Skip to content

Latest commit

 

History

History
40 lines (28 loc) · 2.65 KB

README.md

File metadata and controls

40 lines (28 loc) · 2.65 KB

PREPARE-Challenge-Acoustic-Track-Code

Official Code of illidanlab team for PREPARE Challenge 2024 (Pioneering Research for Early Prediction of Alzheimer's and Related Dementias EUREKA Challenge 2024).

Overview

featured Early detection of cognitive diseases is critical for effective treatment. While brain imaging has demonstrated high accuracy in disease detection, its widespread use is limited by complex technology and high costs, making it impractical for large-scale screening of adults. The PREPARE Challenge aims to develop effective methods leveraging acoustic biomarkers from patients' speech samples to identify cognitive characteristics. Our approach combines advanced acoustic feature extraction techniques, utilizing Wav2Vec and MFCC features, with robust outlier detection to mitigate noise. These features were then used in conjunction with Support Vector Machines (SVM), resulting in a competitive 10th-place ranking. Our codes are provided in https://github.com/illidanlab/PREPARE-Challenge-Acoustic-Track-Code.

Package dependencies

Use pip install -r requirements.txt to install necessary libraries.

Data Preparation

Download acoustic data from PREPARE Challenge website in https://www.drivendata.org/competitions/299/competition-nih-alzheimers-acoustic-2/ and put it in folder named data.

Demos

Here we provide several demos of results in the project report. You can change the arguments from main.py to try different settings.

  • --acoustic (list of strings, optional, default: ["Wav2Vec", "MFCC"]):

    • Defines the acoustic features to use.
    • Options include: "Wav2Vec", "MFCC".

  • --model (string, optional, default: "SVC"):

    • Specifies the classifier to use.
    • Options include: "MLP", "SVC", "GBC", "RF", "LR", "NB".

  • --remove_noise (flag, optional):

    • Specifies type of noise to remove
    • Options include: "all", "short", "language", "alexa", "svm", "none".

  • --grid_search (flag, optional):

    • If set, perform GridSearch for the best hyperparameters of classifiers.
    • Use as --grid_search to enable.

Examples

  • Use MFCC as feature and Logistic Regression as classifier: python main.py --acoustic MFCC --model LR
  • Remove noisy samples in training data: python main.py --remove_noise
  • Use Wav2Vec + MFCC as features, Support Vector Classifier as classifier, Grid Search for hyper-parameters, and remove all noisy samples: python main.py --acoustic Wav2Vec MFCC --model SVC --remove_noise all --grid_search