Cross-dataset evaluation of dementia longitudinal progression prediction models

References

• Zhang, C., An, L., Wulan, N., Nguyen, K. N., Orban, C., Chen, P., Chen, C., Zhou, J. H., Liu, K., Yeo, B. T. T., 2024. Cross-dataset Evaluation of Dementia Longitudinal Progression Prediction Models, medRxiv

---

Background

Accurate Alzheimer’s Disease (AD) progression prediction is essential for early intervention. The TADPOLE challenge, involving 92 algorithms, used multimodal biomarkers of observed historical visits to predict future clinical diagnosis, cognition, and brain atrophy. The winning algorithm, FROG, employed a Longitudinal-to-Cross-sectional (L2C) transformation to convert variable longitudinal histories into fixed-length feature vectors—contrasting with most existing approaches that model entire longitudinal sequences (e.g., AD Course Map and MinimalRNN). In this project, we extend FROG by unifying its XGBoost components into a single feedforward neural network (L2C-FNN), and systematically evaluate its generalizability by training on the ADNI dataset and testing on three external cohorts—covering 2,312 participants and over 13,200 timepoints. L2C-FNN demonstrates strong and consistent performance across datasets, underscoring its potential for long-term dementia progression prediction.

---

Folder structure

• data_processing: Contains Python files for data splitting (train/validation/test), L2C feature transformation, and preprocessing to prepare data for model input.
• examples: Contains scripts for generating demo data and example use of the models.
• models: Contains Python files for training and evaluating the models.
• predict_only: Contains scripts for predicting dementia progression using pretrained L2C-FNN models on your own dataset.
• readme_figures: Graphical abstract of the paper, providing high-level understanding of pipeline.
• replication: Contains scripts for replicating the results.
• unit_tests: Uses examples demo data to check validity of the code.
• utils: Helper functions for evaluating model performance and comparing against reference results.

---

Code Release

Download stand-alone repository

Since the whole Github repository is too big, we provide a stand-alone version of only this project and its dependencies. To download this stand-alone repository, visit this link: https://github.com/ThomasYeoLab/Standalone_Zhang2025_L2CFNN

Download whole repository

If you want to use the code from our lab's other stable projects (other than Zhang2025_L2CFNN), you would need to download the whole CBIG repository.

• To download the version of the code that was last tested, you can either

  • visit this link: https://github.com/ThomasYeoLab/CBIG/releases/tag/v0.35.0-Zhang2025_L2CFNN

  or

  • run the following command, if you have Git installed

  git checkout -b Zhang2025_L2CFNN v0.35.0-Zhang2025_L2CFNN
  

---

Usage

Environment setup

• Our code uses Python, here is the setup:
  1. Install Miniconda or Anaconda with Python 3.x if you don't have conda
  2. Create conda environment from our yml/txt files in replication/config file by:

     cd $CBIG_CODE_DIR/stable_projects/predict_phenotypes/Zhang2025_L2CFNN
     conda env create -f replication/config/CBIG_L2CFNN_python_env.yml
     bash replication/scripts/CBIG_L2CFNN_install_env.sh
     

Example

• The example of our code is detailed in examples/README.md

Replication

• If you have access to ADNI, AIBL, MACC, and OASIS datasets, you can replicate our results using the instructions detailed in replication/README.md.

Predict_only

• If you would like to use the pretrained ADNI models on your own dataset to perform predictions, please refer to detailed instructions in predict_only/README.md

---

Updates

• Release v0.35.0 (03/06/2025): Initial release of Zhang2025_L2CFNN project

---

Bugs and Questions

Please contact Chen Zhang at [email protected] and Thomas Yeo at [email protected]