A Novel Deep Learning Model as a Donor–Recipient Matching Tool to Predict Survival after Liver Transplantation

Nikolaus Börner, Markus Bo Schoenberg, Philipp Pöschke, Christian Heiliger, Sven Jacob, Dominik Koch, Benedikt Pöllmann, Moritz Drefs, Dionysios Koliogiannis, Christian Böhm, Konrad W. Karcz, Jens Werner, Markus GubaJournal of Clinical Medicine 2022 · DOI

TL;DR

This paper introduces a novel deep learning model designed to predict patient survival after liver transplantation. By leveraging a comprehensive dataset and a unique imputation method, the model achieves high accuracy in predicting outcomes, offering a more transparent and potentially fairer alternative to current organ allocation systems.

What Problem Does This Paper Address?

Liver transplantation is a life-saving procedure, but the scarcity of donor organs creates a significant challenge in allocation. Current systems, like the Model for End-Stage Liver Disease (MELD), primarily rely on urgency, which can lead to the transplantation of organs in futile cases and is susceptible to manipulation. The authors argue that a more accurate prediction of post-transplant survival is crucial for developing a truly fair and effective allocation system that maximizes patient benefit. The core problem is the difficulty in accurately predicting survival outcomes based on donor and recipient characteristics, which is essential for optimizing organ allocation.

Methods

The researchers developed a novel, individually designed deep learning neural network. A key innovation is their multidimensional medical combined imputation (MMCI) algorithm, created to handle the heterogeneous and often incomplete nature of transplant data. This pipeline of interconnected imputation methods aims to estimate missing values with high accuracy, outperforming established methods like missForest, k-NN, and MICE. The neural network itself is structured with multiple layers to learn complex correlations between input variables (donor, recipient, and transplant-specific data) and the outcome (survival). Hyperparameters were tuned to balance model complexity and prevent underfitting or overfitting, using metrics like accuracy, cross-entropy loss, F1 score, and AUC score for monitoring.

Datasets

The study utilized data collected from all liver transplant patients at a university transplant center in Munich/Erlangen between 2004 and 2019. This resulted in a dataset of 529 transplantations, encompassing 1058 donor-recipient observations. The data included recipient demographics, underlying disease, MELD score, and laboratory values; donor demographics, cause of death, and laboratory values; and transplant-specific data such as ischemia time and graft quality. The dataset was split into an 80% training set, a 10% cross-validation set, and a 10% test set (52 transplantations) for evaluation.

Key Insights and Ideas

  • The Power of MMCI: The development of a robust imputation method (MMCI) is a significant contribution, addressing a common bottleneck in medical data analysis and enabling the use of more comprehensive datasets.
  • Transparency in AI: The authors emphasize the importance of a transparent and interpretable deep learning framework, particularly in clinical decision-making, to address concerns about "black-box" AI models.
  • Predicting Multiple Timepoints: Unlike some previous models that predict survival at a single timepoint, this deep learning model can predict outcomes at various stages (e.g., in-hospital, 3-month, 12-month), offering a more nuanced view of patient prognosis.
  • AI for Fairer Allocation: The ultimate goal is to move beyond urgency-based allocation towards a utility-based system that maximizes patient survival, with AI playing a central role in achieving this.

Key Results

The deep learning model demonstrated strong performance:

  • Overall Accuracy: 95.8% for combined outcome parameters.
  • In-Hospital Mortality Prediction: 94.3% accuracy.
  • Overall F1 Score: 0.899.
  • Overall AUC Score: 0.940. This AUC score is highlighted as particularly valuable in the clinical context of transplantation.

The study also found that while demographic and disease-specific variables showed no significant differences between training and test sets, some laboratory values (potassium) and the donor risk index (DRI) did show significant differences.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Study: The study is based on retrospective data from two German transplant centers.
  • Generalizability: While cross-validation was used to mitigate bias, external validation in a randomized clinical trial, ideally multicenter, is necessary for broader generalizability.
  • "Black-Box" Nature: Despite efforts towards transparency, the inherent nature of deep learning models as "black boxes" remains a challenge.

Future work should focus on external validation and ultimately integrating this model into a clinical decision support system for organ allocation. The authors propose developing an AI-assisted utility-based allocation concept that leverages this predictive power to optimize organ utility and improve patient outcomes.

Summary updated

Abstract

With the achieved results, the network serves as a reliable tool to predict survival. It adds new insight into the potential of deep learning to assist medical decisions. Especially in the field of transplantation, an AUC Score of 94% is very valuable. This neuronal network is unique as it utilizes transparent and easily interpretable data to predict the outcome after liver transplantation. Further validation must be performed prior to utilization in a clinical context.
BibTeX
@article{borner2022novel,
  author = {Nikolaus Börner and Markus Bo Schoenberg and Philipp Pöschke and Christian Heiliger and Sven Jacob and Dominik Koch and Benedikt Pöllmann and Moritz Drefs and Dionysios Koliogiannis and Christian Böhm and Konrad W. Karcz and Jens Werner and Markus Guba},
  title = {{A Novel Deep Learning Model as a Donor–Recipient Matching Tool to Predict Survival after Liver Transplantation}},
  journal = {Journal of Clinical Medicine},
  year = {2022},
  doi = {10.3390/jcm11216422},
}
Figure