TEMPO

This repository contains codes, data and models to reproduce the results reported in the submission of TEMPO: Transformers for Temporal Disease Progression from Cross-Sectional Data.

Additionally,

• https://github.com/hongtaoh/TEMPO_lowdim contains the reproducible codes for the low dimensional experiments for the benchmarking algorithms.
• https://github.com/hongtaoh/TEMPO_highdim contains the reproducible codes for the high dimensional experiments for the benchmarking algorithms.

Table of Contents

• Installation
• Hardware Requirements
• Quick Start
• Step-by-Step Guide
• Running on ADNI Data

Installation

First, git clone this repository. Then create a virtual environment and install dependences:

uv venv --python python3.11
source .venv/bin/activate
uv pip install -r requirements.txt

Hardware Requirements

The scripts are optimized for GPU execution. While CPU execution is supported, expect significantly longer runtimes.

Quick Start

From the root directory, run these three commands in order:

bash gen.sh           # Generate all synthetic data (lowdim + highdim, train + test)
bash run_all.sh       # Train all models (lowdim then highdim)
bash postprocess.sh   # Run inference, flatten results, and generate plots

Before training, set the gpus: block in config.yaml to match your available GPUs:

gpus:
  0: [exp1, exp2, exp3]
  1: [exp4, exp5, exp6, exp7, exp8, exp9]

That's it. No manual commenting/uncommenting of config files required.

Step-by-Step Guide

1. Generating Synthetic Data

bash gen.sh

This generates all four data directories (train, test, train_highdim, test_highdim). Already-existing directories are skipped automatically.

To regenerate from scratch, delete the directories first:

rm -rf train test train_highdim test_highdim
bash gen.sh

To generate only one dimension:

python3 gen.py --mode lowdim
python3 gen.py --mode highdim

Additional parameters (number of datasets, participants, experiments, etc.) can be modified in config.yaml.

2. Training

Set the gpus: block in config.yaml to assign experiments to GPUs, then run:

bash run_all.sh

This trains lowdim models first (saving to models/), waits for completion, then trains highdim models (saving to models_highdim/). Training logs are saved to the logs/ folder. If a models directory already exists and is non-empty, that step is skipped automatically.

If running on a remote server or want training to survive terminal close, use:

mkdir -p logs && nohup bash run_all.sh > logs/run_all.log 2>&1 &

To retrain from scratch:

rm -rf models models_highdim
bash run_all.sh

To train only one dimension:

bash run_parallel.sh --train_data_dir ./train --models_dir ./models
bash run_parallel.sh --train_data_dir ./train_highdim --models_dir ./models_highdim

To stop running jobs:

bash kill.sh

3. Inference, Results Processing, and Visualization

bash postprocess.sh

This runs three steps automatically:

1. Inference — evaluates trained models on all test experiments, saving results to:

   • results/cross_experiment_results_12bio.json
   • results/cross_experiment_results_100bio.json

2. Flatten results — extracts diagonal (same-experiment) results into:

   • results/12bio_tempo.csv
   • results/100bio_tempo.csv

3. Plots — merges TEMPO results with benchmark algorithm results and generates figures in plots/12bio/ and plots/100bio/.

Benchmark algorithm results are available in the benchmark_algo_results/ folder.

Running on ADNI Data

You'll need to request access to the ADNI data through https://adni.loni.usc.edu/data-samples/adni-data/. Download the ADNIMERGE.csv. Put it in the root directory of this repository. Then

python process_adni.py --raw ADNIMERGE.csv

You will see adni.csv and also id_dx.json. Then, run

python3 run_adni.py

Results will be saved to the adni_results folder.