Audio Respiratory Disease Prediction

Inspiration

One of top 3 death causes by coronavirus Covid-19 is Chronic obstructive pulmonary disease brieflly COPD[4]. COPD' diagnosis happens by breath oscultation. We would like to create an intelligent software that can provide medical diagnosis by just oscultating the breath recorded by a smartphone.

What it does

Predicts whther a recorded breath is healthy or is affected by some major pathologies such as COPD.

How we built it

The dataset is taken from:

Α Respiratory Sound Database for the Development of Automated Classification Rocha BM, Filos D, Mendes L, Vogiatzis I, Perantoni E, Kaimakamis E, Natsiavas P, Oliveira A, Jácome C, Marques A, Paiva RP (2018) In Precision Medicine Powered by pHealth and Connected Health (pp. 51-55). Springer, Singapore.

Here is the pipeline:

1- 4k resampling

2 - filtering ebu normalization

3 - blind source separation

4 - segmentation

5 - noise removal with stacked denoising autoencencoders

6 - data augmentation via VAE / GAN && or ratation, shifting & padding

7 - mel spectrogram, gammatone spectrogram, wavelet spectrogram, optimized s transform each of 128x128. The resulting image is composed by stacking horizontally and vertically the 4 spectrogram images into a 256x256 color image. Another solution which gave better results was on working on novel 1D ad hoc modified ResNet especially designed for audio.

8 - hand made modified ResNet for classification

The code runs on google colab but will be soon released on github after the pre-print of the paper is uploaded on preprinting services such as ArXiv.

Challenges we ran into

VAE/GAN needs lots of data

the majority of reviewed papers use 70-30 split without using the official test set provided by the challenge. Thus they don't use a inter-patient separation scheme, revealing wrong results.

Accomplishments that we're proud of

The segmentation works!! For the classification at moment I have 95% accuracy in binary classification (healthy vs pathology) with interpatient separation scheme testing on the official test set provided originally from the paper.

What we learned

Probably managing Audio deep learning as a 1D time series, instead of transforming it in images (as done in the majority of the research reviewed) is effective, but requires a specialized network architecture.

What's next for Audio Respiratory Disease Prediction

build a product but for now...

Here is the demo

https://159.203.68.29:5000

Since the https is not official, please click on proceed in any case (insecure). Thanks

MEDICAL DISCUSSION

Cardiac injury is a common condition among patients with COVID-19, and it is associated with a higher risk of in-hospital mortality.The findings presented in the research on 416 patients in Wuhan, Shi and colleagues [2], highlighted the need to consider cardiac complication in COVID-19 management. The patients with Covid-19 have respiratory distress and low blood oxygen levels, consequently they have high risk of ischemia or heart attack that compromises myocardial contractility and this situation can cause severe arrhythmia.

Respiratory diseases detection was conducted in the following way: We used the data from paper [1] Filter noise and remove background sounds We used several segmentation techniques based on spectrograms of respiratory audio Segmentation is necessary to understand when the first respiratory cycle starts. We used a custom deep learning neural network to predict binary healthy/unhealthy patients using only audio recorded by smartphone Developed a web service and web page to record and show the result. The accuracy is tested on the official test set in [1]

Why is early detection of COPD important for covid19?

SARS-CoV-2 uses the angiotensin converting enzyme II (ACE-2) as the cellular entry receptor to infect the lower respiratory tract. ACE-2 expression in lower airways is increased in patients with COPD and with current smoking [3] . ACE-2 is expressed in a variety of different tissues including both the upper and lower respiratory tract and myocardium. Importantly, nearly all deaths have occurred in those with significant underlying chronic diseases including COPD, and cardiovascular diseases These findings highlight the importance of increased surveillance of these risk subgroups for prevention and rapid diagnosis of this potentially deadly disease.

References

1- Rocha, B. M., Filos, D., Mendes, L., Vogiatzis, I., Perantoni, E., Kaimakamis, E., ... & Paiva, R. P. (2018). Α respiratory sound database for the development of automated classification. In Precision Medicine Powered by pHealth and Connected Health (pp. 33-37). Springer, Singapore.

2- Shi S, Qin M, Shen B, et al. Association of Cardiac Injury With Mortality in Hospitalized Patients With COVID-19 in Wuhan, China. JAMA Cardiol. Published online March 25, 2020. doi:10.1001/jamacardio.2020.0950

3- ACE-2 Expression in the Small Airway Epithelia of Smokers and COPD Patients: Implications for COVID-19. Janice M Leung, Chen Xi Yang, Anthony Tam, Tawimas Shaipanich, Tillie L Hackett, Gurpreet K Singhera, Delbert R Dorscheid, Don D Sin. Published in European Respiratory Journal doi: 10.1183/13993003.00688-2020

4 - Halpin, D. M., Faner, R., Sibila, O., Badia, J. R., & Agusti, A. (2020). Do chronic respiratory diseases or their treatment affect the risk of SARS-CoV-2 infection?. The Lancet Respiratory Medicine.