14 February 2022

Introduction

  • Histopathological diagnosis of graft rejection faces serious limitations in terms of accuracy and risk prediction.

  • High-throughput gene expression profiling can provide evidence of graft rejection before a clinical phenotype becomes apparent.

  • Combining decisions from a diverse set of algorithms can result in more stable predictions.

  • Microarray data are susceptible to the curse of dimensionality.

  • Determining the best set of hyperparameters for an ensemble can be challenging

Workflow

Experiments

  • Intercom Study
  • Molecular Microscope
  • Callemeyn Study

Intercom Study

Intercom Optimisation Results

Intercom Classification Results

Log-loss

Molecular Microscope Study

Molecular Microscope Optimisation Results

Molecular Microscope Classification Results

Log-loss

Callemeyn Study

Callemeyn Optimisation Results

Callemeyn Classification Results

Log-loss

(re)ComBat Analysis

Used (re)ComBat to remove batch effects when combining the datasets to increase the size of the data set being used to analyse the classification capability of the (SA)BIO pipelines.

reComBat

Informing batch and class (desired variation):

Preliminar reComBat Optimisation Results

–> –> –> –> –> –> –> –> –> –>

–> –> –> –> –> –> –> –> –> –>

–> –> –> –> –> –> –> –> –> –>

CTGAN Experiment

Conditional tabular GAN from The Synthetic Data Vault to generate ABMR samples.

Current Work

  • More replicates of the reComBat experiment – running

  • Multiclass pipeline – running tests on molecular microscope data (ABMR, TCMR (+MIXED), CO)

  • CTGAN approach to generate ABMR synthetic data and compare with ADASYN, SMOTE(Tomek)

Thanks!