madress-2023

1. Overview

Source code for LCN submission for ADReSS-M challenge (formerly called MADReSS).

The model was trained on 228 English samples of a picture description task and was transferred to Greek using only 8 samples. We obtained an accuracy of 82.6% for AD detection, a root-mean-square error of 4.345 for cognitive score prediction, and ranked 2nd place in the competition out of 24 competitors.

AD Detection Models (%) [Link]

• Model0 - 76.1%
• Model1 - 71.7%
• Model2 - 82.6%
• Model3 - 80.4%
• Model4 - 73.9%

Cognitive Score Prediction Models (RMSE) [Link]

• Model0 - 4.716
• Model1 - 4.713
• Model2 - 4.816
• Model3 - 4.345
• Model4 - 4.837

2. Installation

Note: this software was developed for Linux.

Clone Repository

git clone https://github.com/lcn-kul/madress-2023.git
cd madress-2023

Create Virtual Environment

make create_environment
source venv/bin/activate
make requirements

3. Reproducing Results

Run the following commands to reproduce the results.

1. Download challenge data

See the raw data for more information.

2. Process raw CSVs

make csvs

3. Do training

This step performs

• eGeMAPS feature extraction
• English pretraining
• English+Greek fine-tuning
• model-averaging
• test set prediction

for the AD and MMSE models. The feature extraction takes 10-15 minutes and the other steps will take another 10-15 minutes.

make train

4. Prepare submission files

Insert predictions from the folders

• models/trained_model_[config]_avg/prediction.txt

into the corresponding submission format file:

• AD : data/raw/ADReSS-M-format_task1.csv
• MMSE : data/raw/ADReSS-M-format_task2.csv