The long-term survival of liver transplant recipients beyond one year is significantly compromised by an increased risk of cancer, cardiovascular mortality, infection and graft failure. There are currently limited clinical tools to identify patients at risk of these complications, which would flag them for screening tests and life-saving interventions.

We hereby propose Deep Learning models designed for longitudinal data that reliably predicts an updated clinical outlook for individual patients. Here is an example of how our top-performing Transformer model estimates the risk progression of a patient more than 20 years post-transplant, accurately outlining the top complications.

First, install dependencies.

# clone project   
git clone git@github.com:bowang-lab/Transplant_Time_Series.git

# install project environment
cd Transplant_Time_Series
conda env create -f environment.yml

Next, download the transplant data from the SRTR Database and process it accordingly.

https://www.srtr.org/about-the-data/the-srtr-database/

Then, activate the environment and run the algorithms.

conda activate transplant

# run model under different modules, add hyperparameter accordingly
python train.py --epochs 30 --batch 256


# check `Models/<Transformer>/train.py` for more args

A DL-based Transformer model was developed and trained on a set of 42,146 LT recipients (median age 53, IQR 45-59 years; 40.8% women) from the publicly available Scientific Registry of Transplant Recipients (SRTR).

The transferability of the model was further evaluated by testingfine-tuning on a local dataset from University Health Network in Toronto, Canada, consisting of 3,269 patients (median age 54, IQR 46-61 years; 33.0% women).

The area under the receiver operating characteristic curve (AUROC) for the top-performing Transformer Model across all outcomes in the SRTR dataset was 0.804, 99% CI [0.795, 0.854] (1 year) and 0.733, 99% CI [0.729, 0.769] (5 years). In the UHN dataset, the top deep learning AUROC was 0.807, 99% CI [0.795, 0.842] (1 year) and 0.722, 99% CI [0.705, 0.764] (5 years). AUROCs ranged from 0.695 for 5-year infection death to 0.856 for 1-year graft failure.

We compared the performance of our Transformer model with other DL-based models (Temporal Convolutional Network model, Recurrent Neural Network model) as well as traditional logistic regression (LR) models. The Transformer model outforms the LR models by significant margin.