Machine learning model integrating transient elastography and clinical data for prediction of graft fibrosis after liver transplantation

Abstract

BACKGROUND

Graft fibrosis compromises graft survival in liver transplant (LT) recipients. Early detection is necessary for prevention and treatment. Biopsy is the gold standard for fibrosis assessment; but its invasiveness prevents frequent monitoring.

AIM

To develop a novel non-invasive machine learning (ML) tool combining clinical data with liver stiffness measurements (LSMs) from transient elastography (TE) for personalized graft fibrosis prediction.

METHODS

We performed a retrospective, single-center study including 197 adult LT patients with TE measurements matched to biopsies between 2014-2023. Variables included patient demographics, comorbidities, laboratory values, and TE-derived LSM as a quantitative imaging-derived parameter. Biopsy-confirmed fibrosis stages were dichotomized as significant (≥ F2) vs no or minimal (F0-1). An extreme gradient boosting (XGBoost) ML model was developed, and feature importance assessed using SHapley Additive exPlanations.

RESULTS

The multimodal XGBoost model achieved an area under the receiver operating characteristic curve of 0.90 (95% confidence interval: 0.79-0.99) in predicting significant graft fibrosis. TE-derived LSM was the most influential predictor, followed by graft age, aspartate aminotransferase levels, age, and body mass index. XGBoost outperformed other conventional ML algorithms. Predictions were generated for a test set. Subgroup analysis showed elevated body mass index was associated with increased LSM and greater variability in TE readings, suggesting reduced reliability of non-invasive fibrosis assessment in obese recipients.

CONCLUSION

A multimodal XGBoost model reliably and accurately diagnosed significant graft fibrosis in LT recipients, providing personalized predictions. Individualized SHapley Additive exPlanations analysis revealed LSM measurement as the strongest fibrosis indicator, with graft and recipient age being the significant contributors among other patients.

Key Words: Artificial intelligence; Liver transplantation; Graft fibrosis prediction; Transient elastography; Machine learning; Liver stiffness measurements

Core Tip: Accurate non-invasive assessment of graft fibrosis after liver-transplantation remains challenging. In this study, we developed an extreme gradient boosting-based machine learning model integrating transient elastography-derived liver stiffness measurements with clinical and laboratory variables to predict clinically significant graft fibrosis. The model demonstrated strong diagnostic performance and highlights the potential of multimodal data integration to improved non-invasive fibrosis assessment in transplant recipients. External validation in multicenter cohorts will be required before clinical implementation.

INTRODUCTION

Liver transplantation is a lifesaving, definitive treatment for end-stage liver disease[1]. However, the long-term outcomes of liver transplantation beyond a year post-transplant have plateaued in recent years[2]. Graft fibrosis may arise from recurrent or de novo injury to the graft, including recurrent disease, immune-mediated injury, long-term immunosuppression, or the onset of metabolic dysfunction-associated steatohepatitis[3]. Approximately one-third of liver transplant (LT) patients are affected by some degree of graft fibrosis, and about 25% may progress to graft cirrhosis requiring re-transplantation or resulting in compromised survival[3,4].

Early detection of insidious graft fibrosis is crucial, as timely intervention can prevent progression to cirrhosis and graft failure[5]. However, diagnosing fibrosis in a transplant patient can be challenging, especially in patients with obesity, where the accuracy and reliability of transient elastography (TE) findings are of limited accuracy and can overestimate fibrosis[6,7]. Current practice relies on a combination of biochemical monitoring with routine laboratory tests, imaging with TE and histological assessment with liver biopsy as the gold standard for diagnosing and staging graft fibrosis[3]. However, liver biopsy is an invasive procedure with inherent risks such as bleeding, infection, and patient discomfort, limiting its repeated use for routine surveillance in clinical practice. Moreover, biopsy sampling variability and inter-observer discrepancies further challenge its reliability and consistency[8]. These limitations highlight the need for alternative non-invasive, reliable, and repeatable methods for assessment of graft fibrosis over time.

While various laboratory tests, biomarkers, and imaging techniques have been explored, TE (Fibroscan) has emerged as a reliable non-invasive tool for the repeated evaluation of liver fibrosis in patients with chronic liver disease[9,10]. Similarly, several studies have also shown TE is a highly accurate diagnostic tool for staging fibrosis and steatosis in LT recipients, outperforming other non-invasive tests such as serum biomarkers, including aspartate aminotransferase (AST) to platelet ratio index (APRI) and fibrosis-4 (FIB-4)[7,11-14], and providing a valuable alternative to liver biopsies[15-17]. However, confounders such as elevated body mass index (BMI), acute inflammation, variations in the anatomic location of the graft, hepatic congestion or cholestasis can falsely elevate stiffness and affect reliability of results of TE[7,15,16]. Additionally, it is unclear how well TE works in living donor recipients and those with obesity, where readings may be less reliable. Furthermore, there is no universal consensus on definitive cutoff values for fibrosis staging in LT recipients undergoing TE[18].

Physicians often attempt to integrate different modes of data to provide an assessment as to whether a LT recipient has significant graft fibrosis. In this study, we create a multi-factorial scoring system that integrates longitudinal clinical, and laboratory data with serial TE findings, including liver stiffness measurement (LSM) and controlled attenuation parameter (CAP) score, to accurately predict significant fibrosis and generate individualized predictions for individual LT recipients.

MATERIALS AND METHODS

Study design and population

This single-center retrospective cohort analysis included 197 adult LT recipients (aged 18 years or over) who underwent a liver biopsy within 365 days of a TE assessment from 2014-2023. Patients who did not have TE findings, who had undergone multiorgan transplant or re-transplantation were excluded from the study. For patients with multiple liver biopsies post-transplant, only the biopsy sample findings closest to the TE assessment were collected and the TE measurement closest to the biopsy date was used for analysis. The derivation of the study cohort is shown in Figure 1.

Figure 1 Flow diagram. The derivation of the final cohort of 197 liver transplant recipients with biopsy-Fibroscan pairs within 1 year, identified from 7367 biopsies performed between 2014 and 2023, after applying exclusion criteria for biopsies < 6 months post-transplant, missing laboratory data within 1 year, and absence of a Fibroscan within 1 year of biopsy.

Exposure and outcomes

Clinical data were retrospectively collected from the electronic medical records of Ajmera transplant center at University Health Network, Toronto, Canada. Variables collected for analysis included demographics and clinical characteristics: Recipient age, graft age, sex, BMI, type of liver transplantation (deceased vs living), pre-transplant comorbidities (hypertension, diabetes, cardiovascular disease, alcohol-associated liver disease, primary indication for LT and post-transplant comorbidities.

LSM was recorded in kPa and the CAP in decibels per meter (dB/m), both obtained using TE. LSM reflects hepatic fibrosis, while CAP (range 100-400 dB/m) quantifies hepatic steatosis. Both values were obtained from Fibroscan 502 Touch model, with the probe automatically selected based on the patient’s body weight. A minimum of 10 valid measurements, with a success rate greater than 60% and interquartile range (IQR) below 30% were considered reliable and included in the study.

Laboratory values taken closest to liver biopsy/TE time point were also recorded, including alanine aminotransferase, AST, alkaline phosphatase (ALP), gamma glutamyl transferase, bilirubin, albumin, international normalized ratio, creatinine, and components of complete blood count (hemoglobin, white blood cell count and platelet count). Only variables with less than 10% missingness were included in the study, and missing values were imputed using median values.

Primary outcomes

Primary outcome was graft fibrosis stage, as determined by liver biopsy. Fibrosis was dichotomized into “significant fibrosis” (stage 2 and above) vs “no or non-significant fibrosis” (stage 0-1). The secondary outcome was the development of a deep learning model that could accurately predict graft fibrosis stage with imaging, patient demographics, and laboratory results.

Ethical considerations

The study was approved by the Ethics Committee of University Health Network (Approval No. 21-6170.7) which waived the requirement for informed consent from patients due to the exclusive use of electronic medical record data.

Statistical analysis

Baseline characteristics were determined using median (IQR) for continuous variables, and proportions for categorical variables. Univariate analysis was used to assess the association between each variable and significant fibrosis: χ² tests for categorical variables and Mann-Whitney U tests for continuous variables. A P-value < 0.05 was considered statistically significant. Variables with significant associations were selected as inputs for machine learning (ML) analysis.

To explore the effect of BMI on liver stiffness (measured in kPa), patients were stratified into twelve groups based on fibrosis status (significant vs non-significant or no fibrosis) and BMI categories: < 18.5, 18.5-24.9, 25.0-29.9, 30.0-34.9, 35.0-39.9, and ≥ 40. For each group, the median and SD of kPa readings were calculated.

ML analysis: Four ML models were developed to predict the risk of significant fibrosis (stage ≥ 2) for each TE-biopsy pair: Logistic regression (LR), random forest classifier, support vector machine (SVM), and extreme gradient boosting (XGBoost). Patients were classified into “significant fibrosis (fibrosis stage 2 and above)” or “no or non-significant fibrosis” groups based on a risk threshold of 0.5. A grid search approach was employed to identify the optimal combination of hyperparameters for each model. Model performance was evaluated in terms of sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC). To address class imbalance synthetic minority over-sampling technique was applied (https://arxiv.org/pdf/1106.1813). 95% confidence intervals (CIs) were computed using 200 bootstrap iterations.To interpret model predictions, SHapley Additive exPlanations (SHAPs) were applied to the best-performing model. SHAP values were used to quantify the contribution of individual features to model decisions both at the cohort and individual levels.

Rule-of-thumb - thresholding for fibrosis prediction: To facilitate clinical application, scatter plots were generated for each variable identified as important by SHAP analysis in relation to LSM (kPa). Based on the observed patterns, simple threshold rules were proposed to support clinical decision-making and help identify patients at low likelihood of significant fibrosis.

RESULTS

Cohort characteristics

A total of 197 adult LT recipients were included in the analysis [125 males (63.5%), 72 (36.5%) female]. All patients underwent at least one liver biopsy and TE within 365 days. Of these 168 patients (85.2%) did not have any significant fibrosis on biopsy (stage 0-1), while 29 (14.7%) had significant fibrosis on biopsy (stage ≥ 2).

Association of clinical variables with fibrosis

In the evaluation of TE values, LSM (kPa) were significantly higher in patients with significant fibrosis compared to those with no significant fibrosis [median: 14.8 kPa (IQR: 8.0-20.9) vs 7.7 kPa (IQR: 4.6-8.0), P < 0.001]. CAP (in dB/m) which qualifies hepatic steatosis showed no statistically significant difference between the fibrosis groups. Regarding the clinical data, graft age was significantly greater in patients with fibrosis [median: 7.1 years (IQR: 2.6-8.9 years)] compared to those with no or non-significant fibrosis [median: 2.1 years (IQR: 0.3-2.3 years), P < 0.001]. In contrast, recipient age was significantly lower in patients with significant fibrosis [median: 47.3 years (IQR: 38.8-56.1 years)] compared to those with no or non-significant fibrosis [median: 55.4 years (IQR: 49.7-63.0 years), P < 0.001]. There were no significant differences between the groups in terms of sex, presence of pre-transplant comorbidities (e.g., cardiovascular disease, diabetes), transplant indications (e.g., acute liver failure, alcohol-related cirrhosis, or viral hepatitis), or type of liver transplantation (deceased vs living donor). Similarly, laboratory variables such as serum albumin, bilirubin, platelets, creatinine, white blood cell count, sex, international normalized ratio, did not show statistically significant differences between the groups (Table 1).

Table 1 Patient demographic and clinical characteristics, median (interquartile rage)/n (%).

Variable	Fibrosis < 2, (n = 168)	Fibrosis ≥ 2, (n = 29)	P value	Missingness
Age at transplant	58.8 (49.7-63.0)	49.0 (38.8-56.1)	< 0.001	0 (0.0)
Graft age	1.0 (0.3-2.3)	3.4 (2.6-8.9)	< 0.001	5 (2.5)
Male	107 (63.7)	18 (62.1)	0.867	0 (0.0)
Body mass index	26.9 (25.8-31.1)	26.5 (24.5-27.6)	0.032	0 (0.0)
Type of transplant: Living donor	62 (36.9)	6 (20.7)	0.091	0 (0.0)
Fibroscan
CAP	258.2 (211.7-318.2)	240.0 (194.0-290.0)	0.087	7 (3.5)
LSM (kPa)	6.0 (4.6-8.0)	12.1 (8.0-20.9)	< 0.001	2 (1.0)
Albumin	41.0 (38.0-44.0)	41.0 (36.0-42.0)	0.187	0 (0.0)
Bilirubin	12.0 (9.0-18.2)	15.0 (12.0-22.0)	0.065	0 (0.0)
INR	1.0 (1.0-1.1)	1.1 (1.0-1.2)	0.405	0 (0.0)
ALP	112.0 (85.7-163.0)	134.0 (108.0-229.0)	0.001	0 (0.0)
ALT	36.5 (23.0-64.2)	52.0 (28.0-67.0)	0.713	0 (0.0)
AST	27.0 (19.7-41.2)	41.0 (28.0-58.0)	0.012	0 (0.0)
Creatinine	102.5 (86.0-126.0)	90.0 (84.0-110.0)	0.224	0 (0.0)
Hemoglobin	127.0 (113.5-136.0)	134.0 (122.0-145.0)	0.017	0 (0.0)
Platelet	164.0 (131.7-199.5)	156.0 (112.0-225.0)	0.618	0 (0.0)
White blood cells	5.9 (4.4-7.3)	6.2 (4.3-7.7)	0.916	0 (0.0)
MELD	9.0 (7.0-11.0)	10.0 (7.0-12.0)	0.082	0 (0.0)
Indication for transplant
AIH	2 (1.2)	2 (6.9)	0.044	0 (0.0)
ALF	0 (0.0)	0 (0.0)		0 (0.0)
ALD	19 (11.3)	5 (17.2)	0.369	0 (0.0)
HBV	11 (6.5)	1 (3.4)	0.521	0 (0.0)
HCV	44 (26.2)	9 (31.0)	0.589	0 (0.0)
Malignancy	0 (0.0)	0 (0.0)		0 (0.0)
Other	37 (22.0)	8 (27.6)	0.512	0 (0.0)
Unknown	6 (3.6)	1 (3.4)	0.973	0 (0.0)
Pre-transplant cardiovascular disease (yes/no)	34 (20.2)	7 (24.1)	0.635	0 (0.0)
Pre-transplant hypertension	37 (22.0)	2 (6.9)	0.059	0 (0.0)
Pre-transplant diabetes	35 (20.8)	2 (6.9)	0.076	0 (0.0)
Post-transplant hypertension	113 (67.2)	19 (65.5)	0.854	0 (0.0)
Post-transplant diabetes	117 (69.6)	18 (62.1)	0.419	0 (0.0)

P-values were calculated using χ² tests for categorical variables and Mann-Whitney U tests for continuous variables. CAP: Controlled attenuation parameter; LSM: Liver stiffness measurement; INR: International normalized ratio; ALP: Alkaline phosphatase; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; MELD: Model for end-stage liver disease; AIH: Autoimmune hepatitis; ALF: Acute liver failure; ALD: Alcohol-related liver disease; HBV: Hepatitis B virus; HCV: Hepatitis C virus.

Liver stiffness and BMI relationship

To evaluate the effect of BMI on LSMs (in kPa), we stratified patients by fibrosis stage and BMI category. In patients with significant fibrosis (stage ≥ 2), liver stiffness increased with higher BMI, particularly beyond a BMI of 30. For example, in patients with BMI 30-35 and significant fibrosis, the mean kPa reached 26.3. Although patients with no significant fibrosis (stage < 2) also exhibited increased kPa at higher BMI, the rise was less pronounced. Liver stiffness accelerates in patients with a BMI over 30, particularly in those with established significant fibrosis (stage ≥ 2). Notably, there is also significantly increased variability of liver stiffness (kPa) at higher BMI. For patients with no significant fibrosis (stage < 2), SDs increased dramatically in the ≥ 35 BMI ranges: ± 11.7 for BMI 35-40 and ± 9.7 for > 40. For patients with significant fibrosis (stage ≥ 2), while there were limited samples (mostly n = 1), high values and variability were also observed. Please refer to Supplementary Table 1 and Supplementary Figure 1 for further details.

ML model performance

We evaluated four ML models, including LR, SVM, random forest, and XGBoost to predict graft fibrosis (stage ≥ 2). Hyperparameters and model configurations are presented in Supplementary Table 2. Performance metrics included AUROC, sensitivity, and specificity.

XGBoost achieved the highest performance with an AUROC of 0.927 (95%CI: 0.799-0.996), followed by random forest (AUROC = 0.922; 95%CI: 0.777-0.984), SVM (AUROC = 0.850; 95%CI: 0.756-0.973), and LR (AUROC = 0.828; 95%CI: 0.748-0.953; Table 2). XGBoost demonstrated high specificity (0.941; 95%CI: 0.860-1.000) and moderate sensitivity (0.667; 95%CI: 0.369-1.000), indicating strong potential for ruling out significant fibrosis (Table 2). To evaluate the incremental value of the multimodal approach, the predictive performance of LSM alone was assessed. Using LSM as the sole predictor yielded an AUROC of 0.865 (95%CI: 0.669-0.902), which was lower than that observed for the multimodal XGBoost model (AUROC of 0.927; Table 3).

Table 2 Multimodal machine learning models used for graft fibrosis classification using all variables.

Model	AUROC (95%CI)	Sensitivity (95%CI)	Specificity (95%CI)
Logistic regression	0.828 (0.748-0.953)	0.667 (0.190-1.000)	0.882 (0.738-0.979)
Support vector machine	0.850 (0.756-0.973)	0.500 (0.222-1.000)	0.912 (0.810-1.000)
Random forest	0.922 (0.777-0.984)	0.667 (0.454-1.000)	0.941 (0.752-0.962)
XGBoost	0.927 (0.799-0.996)	0.667 (0.369-1.000)	0.941 (0.860-1.000)

XGBoost: Extreme gradient boosting; AUROC: Area under the receiver operating characteristic curve; CI: Confidence interval.

Table 3 Machine learning models used for graft fibrosis classification using liver stiffness measurement value alone.

Model	AUROC (95%CI)	Sensitivity (95%CI)	Specificity (95%CI)
Logistic regression	0.796 (0.719-0.936)	0.500 (0.029-0.833)	0.735 (0.759-1.000)
Support vector machine	0.787 (0.226-0.837)	0.833 (0.000-0.429)	0.618 (0.898-1.000)
Random forest	0.843 (0.649-0.898)	0.833 (0.200-0.875)	0.676 (0.649-0.898)
XGBoost	0.865 (0.669-0.902)	0.833 (0.133-0.817)	0.647 (0.740-0.958)

XGBoost: Extreme gradient boosting; AUROC: Area under the receiver operating characteristic curve; CI: Confidence interval.

Model explainability using SHAP

To interpret the model’s predictions, SHAP analysis was performed on the XGBoost classifier. The direction and magnitude of association between key quantitative input variables and the likelihood of significant fibrosis are illustrated in Figure 2 and Supplementary Figure 2. The most influential predictors of significant fibrosis included higher liver stiffness (kPa), older graft age, elevated hemoglobin levels, higher AST, and a history of hypertension. In contrast, protective factors associated with a lower likelihood of significant fibrosis were older recipient age, elevated ALP, and receipt of a liver from a living donor. BMI despite its association with higher LSMs, acts as a protective factor in SHAP analysis, indicating a negative contribution to the likelihood of significant graft fibrosis.

Figure 2 Mean SHapley Additive exPlanation values. Feature importance plot for the 10 variables most associated with significant graft fibrosis for the best performing model (extreme gradient boosting). The higher the average SHapley Additive exPlanation value, the higher the contribution of the individual feature in predicting graft fibrosis. The direction of the association between top-ranked input variables and graft fibrosis is shown with arrows. Features highlighted in red signify an increased risk of graft fibrosis, while those depicted in blue indicate protective factors against fibrosis. BMI: Body mass index; HGB: Hemoglobin; ALP: Alkaline phosphatase; AST: Aspartate aminotransferase; HTN: Hypertension; Y: Yes; N: No; 1Diag-AIH: Type 1 autoimmune hepatitis diagnosis; SHAP: SHapley Additive exPlanations.

We also generated additive force plots for the XGBoost model to illustrate the effect of each feature to development of significant graft fibrosis in individual patients (Figure 3; Supplementary Figure 3). These plots illustrate the direction and magnitude of each feature’s impact on a given patient’s predicted outcome. To contextualize these local explanations, we compared each patient’s feature values with the group-level statistics (median and IQR) presented in Table 4. For example, the patient shown in Figure 3A had a LSM of 3.9 kPa. This value is below the IQR of the “no significant fibrosis” group, corresponding to a large negative SHAP value of -2.78 for LSM, indicating a strong contribution against the prediction of fibrosis. Similarly, this patient had a graft age of 0.7 years, lower than the median of the “no fibrosis” group, which contributed a negative SHAP value of -1.55. The patient’s chronological age (62.5 years) was also higher than the median age in both groups, and correspondingly received a SHAP value of -1.17, suggesting older recipients were less likely to develop significant fibrosis in our cohort.

Figure 3 Contributing factors in extreme gradient boosting decision-making for 4 random patients by SHapley Additive exPlanation analysis. Features highlighted in red signify an increased risk of graft fibrosis, while those depicted in blue indicate protective factors against fibrosis. A: No fibrosis (model prediction: No fibrosis): A 63-year-old patient, eight months post-deceased donor liver transplant, with no pre-transplant or post-transplant hypertension, diabetes, or cardiovascular disease. The indication for transplant was unknown. Body mass index (BMI) at time of Fibroscan was 26.9. Liver stiffness measured 3.9 kPa. Biopsy showed no fibrosis; B: No fibrosis (model prediction: No fibrosis): A 59-year-old patient, four years post-deceased donor liver transplant for hepatitis C. No pre-transplant comorbidities but developed both hypertension and diabetes post-transplant. BMI at time of Fibroscan was 21.3. Liver stiffness was 4.9 kPa. Biopsy showed no fibrosis; C: Significant fibrosis (model prediction: Fibrosis): A 41-year-old patient, 2.8 years post-deceased donor liver transplant for autoimmune hepatitis. No pre-transplant comorbidities. Developed both hypertension and diabetes post-transplant. BMI at time of Fibroscan was 21.4. Liver stiffness measured 11.9 kPa. Biopsy confirmed significant fibrosis; D: Significant fibrosis (model prediction: Fibrosis): A 42-year-old patient, 15 years post-deceased donor liver transplant for hepatitis C. No pre-transplant hypertension, diabetes, or cardiovascular disease. Developed both hypertension and diabetes post-transplant. Body mass index at time of Fibroscan was 26.9. Liver stiffness was 7.7 kPa. Biopsy showed significant fibrosis. SHapley Additive exPlanation analysis showed that graft age was the most important predictor of significant fibrosis in this patient, as opposed to the patient examples where liver stiffness measurement was most important. LSM: Liver stiffness measurement; BMI: Body mass index; AST: Aspartate aminotransferase; HGB: Hemoglobin; ALP: Alkaline phosphatase; HTN: Hypertension; Y: Yes; N: No; AIH: Autoimmune hepatitis.

Table 4 Representative SHapley Additive exPlanation case examples, median (interquartile rage)/n (%).

Variable	Group characteristics		Value (SHAP value)
	Fibrosis < 2 (n = 168)	Fibrosis ≥ 2 (n = 29)	Case 1	Case 2	Case 3	Case 4
LSM (kPa)	6.0 (4.6-8.0)	12.1 (8.0-20.9)	3.9 (-2.78)	4.9 (-3.29)	7.7 (-0.47)	5 (-3.34)
Graft age (year)	1.0 (0.3-2.3)	3.4 (2.6-8.9)	0.7 (-1.55)	4.0 (+0.62)	14.8 (+2.39)	4.7 (+0.77)
Age (year)	58.8 (49.7-63.0)	49.0 (38.8-56.1)	62.5 (-1.17)	59.0 (-1.43)	42.0 (+1.70)	59.0 (-1.4)
BMI	26.9 (25.8-31.1)	26.5 (24.5-27.6)	26.9 (+0.27)	21.3 (+0.80)	26.9 (+0.64)	24.5 (+0.92)
HGB	127.0 (113.5-136.0)	134.0 (122.0-145.0)	131 (+0.13)	144 (+0.34)	158 (+0.57)	124 (-0.11)
ALP	112.0 (85.7-163.0)	134.0 (108.0-229.0)	77 (-0.11)	125 (+0.35)	52 (-0.11)	175 (-0.21)
Living donor	62 (36.9)	6 (20.7)	No (+0.07)	No (+0.21)	No (+0.22)	No (+0.21)
AST	27.0 (19.7-41.2)	41.0 (28.0-58.0)	10 (-0.21)	29 (-0.12)	22 (-0.10)	38 (+0.14)
HTN pre-transplant (yes/no)	37 (22.0)	2 (6.9)	No (-0.04)	No (-0.01)	No (-0.01)	No (-0.01)

LSM: Liver stiffness measurement; BMI: Body mass index; HGB: Hemoglobin; ALP: Alkaline phosphatase; AST: Aspartate aminotransferase; HTN: Hypertension; SHAP: SHapley Additive exPlanation.

In contrast, the patient’s hemoglobin level was 131 g/L - closer to the median of the “significant fibrosis” group - resulting in a small positive SHAP value of +0.13. Similarly, the patient’s BMI was 26.9, aligning more closely with the concentrated IQR of the fibrosis group (median 26.5), resulting in a positive SHAP contribution of +0.27. These findings illustrate that both the direction and magnitude of each SHAP feature are influenced by how closely the patient’s value aligns with the distribution of that feature in the fibrosis vs non-fibrosis groups.

This interpretability framework not only enhances transparency of the model’s decision-making but also provides clinical insight into how patient-specific characteristics influence risk classification.

Practical guideline

To further explore the relationships between key predictors and liver stiffness, we generated two-dimensional scatter plots of kPa vs variables identified by SHAP (Figure 4). Each point represents one patient, colored by fibrosis stage (0-4). Cutoff values for TE in LT recipients range widely, from 4.7 kPa to 12.3 kPa. While a cut of 7.4 kPa and above is commonly considered indicative of significant fibrosis (stage 2 and above) in LT recipients[18], our plots demonstrate that this cutoff is insufficient when considered in isolation. For variables such as graft age, ALP, and AST, significant fibrosis can still occur above or below this threshold, suggesting a more nuanced, multivariate approach is needed. However, the 7.4 kPa threshold appeared more reliable when stratifying by patient age, BMI, hemoglobin, and donor type.

Figure 4 Rule-of-thumb scatter plots. Scatter plots showing liver stiffness measurements (kPa) in relation to key predictive variables: A: Graft age; B: Aspartate aminotransferase; C: Patient age; D: Body mass index; E: Alkaline phosphatase; F: Hemoglobin; G: Donor type. Each point represents an individual patient. AST: Aspartate aminotransferase; BMI: Body mass index; ALP: Alkaline phosphatase; HGB: Hemoglobin.

To identify low-risk patients in a more intuitive manner, we constructed a three-dimensional scatter plot combining kPa, graft age, and ALP. Patients clustering near the origin (i.e., low kPa, low graft age, and low ALP) were predominantly those without significant fibrosis (stage 0-1). Based on this, we fitted a simple ellipsoid to define a low-risk zone, with radii defined as follows: (1) Graft age radius ≤ 8 years; (2) LSM (kPa) ≤ 9; and (3) ALP ≤ 600.

We derived the following formula for clinical use: Patients falling within this ellipsoid were considered unlikely to have significant fibrosis. This rule-of-thumb classifier demonstrated a sensitivity of 1.00, specificity of 0.68, and overall accuracy of 0.73.

DISCUSSION

In this study, we developed and internally evaluated an XGBoost-based ML model integrating clinical variables with a quantitative imaging-derived parameter from TE (LSM) to monitor and predict clinically significant graft fibrosis (stage 2 and above) in LT recipients. The model’s diagnostic performance was compared with three other ML models (LR, random forest classifier and SVM), each incorporating clinical, demographic, and TE-derived LSMs, using liver biopsy as the reference standard. The XGBoost model outperformed the other models, achieving the greatest AUROC value of 0.90 (95%CI: 0.79-0.99), with SHAP values indicating TE findings, graft age, patient age, BMI as the most impactful variables. We additionally derived a simple clinical rule integrating graft age (≤ 8 years), liver stiffness (kPa ≤ 9), and ALP levels (≤ 600) to identify patients at low likelihood of significant fibrosis with high sensitivity (100%), but moderate specificity (60%) and overall accuracy (65%). This proposed ellipsoid low-risk rule should be interpreted as exploratory and hypothesis-generating. As the rule was derived from the same cohort used for model development and requires external validation to explore reproducibility and clinical utility prior to clinical application.

The use of the multimodal artificial intelligence tools in clinical practice remains in its early stages and could potentially reshape the post-transplant fibrosis screening care pathway. Liver biopsy remains the gold standard for fibrosis staging in the posttransplant population; however, it is invasive, resource intensive and is associated with risks including sampling error, infection, bleeding and pain. Our findings suggest that a multimodal ML model that incorporates TE-derived LSMs with clinical data and patient demographics could approximate biopsy-based fibrosis staging, particularly when used in conjunction with clinical judgement. Importantly, the diagnostic thresholds for TE can vary across devices and patient subtypes. Factors such as obesity, inflammation, and hepatic congestion may impact measurement accuracy, underscoring the need for context-specific interpretation of TE cutoffs in LT recipients. This variability reinforces the potential benefit of a multimodal ML model that integrates multiple clinical inputs to enhance diagnostic precision. Such a model could reduce dependence on liver biopsies while enabling more frequent non-invasive fibrosis monitoring, potentially allowing earlier detection of significant graft fibrosis in LT patients. Significant fibrosis, if unrecognized, can progress to cirrhosis, graft failure and even requirement of re-transplantation[3]. A non-invasive tool that can identify patients at risk of graft fibrosis may allow for early therapeutic adjustments and closer follow up which may delay or prevent progression to significant graft injury or graft failure. The model utilizes routinely collected, non-invasive data, making it feasible for integration into clinical workflows. Furthermore, this ML tool could be especially useful in patients that are not candidates for frequent biopsies and useful in resource- limited settings.

The findings of this study build upon recent work in hepatology focusing on fibrosis prediction tools. Studies have shown XGBoost models are successful at diagnosing advanced liver fibrosis (stage 3-4) in patients with non-alcoholic steatohepatitis[19,20]. These models have been shown to outperform traditional scoring systems like FIB-4 and APRI in predicting liver fibrosis[19,20]. Xiong et al[19] showed the XGBoost model achieved an AUROC of 0.934 in the training cohort and 0.917 in the validation cohort, significantly outperforming APRI (AUROC = 0.803 and AUROC = 0.737) and FIB-4 (AUROC = 0.811 and AUROC = 0.752) in their respective cohorts. Similarly, Wu et al[20] reported that an XGBoost model achieved the best AUROC score of 0.836 for advanced fibrosis (stage 3-4) compared to other models, including LR, random forest classifier, SVM, and decision tree. However, these studies focused on the role of ML prediction tools for liver fibrosis in non-transplant patients. Within LT research, Azhie et al[21] conducted a comprehensive retrospective study that explored the use of a weighted long-short-term memory model, along with other DL and ML models trained on longitudinal, clinical and laboratory data to predict significant fibrosis (≥ stage 2) in LT recipients. The weighted LSTM model consistently outperformed the other models and achieved an AUROC of 0.798 (95%CI: 0.790-0.810)[21]. In a subset analysis involving 149 patients who underwent TE within a year of liver biopsy, the weighted LSTM model achieved an AUROC of 0.705 (95%CI: 0.687-0.724), comparable to the AUROC of 0.685 (95%CI: 0.662-0.704) for TE alone[21]. In contrast, our study integrated TE-derived LSMs with clinical and laboratory variables within an explainable ML framework, allowing identification of individual predictor contributions through SHAP analysis. This combined approach yielded a higher AUROC of 0.927 (95%CI: 0.79-0.96; Table 2) for predicting significant graft fibrosis, although the relatively wide CI and limited fibrosis cohort suggest that these findings should be interpreted cautiously. To better evaluate the incremental value of the multimodal approach, we also assess model performance using LSM alone which yielded a AUROC of 0.865 (95%CI: 0.669-0.902; Table 3). While LSM was a strong individual predictor of fibrosis, the multimodal model incorporating additional clinical variables demonstrated improved overall discrimination. These findings suggest that integrating clinical context with TE-derived measurements may provide a more robust assessment than relying on LSM alone, particularly in transplant recipients where fibrosis risk is influenced by multiple factors. The superior performance of our model demonstrates the additive value of incorporating TE data with clinical parameters to enhance the accuracy of non-invasive fibrosis assessment and provides an interpretable multimodal model for fibrosis prediction. Furthermore, Qazi Arisar et al[22] conducted a pilot study that aimed to develop a radiomics-based model for predicting significant graft fibrosis (≥ F2) in LT recipients. Using a least absolute shrinkage and selection operator regression model, they evaluated three predictive models: Radiomics data from contrast-enhanced computed tomography scans, clinical-only, and a combined radiomics-clinical model. The combined model achieved an AUROC of 0.811 (95%CI: 0.670-0.921), outperforming both the clinical-only model (AUROC = 0.793) and the radiomics-only model (AUROC = 0.664)[22]. In comparison, our study yielded a higher AUROC using TE findings in combination with clinical data. These findings highlight the potential of multimodal data integration in ML models to improve the non-invasive prediction of graft fibrosis.

Previous studies have demonstrated that TE measurements may be influenced by factors such as obesity, inflammation, and hepatic congestion[7,15,16]. Consistent with these observations, our subgroup analysis revealed a significant association between BMI and elevated liver stiffness readings, highlighting potential confounding effects in obese LT recipients. In patients with significant graft fibrosis, kPa values increase sharply in patients with BMI over 30, suggesting that obesity may affect transient TE readings. The variability in LSM significantly widens in high-BMI groups, indicating potential diagnostic uncertainty. This suggest that while higher BMI is associated with increased LSM, this may reflect measurement related variability rather than true fibrosis severity in this population. These findings highlight the importance of vigilant fibrosis assessment in obese LT recipients as they are at higher risk for fibrosis as compared to non-obese recipients[23], and non-invasive diagnostic imaging results may be harder to interpret in this population warranting more reliable diagnostic strategies such as additional imaging or consideration of biopsy.

We acknowledge the limitations of this study including a relatively small cohort size from a single center. In addition, the temporal interval of up to 365 days between TE and biopsy may introduce potential misclassification bias, as fibrosis stage may change over time. Also, using liver biopsy as the reference method for the ML model has some known limitations due to intraobserver and interobserver variability[24]. Additionally, the study’s retrospective nature and single-center setting may limit the generalizability of the findings. Variations in patient populations, clinical practices, and data collection methods across different centers could affect the model’s performance when applied elsewhere. In addition, the time interval between TE and liver biopsy was allowed to extend up to 365 days, which may introduce potential misclassification bias if fibrosis progression or regression occurred during that period. External validation was not feasible within this study due to the limited availability of independent datasets containing LT patients with paired TE and biopsy data; therefore, future multicenter studies will be necessary to assess model generalizability. Similarly, including only patients who underwent both TE and liver biopsy may introduce selection bias, as these patients might differ systematically from those who did not undergo both procedures. This could impact the model’s applicability to the broader transplant population. Furthermore, our feature selection strategy relied on initial univariate screening step to reduce the number of candidate predictors given the relatively small sample size. While this pragmatic approach helped limit model complexity, it may exclude potentially informative nonlinear predictors. Finally, given the relatively small sample size (197 patients) and the complexity of the XGBoost model, there is a risk of overfitting. Cross-validation and external testing are needed to decrease this risk.

CONCLUSION

Our study findings support the development of ML-powered, non-invasive tools that predict significant graft fibrosis using TE and clinical data. Additionally, future work should focus on developing a real-time risk stratification tool integrated within electronic health records. If externally validated and integrated into routine practice, this approach could enable earlier detection, less invasive monitoring, and more personalized surveillance strategies, ultimately improving long-term graft outcomes.