 3.49% using Smye and Daugirdas methods, respectively.
The ANN model achieved a correlation coefficient of 0.97 (P <0.0001), while the Smye and Daugirdas methods yielded R = 0.81 and 0.93, respectively (P <0.0001); the errors of the Smye method were more significant than those of the other methods and resulted in a considerable bias in all cases, while the predictive accuracy for (eq Kt/V) 60 was equally good by the Daugirdas' formula and the ANN. The numerical approximation for the population equations was based on semigroup theory, respectively on the theory of abstract Cauchy problems.
Modeling Erythropoiesis in Dialysis Patients partial differential equations, was adapted to an individual patient. A standard Least-Squares formulation was used to define the cost-functional used for parameter identification.
The system state was approximated by system states of high order differential equations on finite dimensional subspaces of the state space of the original system.
A low approximation dimension was sufficient for obtaining accurate numerical solutions and estimates for the parameters. Random forest indicated higher performance with AUC of the ROC curve and sensitivity higher than 70% in both temporal windows models, proving that random forests were able to exploit non-linear patterns retrieved in the feature space.
Out-of-bag estimates of variable importance and regression coefficients were used to gain insight into the models implemented.
Malnutrition and an inflammatory condition strongly influenced cardiovascular outcome in incident HD patients.
The most important variables in the model were blood test variables such as the total protein content, percentage value of albumin, total protein content, creatinine and C reactive protein. Also, the age of patients and weight loss in the first six months of renal replacement therapy were highly involved in the prediction.
Among the built models random forests showed the best predictive performance. Chen, W. L. 2014 A rule-based decisionmaking diagnosis system to evaluate arteriovenous shunt stenosis for HD treatment of patients using fuzzy petri nets Rule-based decisionmaking diagnosis system evaluated arteriovenous shunt (AVS) stenosis for longterm HD treatment of patients using fuzzy petri nets (FPNs) Observational study

HD patients
The Burg autoregressive (AR) method was used to estimate the frequency spectra of a phonoangiographic signal and identify the characteristic frequencies.
A rule-based decision-making method, FPNs, was designed as a decision-making system to evaluate the degree of stenosis (DOS) in routine examinations.
The examination results indicated that the proposed diagnosis system had greater efficiency in evaluating AVS stenosis.

2014
Optimization of anemia treatment in HD patients via reinforcement learning Methodology based on reinforcement learning (RL) to optimize ESA therapy.
RL were formulated as Markov decision processes (MDPs) Observational study

HD patients
Simulation results showed that the performance of Q-learning was lower than fitted Q iteration (FQI) and the protocol.
FQI achieved an increment of 27.6% in the proportion of patients that were within the targeted range of hemoglobin during the period of treatment.
The quantity of drug needed was reduced by 5.13%, which indicated a more efficient use of ESAs.
The RL algorithm employed in the proposed methodology was fitted Q iteration, which stood out for its ability to make an efficient use of data. RL could represent an alternative to current protocols. Nigwekar, S. U. 2014 Quantifying a rare disease in administrative data: the example of calciphylaxis Mortality rates among calciphylaxis patients were noted to be 2.5-3 times higher than average mortality rates for chronic HD patients.
By developing and successfully applying this novel algorithm a significant increase in calciphylaxis incidence was identified. Barbieri In CKD patients, Machine Learning (ML) was explicitly taken into account in order to produce a general and reliable model for the prediction of ESA/Iron therapy response.
The ML model makes use of both human physiology and drug pharmacology, yielding Mean Absolute Errors (MAE) of the Hemoglobin (Hb) prediction around or lower than 0.6 g/dl.
ML had predicted improvement based on red blood cell dynamics and drug kinetics.
ML improved the anemia management in dialysis and was suitable for the application in daily clinical practice. Barbieri The analysis revealed a strong association between PTH and phosphate that was superior to that of PTH and Calcium.
RF assumed that changes in phosphate would cause modifications in other associated variables (calcium and others) that may also affect PTH values.
Using RF the correlation coefficient between changes in serum PTH and phosphate was 0.77, p<0.001; thus, the power of prediction markedly increased. The effect of therapy on biochemical variables was also analyzed using this RF. All methods indicated an advantage in at least one area over the traditional paper expert system used by most dialysis facilities.
This study showed improvements in the percentage of hemoglobin measurements within target range, decreased within-subject hemoglobin variability, decreased erythropoiesis-stimulating agent dose, and decreased transfusion rates. pediatric patients Artificial intelligence dry weight was higher (28.6%), lower (50%), or identical to nephrologist dry weight.
Mean difference between artificial intelligence and nephrologist dry weights was 0.497 kg (- 1.33 to + 1.29 kg).
In patients for whom artificial intelligence dry weight was lower than nephrologist dry weight, systolic blood pressure significantly decreased after dry weight decrease to artificial intelligence dry weight (77th to 60th percentile, p = 0.022); anti-hypertensive treatments were successfully decreased or discontinued in 28.7% of cases.
In patients for whom artificial intelligence dry weight was higher than nephrologist dry weight, no hypertension was observed after dry weight increase to artificial intelligence dry weight.
Neural network predictions outperformed those of experienced nephrologists in most cases, proving that artificial intelligence was a powerful tool for predicting dry weight in HD patients. State-of-the-art AI was adopted in other complex decision-making tasks for dialysis patients and may help personalize the multiple dialysis-related prescriptions affecting patients' intradialytic hemodynamics.
A multiple-endpoint predicted session-specific Kt/V, fluid volume removal, heart rate, and BP based on patient characteristics, historic hemodynamic responses, as well as dialysis-related prescriptions.
The accuracy and precision of this preliminary model was encouraging and may be used to anticipate patients' reactions through simulation so that the best strategy can be chosen based on clinical judgment or formal utility functions. The development of novel wearable dialysis devices and the improvement of clinical tolerance will need contributions from new branches of engineering such as AI and ML for the real-time analysis of equipment alarms, dialysis parameters, and patientrelated data with a real-time feedback response.
Emerging technologies derived from AI, ML, electronics, and robotics may offer great opportunities for dialysis therapy, but much innovation is needed before we achieve a smart dialysis machine able to analyze and understand changes in patient homeostasis and to respond appropriately in real time.

Author
Year Title of study Type of AI/ML algorithm used

Type of study & No PD patients
Results Zhang, M. 2005 Selection of PD schemes based on multi-objective fuzzy pattern recognition Multi-objective fuzzy pattern recognition and its application in the selection of PD schemes

Not available
The method was first applied to the field of PD.
The conclusion showed that this method is compliant with doctors' opinions. It provided a new idea for research in this field. At the same time, since the method was simple and easy to use, it can be of wide application. Chen, C. A.

2006
Neural network modeling to stratify peritoneal membrane transporter in predialytic patients ANN model for predialytic stratification of uremic patients on the basis of peritoneal membrane transport status.
Observational study

PD patients
The ANN model demonstrated the usefulness of the model to stratify predialytic patients into H (high average transporters) and L (low average transporters) groups by its significant discrimination (AUC=0.812>0.7) and best fitted calibration (p value of H-statistic=0.421>0.05).
The evaluation of peritoneal membrane transport status helped clinicians offer their uremic patients better therapeutic options.
The ANN model appeared to be a promising tool for stratifying predialytic patients on the basis of peritoneal membrane transport status and helped clinicians make decisions about which dialysis modality was suitable. RT and IBK had the best results compared to all the other algorithms mentioned specially because they were the only algorithms whose Sensitivity, Specificity and Accuracy were higher than 95%.
Using this algorithm would provide the medical staff with a high quality with a low error informed diagnosis. The best classification algorithm was a Generalized Linear Model, which achieved AUC values above 96% using a small subset of the original variables following a feature selection approach.
This approach allowed to increase the interpretability of the combinations of traditional factors, advanced CKD factors and PD factors, all related with a cardiovascular risk profile. Remote patient management (RPM) had the potential to improve outcomes in PD through telehealth platforms that facilitated virtual clinical presence.
RPM enabled patient-generated clinical documentation and feedback mechanism and promoted self-monitoring.
RPM enabled the clinicians to closely monitor and detect early issues, provide feedback in real-time, and initiate early interventions prescription modifications and contextual clinical decision support.
ML and AI algorithms would help detect patterns and predict impending complications (e.g. luid overload, heart failure or peritonitis), allowing early detection to avoid hospitalizations. Table 3.

Author
Year ANN identified a correlation between the selected variables and reduced the initial set of 33 variables to a set of several variables, which were best correlated with chronic rejection (CR) progression.
This method provided the same validity of the identification, although the number of features was reduced. ANN seemed more reliable in the prediction of the CR course than the usual statistical methods. Fritsche, L. The accuracy of the algorithm increased steadily with the size of the available case base. With the largest case bases, the casebased algorithm reached an accuracy of 78 +/-2%, which was significantly higher than the performance of experienced physicians (69 +/-5.3%) (p < 0.001).
The new case-based reasoning algorithm with dynamic time warping as the measure of similarity allowed the extension of the use of automatic laboratory alerting systems to conditions in which abnormal laboratory results were the norm and critical states could be detected only by recognition of pathological changes over time. Stachowska, E. 2006 The The MD diet would be ideal for posttransplantation patients without serious pathologic dyslipidemia. In the case of patients with substantial dyslipidemia, appropriate pharmacologic treatment lowering proatherosclerotic lipid levels should be used in combination with the MD.
Artificial neural networks (ANNs) were a useful tool in modeling biological parameters, showing the dynamics of the studied interactions in a very detailed way. ANN was the most suitable method for investigations with many variables, interconnected nonlinearly; it also allowed for a more general approach to biological problems.
To ensure the predictive power of this method for new cases, the representative database was indispensable, and ANN proved to be a prospective tool for reliable, quick assessments and predictions. Santori, G. 2007 Application of an ANN model to predict delayed decrease of serum creatinine in pediatric patients after kidney transplantation Effectiveness of ANN model to predict a delayed decrease of serum creatinine Retrospective study 107 pediatric kidney recipients The neural network showed sensitivity and specificity for the whole patient cohort were 0.875 and 0.87, respectively, whereas using logistic regression sensitivity and specificity yielded 0.37 and 0.94, respectively.
The neural network model seemed to predict a delayed of decrease in serum creatinine among pediatric kidney recipients.
The availability of the source code may allow the development of stand-alone neural networks to validate this model in prospective studies.

Sharma, D. 2008
An intelligent multi-agent design in healthcare management system The multi-agent approach for the development of complex e-health systems

Not available
The use of intelligent multi-agent approach was important for developing e-health systems, for the prediction of kidney transplant outcomes and the management of chronic diseases such as diabetes.
The kidney transplant outcome prediction was based on the use of a novel classification approach which was a combination of initial data preparations, preliminary classification by ensembles of neural networks, generation of new training data based on criteria of highly accuracy and model agreement, and decision trees. Greco, R. The classification algorithm produced a decision tree that allowed to evaluate the interactions between ARE (acute rejection episode), DGF (delayed graft function), CAN (chronic allograft nephropathy), and BMI on graft outcomes, producing a validation set with 88.2% sensitivity and 73.8% specificity.
This model was able to highlight that subjects at risk of graft loss experienced one or more events of ARE, developed DGF and CAN, or had a BMI > 24.8 kg/m2 and CAN.
The use of decision trees in clinical practice may be a suitable alternative to the traditional statistical methods, since it may allow one to analyze interactions between various risk factors beyond the previous knowledge. CIT was the first approach of more accurate forecasting models for the medication blood level of the respective medication.
The next step could be to add weights, data of rejections or other covariates, which may be able to improve the model even further, but this data was not easily accessible. Decruyenaere, A. 2015 Prediction of delayed graft function after kidney transplantation: comparison between logistic regression and ML methods ML methods in the prediction of delayed graft function (DGF).
Cohort study

KT patients
The discriminative capacities of LDA (linear discriminant analysis), linear SVM, radial SVM (support vector machines) and LR (logistic regression) were the only ones above 80%.
None of the pairwise AUROC comparisons between these models were statistically significant, except linear SVM outperforming LR. The patient donor information in the Excel spreadsheets was processed by a code developed in Visual Basic (VB) language used for Excel macros. The VB code generated all possible 2wise and 3-wise transplantation combinations between patients and donor based on their information (blood group compatibility, age compatibility, etc).
Decision support to use in transplantation centers, facilitating their operations. system A mathematical model developed in Cplex mathematical modeling language used data prepared by VB code and calculated the optimal selection of the transplantation combinations.
Cplex code output the results in the same spreadsheet organ transplantation coordinators used. The decision support system could be modified according to the matching preferences of transplantation centers and it could be used as a simulation tool for analyzing different allocation methods. The integrative analyses (microRNA and gene expression profiling from the same biopsy sample) identified the induction of regulators with demonstrated roles in the downregulation of inflammatory pathways and maintenance of tissue homeostasis in tolerance-induced FCRx (bioengineered stem cell product) samples compared with SIS (standard immunosuppression) samples).
This study highlighted the utility of molecular intragraft evaluation of pathways related to FCRx-induced tolerance and the use of integrative analyses for identifying upstream regulators of the affected downstream molecular pathways. Differentially expressed genes (DEGs) of AR were mainly located on membranes and impacted the activation of receptors in immune responses.
In the PPI network, Src kinase, lymphocyte kinase (LCK), CD3G, B2M, interferon-γ, CD3D, tumor necrosis factor, VAV1, and CD3E in the T cell receptor signaling pathway were selected as important factors, and LCK was identified as the hub protein.
co-citation network and to select the hub protein.
Bioinformatics methods to identify the key biomarker of acute rejection (AR) after kidney transplantation LCK, via acting on T-cell receptor, might be a potential therapeutic target for AR after kidney transplantation. The most accurate model included 3 static covariates of recipients' gender, donors' age and gender as well as 11 dynamic covariates of recipients including current age, time since transplant, serum creatinine, fasting blood sugar, weight and blood pressures available at each visit time.
The performance of prediction models in the validation cohort was improved when history window of time dependent variables' recent values was increased from 1 to 10 (an MSE decline from 161 to 99). The best performed model was able to dynamically predict a future eGFR value for kidney recipients' upcoming visits.