Key words : Artificial intelligence, Graft rejection, Machine learning, Outcome prediction

Introduction

Liver transplant (LT) is considered the only available therapeutic option for many patients with end-stage liver failure. Still, over half of patients that undergo LT will have at least 1 significant complication during the first year posttransplant. In this context, the criteria for assigning an organ should include the appropriate balance between benefit and utility, in which the concept of “survival benefit” plays a pivotal role.¹ High rates of LT complications could also be attributed to an increase in organs transplanted from marginal donors (eg, older age, asystole donors, donors with expanded criteria) and/or the clinical characteristics of the recipient (eg, age, disease state, comorbidities).

Comorbidity, coined by Alvan Fenstein in 1970, can have several interpretations. More specifically, multimorbidity could be defined as the presence of different diseases or conditions that coexist with a main chronic disease.^2,3 Indications for LT have progressively expanded, which, in addition to the admission of older patients on transplant wait lists, have resulted in an growing number of comorbidities among transplant recipients.⁴ This has led to a myriad of studies devoted to determining how specific comorbidities affect LT outcomes, for example, coronary disease,^5-7 chronic kidney disease,⁸ diabetes,⁹ and nonhepatic cancer.¹⁰ However, so far, most of the indicators described to adjust the donor-recipient allocation have not only ignored recipient comor-bidities but share some common flaws, such as limited external validity when applied to populations other than those initially described^11,12 and the use of the logistic regression as the principal statistical tool.^13,14

In addition to conventional methods, the imple-mentation of techniques from artificial intelligence, such as big data analysis and machine learning, may significantly enhance clinical research and subsequently improve clinical practice in the future.¹⁵ Of note, artificial neural networks (ANNs) have emerged as an alternative multivariate analysis method, which could better address the complexity found among patients with different comorbidities.

In this study, we aimed to explore and analyze the predictive value of the recipient comorbidities in liver graft survival within the first year post-transplant by conventional and artificial intelligence methods.

Materials and Methods

Data sources

The study included prospectively collected data from all the patients who received LTs (first transplant) at the Lozano Blesa University Clinical Hospital (Aragón, Spain) from 2010 to 2021. The study complied with Organic Law 15/1999 on Personal Data Protection and was approved by the Aragon Ethics Committee (CEICA, for its acronym in Spanish) (Act 1/2022).

Variables and event

Three large groups of variables were collected (Table 1).¹⁶ Donor data included age, sex, cause of death, and donation after cardiac death or after brain death. Recipient data included demographic and anthropometric characteristics of the recipient. Characteristics of liver disease included list entry code, etiology of liver disease, Model for End-Stage Liver Disease (MELD) score, Child-Pugh score, and recipient comorbidities. Transplant data included surgery variables (time on wait list, cold ischemia time). Evolution variables included patient death. Survival time and cause of death were collected according to the Spanish Liver Transplant Registry (RETH) categorization and the graft function in case of death (death with/without functioning graft, retransplant).

The event under study was loss of the graft in the first year, either by retransplant or death due to any cause with graft dysfunction.

Conventional statistics

For quantitative variables, we obtained central tendency parameters (arithmetic, geometric, and harmonic mean and mode), dispersion measures (standard deviation, standard error, coefficient of variation, range, and variance), and shape measures (kurtosis coefficient or flattening and coefficient of asymmetry). For qualitative variables, we calculated the distribution of responses per group.

We used the Pearson chi-square derived from contingency tables for qualitative variables and the t test or nonparametric test for quantitative variables, depending on their normality distribution.

Artificial neural network

A predictive model was created using an ANN (multilayer perceptron), which included the statis-tically significant variables associated with graft survival according to the RETH report, based on data from 28?609 patients (Cox analysis),¹⁷ and the comorbidities collected with a prevalence >2%.¹⁶

In the exploratory ANN model, the data were randomly divided into training (70%) and testing (30%) samples. The hyperbolic tangent activation function was used in the hidden layer and the softmax function in the output layer. The training parameters were as follows: batch, scaled conjugate gradient as the algorithm, initial lambda 0.0000005, initial sigma 0.00005, and center of interval 0.

Sensitivity analysis of the artificial neural network

A sensitivity analysis, also known as information value (IV) analysis, was performed to retrieve the optimal variables in the construction of the ANN model.¹⁸ An IV value greater than 0.03 was considered clinically important (predictive), values between 0.03 and 0.1 were considered somewhat predictive, and values greater than > 0.1 were considered highly predictive.

Programs

For data treatment, we used the IBM SPSS statistics package version 26.0. For the design and validation of the ANN, we used IBM Neural Network program version 25.0. A Wald P < .05 was considered significant.

Results

General description

The general description of the series studied is shown in (Table 2). Most patients were male (75.5%), the mean age of the patients was 54.8 ± 9.6 years, the main cause of the transplant was cirrhosis (86.7%), and 67.4% of the patients had some associated comorbidities.

Description of groups: event

Graft loss due to retransplant or death with dysfunction occurred in 14% of cases. In (Table 3), the 2 event groups are shown in terms of the variables from the RETH report with regard to statistical significance for liver graft survival. We found that only donor age (52.8 ± 17.5 vs 57.0 ± 17.1 years; P < .05) and liver infection caused by hepatitis C virus (30.7% vs 46.4%; P < .01) were significant between groups.

Description of groups: comorbidities

(Table 4) shows a description of the 2 event groups in relation to all the comorbidities analyzed. Of all the comorbidities studied, only the use of antiplatelet and/or anticoagulant agents (4.5% vs 11.9%; P < .01) and portal thrombosis (30.7% vs 46.4%; P < .01) showed significant relation with the occurrence of the event.

Artificial neural network: information value

From the 6 variables recognized by RETH as independent predictors of graft survival and the 5 comorbidities from our series with a prevalence >2% in both groups, an 11:8:2 ANN was deployed (Figure 1). The sensitivity analysis of the variables included in the network was used to estimate the IV parameter and normalized IV (IVn), as detailed in (Table 5). Of note, almost all the variables showed a certain predictive power, with the highest values for recipient age (IV: 0.159; IVn: 100%) and donor age (IV: 0.132; IVn: 83.1%), followed by 3 morbidities. These were antiplatelet and/or anticoagulant (IV: 0.124; IVn: 78.4%) treatment, previous immunosup-pression (IV: 0.110; IVn: 69.6%), and portal thrombosis (IV: 0.105; IVn: 66.3%). In contrast, the presence of associated hepatocellular carcinoma did not show predictive value in graft survival (IV: 0.019; IVn: 11.8%).

Discussion

Many indicators have been described to predict the probability of liver graft survival; however, they typically underperform when extrapolated to the general population (ie, external validation). In this regard, a recent meta-analysis that included 12 articles and published in 2021¹⁹ reviewed the prediction capacity of the Donor Risk Index, Eurotransplant Donor Risk Index, and Balance of Risk scores. The authors concluded that these 3 scores did not discriminate well between graft loss and survival. In addition, the result of the transplant was mainly influenced by donor age, MELD score, and the etiology of the liver failure.

In 2021, a retrospective cohort of 177 patients in Brazil was also published,²⁰ which evaluated Survival Outcomes After Liver Transplant, Balance of Risk, and Donor Risk indices. The Survival Outcomes After Liver Transplant score, which includes recipient data among its variables, was the only one offering an area under the curve >0.7 (0.73), followed by the Balance of Risk index (0.69). Hence, the authors concluded that the scores that consider data from both the recipient and the donor (Survival Outcomes After Liver Transplant and Balance of Risk) consistently offered more accurate predictions of graft survival.

In light of the above, one of the reasons underlying the poor external validity of these indicators could be the lack of sufficient information from the recipient. In the literature, we can observe the emphasis typically given to the presence of high MELD values^21,22 or to specific pathologies that may incur a contraindication for transplantation²³ and the few studies reporting the impact of recipient multiple morbidities. For example, Volk and colleagues²⁴ proposed a modified Charlson Index to predict mortality at 5 years, Cardoso and colleagues²⁵ found 6 variables related to 5-year mortality, and Tovikkai and colleagues,²⁶ in a study conducted in the United Kingdom, observed that there are 4 factors related to the risk of mortality at 90 days (congestive heart failure, history of extrahepatic malignant disease, cardiovascular disease, and chronic kidney disease). However, only cardiovascular disease was an independent risk factor for mortality in all periods studied (90 days, 1 year, and 5 years).

In a 2020 review about the selection criteria for LT recipients, the authors raised the importance of not only considering the cause of liver failure but also including some comorbidities of the recipient, especially those that are now considered relative contraindications for LT and were previously considered absolute.²⁷ Among these pathologies, the authors stressed the importance of portal venous thrombosis, HIV, and morbid obesity, which were also analyzed in our study. Furthermore, they claimed that, in some transplant candidates (eg, patients over 65 years), it should be carefully assessed whether they present a “favorable comorbidity profile.”

In the Danish Comorbidity Liver Transplant Recipient project, a prospective study initiated in Denmark in 2021,²⁸ many of the comorbidities included in ours were considered. However, rather than analyzing the survival of the graft, the authors monitored the survival of the patient. One of the objectives was to determine whether all these comorbidity factors could be used to develop guidelines for detection, monitoring, and treatment in LT.

In our study, we evaluated how comorbidities can influence graft survival, introducing them into an ANN model with 6 independent risk factors, previously identified through the National Multicenter Registry (RETH, with more than 28?000 patients). Of the 11 variables analyzed, 10 were found to have a certain predictive capacity, and 3 of them were comorbidities. Our model showed a C statistic of 0.745 (95% CI, 0.692-0.798; asymptotic P < .001), which was higher than others indicated in previous works (Figure 2).

Arguably, the ANN built in this study outper-forms other strategies by accounting for outliers and nonlinear interactions between the input variables, so that weak or previously unrecognized relationships between the included parameters emerged. In this manner, ANNs can reveal connections between data sets that would not reach significance using conven-tional statistics, as observed in other studies.^29,30

Nonetheless, our study had some limitations. First, the data were collected retrospectively and from a single center, which can lead to a population bias. Second, the sample is rather small, restraining the prevalence of multimorbidity in both groups. Finally, the sensitivity analysis is based on an ANN model that provides a moderate network perfor-mance (C statistic of 0.745; 95% CI, 0.692-0.798), yet in line with other ANN models in the literature.

Conclusions

When developing an ANN model, apart from including consistent variables (such as donor and recipient age), it is also recommended to add recipient comorbidities. In addition, as presented here, some comorbidities can be highly predictive clinical factors in liver graft survival during the first year.

References:

1. Knight M, Barber K, Gimson A, Collett D, Neuberger J, Liver Advisory Group of National Health Service Blood Transplant. Implications of changing the minimal survival benefit in liver transplantation. Liver Transpl. 2012;18(5):549-557. doi:10.1002/lt.23380
   CrossRef - PubMed
   
2. Bernabeu-Wittel M, Alonso-Coello P, Rico-Blazquez M, Rotaeche del Campo R, Sanchez Gomez S, Casariego Vales E. [Development of clinical practice guidelines for patients with comorbidity and multiple diseases]. Aten Primaria. 2014;46(7):385-392. doi:10.1016/j.aprim.2013.11.013
   CrossRef - PubMed
   
3. Uhlig K, Leff B, Kent D, et al. A framework for crafting clinical practice guidelines that are relevant to the care and management of people with multimorbidity. J Gen Intern Med. 2014;29(4):670-679. doi:10.1007/s11606-013-2659-y
   CrossRef - PubMed
   
4. Ravaioli M, Grande G, Di Gioia P, et al. Risk avoidance and liver transplantation: a single-center experience in a national network. Ann Surg. 2016;264(5):778-786. doi:10.1097/SLA.0000000000001887
   CrossRef - PubMed
   
5. D’Avola D, Cuervas-Mons V, Marti J, et al. Cardiovascular morbidity and mortality after liver transplantation: the protective role of mycophenolate mofetil. Liver Transpl. 2017;23(4):498-509. doi:10.1002/lt.24738
   CrossRef - PubMed
   
6. Snipelisky DF, McRee C, Seeger K, Levy M, Shapiro BP. Coronary interventions before liver transplantation might not avert postoperative cardiovascular events. Tex Heart Inst J. 2015;42(5):438-442. doi:10.14503/THIJ-14-4738
   CrossRef - PubMed
   
7. Skaro AI, Gallon LG, Lyuksemburg V, et al. The impact of coronary artery disease on outcomes after liver transplantation. J Cardiovasc Med (Hagerstown). 2016;17(12):875-885. doi:10.2459/JCM.0000000000000207
   CrossRef - PubMed
   
8. Nair S, Verma S, Thuluvath PJ. Pretransplant renal function predicts survival in patients undergoing orthotopic liver transplantation. Hepatology. 2002;35(5):1179-1185. doi:10.1053/jhep.2002.33160
   CrossRef - PubMed
   
9. Kuo HT, Lum E, Martin P, Bunnapradist S. Effect of diabetes and acute rejection on liver transplant outcomes: an analysis of the organ procurement and transplantation network/united network for organ sharing database. Liver Transpl. 2016;22(6):796-804. doi:10.1002/lt.24414
   CrossRef - PubMed
   
10. Brattstrom C, Granath F, Edgren G, Smedby KE, Wilczek HE. Overall and cause-specific mortality in transplant recipients with a pretransplantation cancer history. Transplantation. 2013;96(3):297-305. doi:10.1097/TP.0b013e31829854b7
    CrossRef - PubMed
    
11. Winter A, Feray C, Audureau E, et al. External validation of the Donor Risk Index and the Eurotransplant Donor Risk Index on the French liver transplantation registry. Liver Int. 2017;37(8):1229-1238. doi:10.1111/liv.13378
    CrossRef - PubMed
    
12. Collett D, Friend PJ, Watson CJ. Factors associated with short- and long-term liver graft survival in the United Kingdom: development of a UK donor liver index. Transplantation. 2017;101(4):786-792. doi:10.1097/TP.0000000000001576
    CrossRef - PubMed
    
13. Feng S, Goodrich NP, Bragg-Gresham JL, et al. Characteristics associated with liver graft failure: the concept of a donor risk index. Am J Transplant. 2006;6(4):783-790. doi:10.1111/j.1600-6143.2006.01242.x
    CrossRef - PubMed
    
14. Braat AE, Blok JJ, Putter H, et al. The Eurotransplant donor risk index in liver transplantation: ET-DRI. Am J Transplant. 2012;12(10):2789-2796. doi:10.1111/j.1600-6143.2012.04195.x
    CrossRef - PubMed
15. Nunez Reiz A, Armengol de la Hoz MA, Sanchez Garcia M. big data analysis and machine learning in intensive care units. Med Intensiva (Engl Ed). 2019;43(7):416-426. doi:10.1016/j.medin.2018.10.007
    CrossRef - PubMed
    
16. La Organización Nacional de Trasplantes. El Registro Español de Trasplante Hepático (RETH). http://www.ont.es/infesp/Paginas/RegistroHepatico.aspx
    CrossRef - PubMed
    
17. Sociedad Española de Trasplante hepático. Memoria de Resultados del Registro Español de Trasplante Hepático. 2020; http://www.sethepatico.org.
    CrossRef - PubMed
    
18. Hong WD, Ji YF, Wang D, Chen TZ, Zhu QH. Use of artificial neural network to predict esophageal varices in patients with HBV related cirrhosis. Hepat Mon. 2011;11(7):544-547.
    CrossRef - PubMed
    
19. Lozanovski VJ, Probst P, Arefidoust A, et al. Prognostic role of the Donor Risk Index, the Eurotransplant Donor Risk Index, and the Balance of Risk score on graft loss after liver transplantation. Transpl Int. 2021;34(5):778-800. doi:10.1111/tri.13861
    CrossRef - PubMed
    
20. Torterolli F, Watanabe RK, Tabushi FI, et al. Bar, Soft and Dri post-hepatic transplantation: what is the best for survival analysis? Arq Bras Cir Dig. 2021;34(1):e1576. doi:10.1590/0102-672020210001e1576
    CrossRef - PubMed
    
21. Grat M, Wronka KM, Patkowski W, et al. Effects of donor age and cold ischemia on liver transplantation outcomes according to the severity of recipient status. Dig Dis Sci. 2016;61(2):626-635. doi:10.1007/s10620-015-3910-7
    CrossRef - PubMed
    
22. Schlegel A, Linecker M, Kron P, et al. Risk assessment in high- and low-MELD liver transplantation. Am J Transplant. 2017;17(4):1050-1063. doi:10.1111/ajt.14065
    CrossRef - PubMed
    
23. Schoening W, Helbig M, Buescher N, et al. Eurotransplant donor-risk-index and recipient factors: influence on long-term outcome after liver transplantation - a large single-center experience. Clin Transplant. 2016;30(5):508-517. doi:10.1111/ctr.12714
    CrossRef - PubMed
    
24. Volk ML, Hernandez JC, Lok AS, Marrero JA. Modified Charlson comorbidity index for predicting survival after liver transplantation. Liver Transpl. 2007;13(11):1515-1520. doi:10.1002/lt.21172
    CrossRef - PubMed
    
25. Cardoso FS, Bagshaw SM, Abraldes JG, et al. Comorbidities have a limited impact on post-transplant survival in carefully selected cirrhotic patients: a population-based cohort study. Ann Hepatol. 2015;14(4):505-514.
    CrossRef - PubMed
    
26. Tovikkai C, Charman SC, Praseedom RK, Gimson AE, van der Meulen J. Time-varying impact of comorbidities on mortality after liver transplantation: a national cohort study using linked clinical and administrative data. BMJ Open. 2015;5(5):e006971. doi:10.1136/bmjopen-2014-006971
    CrossRef - PubMed
    
27. Mahmud N. Selection for liver transplantation: indications and evaluation. Curr Hepatol Rep. 2020;19(3):203-212. doi:10.1007/s11901-020-00527-9
    CrossRef - PubMed
    
28. Thomsen MT, Hogh J, Knudsen AD, et al. The Danish comorbidity in liver transplant recipients study (DACOLT): a non-interventional prospective observational cohort study. BMC Gastroenterol. 2021;21(1):145. doi:10.1186/s12876-021-01733-5
    CrossRef - PubMed
    
29. Andersson B, Andersson R, Ohlsson M, Nilsson J. Prediction of severe acute pancreatitis at admission to hospital using artificial neural networks. Pancreatology. 2011;11(3):328-335. doi:10.1159/000327903
    CrossRef - PubMed
    
30. Hong WD, Chen XR, Jin SQ, Huang QK, Zhu QH, Pan JY. Use of an artificial neural network to predict persistent organ failure in patients with acute pancreatitis. Clinics (Sao Paulo). 2013;68(1):27-31. doi:10.6061/clinics/2013(01)rc01
    CrossRef - PubMed