AirwaySymptomDetection

This repository contains related code to the paper below.

How to use the code

To use the code, you need a Python installation together with relevant libraries (librosa, numpy, pandas, flammkuchen, scikit-learn, tensorflow).

Overview

Data preprocessing (Dataset creation)

Data preprocessing can be found in neural_networks/src/dataloader.py. Part of it is generating Mel-spectrograms from 1-D microphone-acoustic and mechano-acoustic data.

Training deep neural networks (DNN mining)

We provide code to train several deep neural network architectures (neural_networks), e.g., ResNet, EfficientNet or RNNs. In analysis, you find a Jupyter notebook for evaluating the trained models.

The file neural_neworks/params.json offers the possibility to specify various hyperparameters, especially with regard to the preprocessing of the data.

Aditionally, the repository provides code for training an autoencoder architecture (neural_networks) for converting from microphone-acoustic to mechano-acoustic Mel-spectrograms.

Evolutionary optimization for wearable deployment

We provide code to run a genetic algorithm (neural_networks/GeneticAlgorithm) to optimize and find a low-size, accurate deep neural network architecture.

Genetic Algorithm model's performance on unseen data and learning capabilities

In analysis, you find two Jupyter notebooks for evaluating the Objective 2 model, which was determined using the Genetic Algorithm.

Used datasets:

• Lei et al., 2020 - predict-new-dataset-mcgill.ipynb
• COUGHVID crowdsourcing dataset - predict-cough-database.ipynb

Explainable AI for mining AI decisions

In analysis, you find Jupyter notebooks for visualizing the class activation maps and results from occlusion experiments.

How to cite this code

Groh et al. "Efficient and Explainable Deep Neural Networks for Airway Symptom Detection in Support of Wearable Health Technology", 2021