Training Data Composition Experiments

This repository contains the scripts and processed results used for the five experiments investigating the influence of antimicrobial susceptibility testing (AST) device selection and training dataset composition on machine learning-based antimicrobial resistance prediction.

Requirements

All scripts are provided as Jupyter notebooks (.ipynb) and were executed using Python 3.14.

The required Conda environment can be created from the provided environment.yml file:

conda env create --name data_comp_env --file environment.yml

After installation, activate the environment using:

conda activate data_comp_env

The machine learning predictions were generated with stackPredAMR, which is available at:

https://github.com/IKIM-Essen/WIN-KID

Repository Structure

The repository is organized by experiment. Each experiment is located in its own directory (e.g. Experiment_1, Experiment_2, ...).

Each experiment contains the following subdirectories:

input_data

Contains the input files used for stackPredAMR together with the scripts that generated these datasets from the original source data.

results

Contains the scripts used for post-processing and analysis, including:

• merging the raw stackPredAMR output files
• calculating the final evaluation metrics
• generating the figures used in the manuscript
• statistical analyses (where applicable)

The merged result files produced by these scripts are also included in this directory.

Some experiments contain additional analysis scripts specific to the corresponding experiment. For example, Experiment 4 includes the implementation of the Radius Neighbors Classifier (neighbors_analysis/radius_neighbors_analysis.ipynb) together with the UMAP visualization (umap_plot.ipynb), while Experiment 5 contains the statistical analysis used for the Wilcoxon signed-rank tests (statistics.ipynb).

results/stackPredAMR_Results

Contains the raw prediction results generated by stackPredAMR for the corresponding experiment.

Data

This repository contains only the data required to reproduce the analyses presented in the manuscript. The underlying machine learning framework (stackPredAMR) is maintained in the separate repository linked above.

Throughout this repository, susceptibility records without documented AST device information are referred to as NA. These records correspond to the Unclassified category described in the manuscript.