[論文レビュー] Multi-Task Learning for Post-transplant Cause of Death Analysis: A Case Study on Liver Transplant

Sirui Ding, Qiaoyu Tan+6 more|PubMed|Mar 30, 2023

Machine Learning in Healthcare参考文献 28被引用数 8

ひとこと要約

The paper introduces CoD-MTL, a tree-distillation multi-task learning framework that jointly predicts post-transplant causes of death (rejection and infection) in liver transplant patients, improving accuracy over traditional models and single-task baselines.

ABSTRACT

Organ transplant is the essential treatment method for some end-stage diseases, such as liver failure. Analyzing the post-transplant cause of death (CoD) after organ transplant provides a powerful tool for clinical decision making, including personalized treatment and organ allocation. However, traditional methods like Model for End-stage Liver Disease (MELD) score and conventional machine learning (ML) methods are limited in CoD analysis due to two major data and model-related challenges. To address this, we propose a novel framework called CoD-MTL leveraging multi-task learning to model the semantic relationships between various CoD prediction tasks jointly. Specifically, we develop a novel tree distillation strategy for multi-task learning, which combines the strength of both the tree model and multi-task learning. Experimental results are presented to show the precise and reliable CoD predictions of our framework. A case study is conducted to demonstrate the clinical importance of our method in the liver transplant.

研究の動機と目的

• Motivate accurate post-transplant cause-of-death analysis to aid organ allocation and personalized treatment.
• Address data imbalance and task correlation by using a multitask framework for related CoD prediction tasks.
• Leverage a tree-distillation strategy to transfer tree-model strengths into neural networks for multi-task learning on EHR data.
• Evaluate the method on a real liver transplant cohort to demonstrate clinical relevance and predictive performance.

提案手法

• Preprocess EHR features by encoding categorical data, concatenating with numerical features, and imputing missing values with zeros.
• Develop a multi-task learning framework (CoD-MTL) with a shared representation and task-specific heads for multiple CoD tasks (rejection and infection).
• Introduce a tree-distillation strategy that distills outputs from GBDT models into a neural network via dense embeddings of tree leaves, enabling end-to-end multi-task learning.
• Train task-specific distilled networks NN_j with a shared representation for all CoD tasks and combine them with a multi-task loss that balances distillation and prediction terms.
• Integrate distillation targets from per-task GBDT models into a unified multi-task neural architecture to improve performance on tabular EHR data.

実験結果

リサーチクエスチョン

• RQ1 RQ1: Can CoD-MTL accurately predict rejection and infection as post-transplant CoDs?
• RQ2 RQ2: Is the CoD-MTL model trustworthy with well-calibrated probability estimates and stable performance?
• RQ3 RQ3: How can CoD-MTL support clinical decision-making in liver transplantation?

主な発見

Model	CoD: Rejection AUROC	CoD: Rejection AUPRC	CoD: Infection AUROC	CoD: Infection AUPRC
Logistic Regression	0.551±0.008	0.482±0.005	0.569±0.013	0.471±0.013
GBDT	0.588±0.008	0.497±0.010	0.611±0.011	0.499±0.014
Random Forest	0.583±0.016	0.504±0.009	0.608±0.009	0.506±0.020
Neural Networks (single task) - MLP	0.571±0.012	0.493±0.008	0.592±0.003	0.483±0.011
Multitask Learning (baseline)	0.595±0.021	0.517±0.015	0.614±0.019	0.515±0.028
CoD-MTL	0.640±0.012	0.557±0.012	0.646±0.007	0.553±0.018

• CoD-MTL achieves higher AUROC and AUPRC than all baselines for both rejection and infection predictions.
• Tree-based models outperform single-task neural nets, and multitask learning further improves performance over single-task baselines.
• CoD-MTL shows up to 16.1% AUROC improvement for rejection and 15.6% AUROC improvement for infection over best baselines; similar gains are seen in AUPRC.
• Calibration analyses indicate CoD-MTL probabilities are well-calibrated with low uncertainty across cross-validation folds.
• Case study illustrations suggest the model can provide nuanced, patient-specific risk profiles to aid clinical decisions.
• Limitations discussed include interpretability and fairness, with future work aiming to address explainable AI and equity constraints.