Validation of the TRAIN-AI Score for the Prediction of Hepatocellular Carcinoma Recurrence After Liver Transplantation
Validation of the TRAIN-AI Score for the Prediction of Hepatocellular Carcinoma Recurrence After Liver Transplantation
Liver transplantation (LT) is the best treatment option for patients with early stages of hepatocellular carcinoma (HCC).1 However, the use of LT depends on maintaining a balance between the risk of post-transplant recurrence or HCC-related death and the equitable distribution of organ donors.2-5 Current selection criteria aim to avoid transplant futility by excluding patients from LT who are at a high risk of tumor recurrence. Selecting patients within the Milan criteria has been shown to provide excellent patient outcomes.6,7 However, these criteria have been challenged by other series showing equivalent outcomes for patients transplanted with a greater tumor burden. A combination of morphologic (i.e., tumor number and size) and biological features has been recently proposed with the intent to implement the patient selection process.8,9 Machine learning represents a statistical tool that can leverage the prognostic abilities of a many clinically available variables. Recently, the TRAIN-AI has been proposed, and a post-transplant HCC recurrence risk calculator using machine learning based on the TRAIN-AI score is available.10 We are seeking to explore the generalizability of this machine learning model to other institutions through a validation study.
Liver transplantation (LT) is the best treatment option for patients with early stages of hepatocellular carcinoma (HCC).1 However, the use of LT depends on maintaining a balance between the risk of post-transplant recurrence or HCC-related death and the equitable distribution of organ donors.2-5 Current selection criteria aim to avoid transplant futility by excluding patients from LT who are at a high risk of tumor recurrence. Selecting patients within the Milan criteria has been shown to provide excellent patient outcomes.6,7 However, these criteria have been challenged by other series showing equivalent outcomes for patients transplanted with a greater tumor burden. A combination of morphologic (i.e., tumor number and size) and biological features has been recently proposed with the intent to implement the patient selection process.8,9 Machine learning represents a statistical tool that can leverage the prognostic abilities of a many clinically available variables. Recently, the TRAIN-AI has been proposed, and a post-transplant HCC recurrence risk calculator using machine learning based on the TRAIN-AI score is available.10 We are seeking to explore the generalizability of this machine learning model to other institutions through a validation study.
Study aims and objective:
The primary objective of this study will be to validate our previously reported TRAIN-AI score using external datasets from other HCC centers.
Study design and methodology:
Validate the TRAIN-AI model by comparing it to other available recurrence risk algorithms on a held-out test set. TRAIN-AI will be compared with Milan Criteria, San Francisco Criteria, Up-to-Seven Criteria, TBS, Metroticket 2.0 Score, HALT-HCC Score, AFP-French model, 5-5-500 Role, NYCA Score, and TRAIN Score.
Study population Adults (≥ 18 years of age) who underwent liver transplant for HCC during the period January 2003 - December 2018.
Inclusion criteria • Patients who underwent liver transplant alone for a diagnosis of HCC. Exclusion criteria
• Patients with incidentally discovered HCC on the explanted liver (i.e. the HCC was not known before the LT)
• Retransplantations or multivisceral transplants
• Patients with tumors other than pure HCC (such as cholangiocarcinoma, mixed HCC-cholangiocarcinoma tumors, fibrolamellar HCC etc.)
Data collection/variables The data required for the analysis are present in the excel spread sheet already sent to the involved centers. The columns in yellow are obligatory for calculating the score.
Data/Statistical analysis:
Data from HCC transplants performed from January 1, 2003 to December 31 2018 will be requested from the invited centers who will obtain them from their records including electronic chart review. The last allowed follow-up of patients included will be December 31, 2023, as this is a retrospective study design. Patient survival will be calculated from the date of LT to patient death (due to any cause). If death does not occur, then the patient will be censored at their last known alive date. The time to recurrence will be calculated from transplantation to the first imaging study (or biopsy if appropriate) that confirmed tumor recurrence. Patient demographics and clinicopathologic characteristics will be described using descriptive statistics using means, medians and proportions, where appropriate. The exact methodology for the calculation of the machine learning-algorithm prediction model, as well the comparisons to previously published models, has been previously outlined in our development cohort study.10 All statistical analyses will be performed using using Python 3.10.1 (libraries: pycox, torch, scikit-learn, and lifelines).
Inclusion Criteria:
Exclusion Criteria: