Virufy aims to build a free smartphone app to detect COVID-19 from cough recordings through machine learning analysis of audio signals, which would allow for mass-scale testing of COVID-19 and could effectively stop the spread of the virus. We are building an end-to-end system that preprocesses cough (audio signals) data and gives a binary output of whether COVID-19 is present or not. This model involves audio engineering and extracting relevant features which shall be hosted on Amazon infrastructure and can be accessed through a REST API.
Over 800,000 people have died as a result of COVID-19 in the US and 5 million worldwide as of this writing. A record number of hospital beds have been occupied due to COVID-19's high infectiousness, exceeding hospital capacity and placing a tremendous burden on global healthcare systems. Despite ongoing global immunization campaigns, distribution attempts in low- and middle-income nations have been limited. Additionally, the efficacy of the current vaccinations in stopping the spread of COVID-19 has been reduced due to the emergence of new viral variations such as Omicron.
Emerging Artificial Intelligence (AI) technologies have demonstrated the capacity to develop quick, inexpensive, and available solutions. There is growing evidence that using machine learning and deep learning techniques, one may predict COVID-19 by listening to infected individuals' coughs. This study advances the development of an audio-based COVID-19 diagnostic tool by demonstrating adequate performance on a sizable and varied dataset. Future work should aim at gathering large numbers of PCR-validated samples and determining the model's generalizability to additional datasets, in addition to enhancing model performance.
The approach to data is different in this study as this is a novel approach where the data is a sound sample. For COVID-19 patients of all ages, in various environments, symptomatic and asymptomatic, and at various times relative to symptom onset, numerous research groups have been collecting sound recordings. There is a significant impact on the industry: a prospective, entirely digital COVID-19 detection approach would enable a smartphone-based quick, fair COVID-19 test with no danger of infection, financial burden, and supply chain problems—all features that are beneficial for reducing COVID-19 spread. Given the significance, it is necessary to investigate and create digital COVID-19 exams that are based on machine learning.
├── data
│ ├── virufy-cdf-coswara
│ ├── virufy-cdf-coughvid
│ ├── virufy-cdf-india-clinical-1
│ ├── model_train_data_set_1_public.csv
│ └── model_val_data_set_1_public.csv
├── documentation
│ ├── virufy-data-exploration.pdf
│ ├── virufy-infographic.pdf
│ ├── virufy-project-proposal.pdf
│ ├── virufy-poster.pdf
│ └── virufy-team-bios.pdf
├── images
│ ├── confusion-matrix.png
│ ├── data-split.png
│ ├── epoch-train-accu.png
│ ├── epoch-val-accu.png
│ ├── how-virufy-works.png
│ ├── iteration-training-loss.png
│ ├── iteration-validation-loss.png
│ ├── metric-results.png
│ ├── proposed-solution.png
│ ├── uw-logo.png
│ └── virufy-logo.png
├── model
│ ├── config.json
│ └── pytorch_model.bin
├── LICENSE
├── README.md
├── hubert_model.ipynb
└── requirements.txt
hubert_model.ipynb - Contains the documentation of the code and all the analyses conducted to arrive at the below results. Due to the dataset being under an NDA, we are unable to provide the input data to help in replicating the same.
- Data source 1: cough audio samples + inferred PCR result.
- Data source 2: cough audio samples + inferred PCR result.
- Data source 3: cough audio samples + inferred PCR result.
-
Three different data sources have been combined and collected through crowdsourcing.
-
The features extracted for this project are the cough recordings and the corresponding PCR test result (positive or negative).
- The dataset is heavily skewed to negative samples.
All the data sets are a property of Virufy and is not to be accessed by public.
This work is licensed under a Creative Commons Attribution 4.0 International License.
Link to the privacy policy
- Our HuBERT model accurately predicts COVID-19 from cough samples with a high true positive rate and low false positive rate.
- However, the model fails to predict positive COVID-19 cases about ⅕ times, which may cause users to overlook preventive measures.
- Validation dataset accuracy improves over iteration, but the model appears to overfit after Epoch 14.
- Lack of data and an imbalanced dataset are major issues in training the model. We only had 754 positive COVID-19 samples, which is not enough for proper tuning.
- We expect HuBERT to generalize better and false positive rate and AUC to increase as new positive samples are collected.
- Although our AUC score falls below the best-performing model's score, we cannot rule out the usefulness of HuBERT until more positive samples are collected.
The use of AI for predicting if someone has COVID-19 could potentially have several benefits. For example, it could help healthcare providers more quickly and accurately diagnose patients, potentially leading to earlier treatment and better outcomes. It could also help healthcare systems prioritize the allocation of resources, such as testing and personal protective equipment, to those who are most in need.
However, there are also potential risks and limitations to using AI for COVID-19 prediction. One major concern is the accuracy of the predictions. AI algorithms are only as good as the data they are trained on, and the accuracy of their predictions can be affected by a variety of factors, such as the quality and diversity of the training data. There is also the potential for bias in the data or the algorithm, which could lead to unfair or inaccurate predictions.
Another potential risk is the ethical concerns around using AI for medical diagnosis. Some people may be uncomfortable with the idea of relying on a machine to make important medical decisions, and there are valid concerns about the potential for errors or misdiagnoses.
Overall, the use of AI for predicting COVID-19 infections could have potential benefits, but it is important to carefully consider the risks and limitations and ensure that any AI systems are developed and used responsibly.
- HuBERT model's poor generalization to unseen data is due to an imbalanced dataset.
- Insufficient positive COVID-19 samples prevent proper tuning of the model.
- Current model's performance is not competitive with the best-known models for this task.
- Viability of the model will not be ruled out until trained on a reasonable number of samples, and the main goal of future work is collecting more positive samples.
[1] Haritaoglu, E. D., Rasmussen, N., Tan, D. C. H., J., J. R., Xiao, J., Chaudhari, G., Rajput, A., Govindan, P., Canham, C., Chen, W., Yamaura, M., Gomezjurado, L., Broukhim, A., Khanzada, A., & Pilanci, M. (2022, March 30). Using deep learning with large aggregated datasets for covid-19 classification from cough. arXiv.org. Retrieved November 22, 2022, from https://arxiv.org/abs/2201.01669.
[2] Nemati, E., Rahman, M. M., Nathan, V., Vatanparvar, K., & Kuang, J. (2020, July). A comprehensive approach for classification of the cough type. In 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) (pp. 208-212). IEEE.
[3] Mouawad, P., Dubnov, T., & Dubnov, S. (2021). Robust detection of COVID-19 in cough sounds. SN Computer Science, 2(1), 1-13.
[4] Chaudhari, G., Jiang, X., Fakhry, A., Han, A., Xiao, J., Shen, S., & Khanzada, A. (2020). Virufy: Global applicability of crowdsourced and clinical datasets for AI detection of COVID-19 from cough. arXiv preprint arXiv:2011.13320.
[5] Fakhry, A., Jiang, X., Xiao, J., Chaudhari, G., Han, A., & Khanzada, A. (2021). Virufy: A multi-branch deep learning network for automated detection of COVID-19. arXiv preprint arXiv:2103.01806.
[6] Feng, K., He, F., Steinmann, J., & Demirkiran, I. (2021, March). Deep-learning based approach to identify covid-19. In SoutheastCon 2021 (pp. 1-4). IEEE.
- PEP 8 – Style Guide for Python Code (Reference link)
- Use of relative path addresses to help in reproducibility
- Use of intuitive variable and function names to ease in understanding
- Appropriate comments and documentation provided for the data aquisition, data processing and data analysis steps
- Description of all data files present in the repository mentioned