A scoping review of artificial intelligence in living donor liver transplantation - current status and untapped potential

Living donor graft allocation

While the majority of living donor liver transplantations are ‘directed’ (pre-existing donor-recipient relationship with an intended recipient of the donation), there has been a growing number of non-directed living liver donors^[3]. Such precious altruistic donations call for the best utility-based allocation; however, there are yet to be any standardised evidence-based approaches.

An American study developed three separate ML random forest (RF) survival models to predict 10-year graft survival for right-lobe, left-lobe, and left-lateral graft transplantations, covering a total of 6,328 recipients^[4], as summarised in Table 1. Decision trees were used to determine which variables influenced graft survival; an importance measure of the variables was then subsequently calculated by determining how the accuracy was affected with exclusion of the variable in the model. The graft-to-recipient weight ratio (GRWR) as a variable was notably not included in this study, and donor-to-recipient body surface area (BSA) was used instead as a surrogate data available from the Organ Procurement and Transplantation Network (OPTN) dataset. Variables identified as relevant in graft survival prediction across all three models included donor-to-recipient BSA, malignant pathology, location at time of transplant [intensive care unit (ICU) or non-ICU], and having moderate ascites. The concordance index (C-index) was 0.70, 0.63, and 0.61 for left-lateral, left, and right-lobe grafts, respectively.

While this model is a step toward guiding graft allocation, this is a complex matching process that also varies in the “real-world” context according to an array of practical variables, such as individual transplant centre programmes, LDLT volume and frequency, and the characteristics of local waiting lists. Further ML analyses with a wider dataset may shed light on optimal allocation recommendations.

Graft steatosis assessment

Hepatic macrovesicular steatosis is a significant factor for declining grafts for transplantation^[5], as grafts with more than 30% steatosis have been shown to have a higher risk of graft failure and early allograft dysfunction^[6,7]. Careful donor selection is important for recipient outcomes, but overly stringent criteria may unnecessarily restrict an already limited pool of donors. A number of studies using mainly RF survival or convolutional neural network (CNN) methods have addressed this clinical issue, as summarised in Table 2.

Lim et al. developed ML models using only non-invasive variables from 1,652 donors for risk prediction of hepatic steatosis. The model aimed to discriminate between donors with ≤ 5% macrovesicular steatosis who would not require further invasive liver biopsy, from those with > 5% who would otherwise require further investigation^[8]. Binary outcome variables were analysed using regularised discriminant analysis (RDA), flexible discriminant analysis (FDA), and mixture discriminant analysis (MDA). Three separate ML algorithms were also used: RF, support vector machine (SVM), and deep neural network (DNN).

The RF model showed the highest accuracy of 100.0% {area under receiver operating characteristic curve (AUROC) 1.00 [95% confidence interval (CI) 1.00-1.00]} (P < 0.01) in the training set, while FDA achieved the highest accuracy in the test set at 80.9% [AUROC 0.87 (95%CI 0.83-0.90)] (P < 0.01). The results may be limited by a lack of external validation, particularly as the results were obtained retrospectively from a single centre. However, the models were developed from a relatively larger sample size, incorporating non-invasive variables including Hounsfield unit on computed tomography (CT) scans that may spare potential donors from avoidable invasive biopsies, and allow a greater number of potential candidates to be considered.

The Banff consensus recommendations were developed to improve standardisation of steatosis assessment on donor livers^[9]; however, there remains significant inter-observer variation in clinical application^[10]. Existing histological techniques may also affect interpretation of steatosis; artefactual white spaces on frozen section preparations mimic steatosis and may lead to overestimation of steatosis^[11]. Given these challenges, four studies developed AI algorithms for steatosis prediction from histological interpretation^[11-14]. Gambella performed analysis with the largest sample size of 292 biopsies and produced the first published deep-learning algorithms (DLAs) for steatosis assessment based on Banff recommendations. The CNN model showed promising concordance with pathologist assessment with an intraclass correlation coefficient of 0.93 (P < 0.001) for the final steatosis score that combined both high-power and low-power assessment.

It was also noteworthy that pathologist assessment tended to overestimate steatosis across the different studies, including both CNN models by Gambella and Sun^[11], and the computer vision artificial intelligence (CVAI) model by Narayan et al.^[12]. Furthermore, two multivariable logistic regression models found that CVAI steatosis scores were significantly associated with early allograft dysfunction with adjusted odds ratio (OR) 1.34 (95%CI 1.03-1.79, P = 0.03), whereas the average of pathologist steatosis scores were not (adjusted OR 1.02, 95%CI 0.99-1.05, P = 0.13) (but may be affected by underpower of the study)^[12]. Nevertheless, given that steatosis interpretation is mainly affected by observer and artefactual errors, the AI model's accuracy may be attributable to the segmentation framework, which enabled individual liver droplets to be identified with greater precision. This is a promising finding, suggesting that the application of AI may lead to greater accuracy in histological analysis and allow a greater number of eligible donors to be accepted.

Living donor graft weight estimation

A number of studies investigated applications of AI for pre-operative graft volumetry and weight estimation as summarised in Table 3. This carries particular clinical significance for adult LDLT in which accurate prediction is important for both donor and recipient - small-for-size syndrome is associated with graft malfunction^[15], while insufficient donor remnant may threaten donor safety.

The ideal GRWR of at least 0.8 or graft volume standard liver volume ratio of at least 0.4 is recommended^[16]. However, the calculation is dependent on accurate liver segmentation and volumetric graft weight estimation. AI and DL or ML have the potential to improve the accuracy of graft volumetry and graft volume-to-weight estimation, and further improve transplant outcomes.

Giglio performed the largest study using data from 872 living donors of right-lobe, left-lobe, or left-lateral sector grafts^[17]. Supervised ML models were trained to predict graft weight and compared with currently available methods, such as multiplying graft volume by a conversion factor of 1.19 mL/g^[18] or using a correlation coefficient of 0.85^[19] for right lobes. ML showed better graft volume estimation than existing methods; the mean absolute error (MAE) was 50 ± 61 g, which was lower than that of the Lemke model (63 ± 53 g; P < 0.001)^[20] and the Yoneyama model (62 ± 63g; P < 0.001)^[19]. Graft weight prediction by ML also showed higher precision for both right and left lobes. For right-lobe grafts, 76% of predictions had absolute errors ≤ 10%, which was superior to the Lemke model (69%; P < 0.001) and the Yoneyama model (70%; P = 0.001). However, predictions for left-lateral grafts were not significantly improved with ML.

The superiority of ML prediction could be explained by additional information gained from donor characteristic data. Existing methods assume linear relationships between graft volume and weight; however, ML may better capture the non-linear impact of donor body mass index (BMI) and age in relation to hepatic steatosis^[21] and actual donor liver density.

Park performed analysis with a focus on automated liver segmentation for right lobe grafts^[22]. Portal venous phase CT data was analysed with 30 donors in the development group with an experienced radiologist for training purposes; the data were then subsequently analysed using GoCDSS software (SmartCareworks Inc., Seoul, Korea) which applied a DLA^[23] for automated liver segmentation. DLA-generated segmentation showed error requiring correction in 28.6% of donors, but was associated with only a small change in volume [Bland-Altman 95% limits of agreement (LOA), 0.05% ± 3.0% of measured liver volume]. Multivariable linear regression analysis was performed to evaluate any confounding effect of hepatic steatosis on graft weight, and a conversion formula for graft weight from CT-measured graft volume was developed using univariable linear regression analysis. Hepatic steatosis was found not to have a significant effect on graft weight in this cohort (coefficient -0.34; P = 0.67) and thus a formula was developed without excluding donors with steatosis. However, few donors in the data had steatosis, which may affect the generalisability of the findings to moderate or severely steatotic donors. The validation group was subsequently used to analyse agreement between estimated and measured graft weights, and the concordance correlation coefficient was 0.83 (95%CI 0.80-0.86, P < 0.001). This study demonstrated benefits in the use of DL to produce highly accurate segmentation without requiring any additional correction in 70% of donors with an average processing time of 1.8 min.

While Park’s algorithm still required manual division of the liver into left and right lobes along Cantlie’s line, Yang et al. developed a fully automatic model with 248 patients^[24]. A UNETR (U-Net Transformer) model was adopted which used a CNN to learn from contextual input (transformer encoder) and predict liver segmentation output (decoder)^[25]. The model achieved a slightly higher accuracy in right lobe graft estimation with a correlation coefficient to actual graft weight of 0.86 [percentage absolute error (PAE) 8.6% ± 4.7%].

Another DL-based 3D residual U-Net model developed by Oh et al. with 114 donors demonstrated better volume estimation than manual assessment by surgeons^[26]. The DL model showed higher correlation to actual graft weight with an R-squared value of 0.76 (P = 0.02) compared with 0.68 (P = 0.04) for the surgeon. While conventional estimation techniques tend to overestimate graft volume due to the assumption of the same density of liver as water, this model interestingly tended to underestimate graft volume by an average of 1.8%. However, the relatively high level of accuracy may still prove clinically beneficial, particularly for avoiding small-for-size graft syndrome.

Kavur et al. further evaluated 12 DL models: 6 semi-automatic and 6 fully-automatic algorithms, but on a comparatively smaller group of 20 donors (8 training and 12 test datasets)^[27]. The automatic models outperformed semi-automatic methods; ensemble analysis was further developed for fusion of individual results - majority voting (MV) and simultaneous truth and performance level estimation (STAPLE) algorithms were combined as inputs in a fusion approach to yield even higher levels of accuracy. The best model was the fully-automatic DL model Istanbul Technical University (ITU)_U-Net that scored 79.6.

Given the highly imaging-based and mathematical nature of graft weight estimation, this is an area that is likely to benefit from AI input; however, expansion with larger datasets across different transplant centres and donor ethnicities and characteristics would likely enhance the generalisability and accuracy of these algorithms.

Prognostication - post-transplantation hepatocellular carcinoma recurrence risk

Fewer studies have been performed on long-term survival outcomes after LDLT. These analyses have particular relevance in the setting of LDLT; expert centres may perform LDLT for patients beyond traditional transplant criteria, which are more strictly adhered to in deceased donor transplantation waiting lists. While expanded criteria allow more patients to benefit from LT, AI models may provide clinical guidance on recurrence risks, to better inform patients, donors, and clinicians on post-transplant outcomes. Studies addressing these outcomes are summarised in Table 4.

Altaf et al. described a deep-learning model to determine the risk of HCC recurrence after LDLT^[28]. A multilayer perceptron model was used to analyse 192 patients who were categorised into low-risk and high-risk post-transplant recurrence groups. Traditionally, tumour dimensions and alpha-fetoprotein (AFP) levels have been used to assess recurrence risk - this model allowed a more dynamic risk assessment by assigning quantitative weight based on Shapley scores. Two AI models were developed; one with pre-transplant factors, and the second incorporating both pre-transplant factors and tumour grade. The AI model using pre-transplant factors had a training and validation area under the curve (AUC) of 0.86 and 0.71, respectively, and accuracy, precision, and F1 scores of 78.9%, 50%, and 55.6%, respectively. The AI model using both pre-transplant factors and tumour grade had training and validation AUCs of 0.87 and 0.77, respectively, and accuracy, precision, and F1 scores of 84.2%, 60%, and 66.7%, respectively, showing better prediction results compared to using pre-transplant factors alone.

Notably, the deep-learning model also showed that response to downstaging (or conversion therapy) was the third most important prognostic factor for HCC recurrence after tumour diameter and AFP levels. This implies that patients beyond traditional criteria may benefit from neoadjuvant therapy to achieve better survival outcomes^[29-35].

Survival outcomes

There remains a lack of a precise model to predict post-transplantation survival. The Model for End-stage Liver Disease (MELD) score has traditionally been used as a prognostic indicator to assess severity of chronic liver disease, but it has limitations in precise estimation of survival and does not capture non-hepatic factors that may affect outcomes^[36-38].

In light of these challenges, Raji et al. described a prognostic deep-learning model to analyse long-term outcomes comparing living donors with deceased donors^[39]. Data was collected from the United Network for Organ Sharing (UNOS) database including 6,180 LDLTs and 135,709 deceased donor transplantations. Data was used to train the DL model Deeplearning4jMultilayer Perceptron (Dl4jMLP) classifier for survival prediction using 23 top-ranked attributes for survival prediction, including donor age, gender, and BMI, recipient MELD/Paediatric End-stage Liver Disease (MELD/PELD) score, ascites, and cold ischaemic time. For the living donor dataset, the model obtained sensitivity, specificity, and accuracy scores of 99.9%, 99.9%, and 99.91%, respectively. The MAE was 0.0009, root mean squared error (RMSE) 0.021, relative absolute error (RAE) 0.26%, and root relative square error (RRSE) 4.99%. The model output was close to actual survival data, which has proved useful in comparison to using the MELD score alone.

A comparison study between LDLT and DDLT was also performed by Gangadhar et al., but with the use of the ML algorithm termed decision path similar matching (DPSM)^[40]. Their data was collected from 72,581 recipients from the Scientific Registry of Transplant Recipients database, including 68,108 DDLT and 4,473 LDLT. DPSM was used for matching to minimise differences between the two groups, and demonstrated a mean 10-year survival gain of 10.3% over DDLT, with the greatest benefit for primary sclerosing cholangitis (12.4% ± 5.3%) and hepatitis C (12.1% ± 4.7%) compared to other aetiologies.

Giglio et al. performed a study using data from the European Liver Transplant Registry (ELTR) to analyse early graft failure within 3 months following adult LDLT; both least absolute shrinkage and selection operator (LASSO) and artificial neural network (ANN) algorithms were investigated^[41]. The ANN showed better performance; after training from the data of 2,073 transplantations, the final ANN showed an AUC of 0.70 at cross-validation, and 0.69 (95%CI: 0.63-0.75) on the test set. The model was also able to stratify recipients into groups with different risks of early graft failure; predicted risk scores were 6.5%, 29%, and 37% for failure risks < 5%, ≥ 30%, and 40%, respectively. However, this model did not capture graft failure risks related to variables pertaining to graft anatomy, including the additional risks associated with multiple biliary or arterial anastomoses and venous reconstructions. Despite these limitations, the study incorporated MELD and Child-Pugh scores, and may be a step toward providing both the clinical team and potential donors with data that may guide their decisions.

Prediction of tacrolimus concentration

Calcineurin inhibitors, particularly tacrolimus, are an important mainstay for immunosuppression; however, their narrow therapeutic index requires careful titration to avoid both toxicity with overexposure and risk of graft rejection in underdosing^[42]. Furthermore, pharmacological studies have produced recommendations for adult patients, but less evidence has been available for paediatric recipients^[43]. Tacrolimus samples have also shown that target range concentration is difficult to achieve in vivo; out of 1,827 plasma samples from 163 infant patients, only 32% were within the target range, and more than half were below the range^[44]. As such, Song et al. developed ML models aimed at prediction of tacrolimus concentration for better clinical outcomes, as summarised in Table 5.

Multivariate regression analysis was used to identify CYP3A5 genotypes in both recipient and donor, GRWR, BMI, urea, haemoglobin, and albumin as statistically significant variables. Different ML models were subsequently developed with covariate combinations selected to give the lowest MAE. Their best performing model was the Ridge model with an ideal rate (predicted value with relative error ≤ 30%) of 55.05% (P < 0.05). While further development may be required to improve predictive accuracy, especially with other potential factors such as drug-drug and drug-food interactions, their work could be an important foundational step in optimising post-transplant immunosuppressant titrations.

Immunosuppression tailoring and operational tolerance

Immunosuppression is key to graft and patient survival, and calls for a delicate balance between prevention of rejection and avoidance of side effects and de novo or recurrent malignancy.

Calcineurin inhibitors, such as tacrolimus, have been the mainstay of immunosuppression, but previous studies have shown that higher levels of exposure are associated with a greater risk of HCC recurrence^[45-47], while mammalian target of rapamycin (mTOR) inhibitors, such as everolimus and sirolimus, may reduce recurrence rates^[48]. Common practice is to minimise tacrolimus exposure through combination regimens with mTOR inhibitors, but studies have yet to demonstrate clear benefits from this combination^[49]. This is an important area of post-transplant management where AI could be used for further analysis to tailor and optimise regimen choices, thereby maximising graft and patient survival outcomes.

Taken a step further, operational tolerance - complete and safe withdrawal of immunosuppressants after transplantation without graft loss - is a potential area of study that could lead to further improvement in survival outcomes. This holds particular potential in liver transplantation due to its immunologically privileged nature compared to other solid organ transplantations^[50]. Prior studies have investigated clinical parameters and biomarkers that may predict candidates for tolerance, such as allograft gene expression patterns, serum immune exhaustion markers, and histological allograft scores; however, much of the evidence remains conflicting^[51]. Indeed, regulatory T cells and complex individualised immunomodulatory expression have already been elucidated as integral in the role of tolerance. However, the complex relationships with inter-recipient variation bring the area into the realm of personalised medicine, where AI analysis might prove beneficial with larger pooled datasets.