Source-linked AI summary
Exploring Automatic Diagnosis of COVID-19 from Crowdsourced Respiratory Sound Data
Chloë Brown, Jagmohan Chauhan, Andreas Grammenos, Jing Han, Apinan Hasthanasombat, Dimitris Spathis, Tong Xia, Pietro Cicuta, Cecilia Mascolo
TL;DR
The paper asks whether crowdsourced respiratory sounds can provide signals for COVID-19 screening despite limited prior evidence from large uncontrolled datasets. It analyzes cough and breathing recordings from COVID-19-positive and control users with shallow classifiers, finding AUC above 80% across three tasks, while emphasizing that the results are preliminary and not sufficient for standalone screening.
Problem
Large-scale evidence is limited on whether crowdsourced respiratory sounds can distinguish COVID-19 from healthy users and users with asthma.
Method
The study analyzes a curated subset of crowdsourced cough and breathing recordings using feature-based, shallow machine-learning classifiers and defined control groups.
Results
Above 80% AUC is achieved across all three binary tasks, including COVID-19 versus healthy users, cough versus healthy cough, and cough versus asthma cough.
Takeaways & Limitations
The findings provide preliminary evidence that analyzed respiratory patterns could aid COVID-19 pre-screening.
Takeaways & Limitations
The results are not solid enough for a standalone screening tool because the analysis uses limited data and lacks ground truth for health status.
Abstract
from arXiv · showhide
Audio signals generated by the human body (e.g., sighs, breathing, heart, digestion, vibration sounds) have routinely been used by clinicians as indicators to diagnose disease or assess disease progression. Until recently, such signals were usually collected through manual auscultation at scheduled visits. Research has now started to use digital technology to gather bodily sounds (e.g., from digital stethoscopes) for cardiovascular or respiratory examination, which could then be used for automatic analysis. Some initial work shows promise in detecting diagnostic signals of COVID-19 from voice and coughs. In this paper we describe our data analysis over a large-scale crowdsourced dataset of respiratory sounds collected to aid diagnosis of COVID-19. We use coughs and breathing to understand how discernible COVID-19 sounds are from those in asthma or healthy controls. Our results show that even a simple binary machine learning classifier is able to classify correctly healthy and COVID-19 sounds. We also show how we distinguish a user who tested positive for COVID-19 and has a cough from a healthy user with a cough, and users who tested positive for COVID-19 and have a cough from users with asthma and a cough. Our models achieve an AUC of above 80% across all tasks. These results are preliminary and only scratch the surface of the potential of this type of data and audio-based machine learning. This work opens the door to further investigation of how automatically analysed respiratory patterns could be used as pre-screening signals to aid COVID-19 diagnosis.
1 INTRODUCTION
The paper investigates whether crowdsourced cough and breathing sounds contain signals that distinguish COVID-19 from healthy controls, asthma, and other coughs. Across three binary classification tasks, preliminary models achieve AUC above 80%.
- Crowdsourced respiratory sounds are analyzed to assess whether COVID-19 can be distinguished from healthy users and users with asthma.
- The dataset supports comparisons between COVID-19-positive users, healthy users, and asthma users, including matched cough comparisons.
- The work presents preliminary evidence that automatically analyzed respiratory sounds may support COVID-19 pre-screening.
- 80% AUC is achieved for distinguishing healthy and COVID-19 sounds in Task 1.
- 82% AUC is achieved when distinguishing COVID-19-positive users with coughs from healthy users with coughs in Task 2.
2 MOTIVATION AND RELATED WORK
The paper situates automated respiratory-sound analysis within a broader shift from clinician-mediated auscultation toward digital sensing and machine learning. It differs from prior COVID-19 studies by using a large, uncontrolled crowdsourced dataset.
- Digital technologies increasingly collect bodily sounds for automated analysis, including respiratory examination and asthma wheeze detection.
- Prior COVID-19 studies analyzed digital-stethoscope lung sounds, phone-recorded coughs, and speech from hospital patients.
- Earlier respiratory-sound datasets were often limited to coughs or smaller samples from specific regions or hospitals.
- This work uses an entirely crowdsourced dataset and must account for self-reported ground truth, different phones, microphones, and environments.
- The app collected samples from more than 7000 unique users, including more than 200 users positive for COVID-19, and enabled repeat sampling.
3 DATA COLLECTION
The study gathers demographic, medical, symptom, location, and respiratory-sound data through web and Android apps, then analyzes a curated subset focused on coughs and breathing. The dataset includes COVID-19-positive, non-COVID, non-COVID-with-cough, and asthma-with-cough groups.
- The web and Android apps collect demographics, medical history, symptoms, COVID-19 testing status, location, coughs, breathing, and speech recordings.
- As of 22 May 2020, the dataset contained 4352 web users and 2261 Android users, with 235 users reporting a positive COVID-19 test.
- Dry cough was the most common single symptom overall, while lack of smell and chest tightness were the most common symptom combination among COVID-19-positive users.
- The analysis uses a curated subset restricted to cough and breathing samples, excluding voice recordings because of COVID-19-positive-user imbalance.
- The control groups comprise 298 non-COVID samples, 32 non-COVID-with-cough samples, and 20 asthma-with-cough samples.
4 METHODS
The paper extracts handcrafted and transfer-learned representations from cough and breathing recordings, then reduces combined features with PCA for shallow classification. The feature pipeline combines temporal, spectral, statistical, and VGGish-derived information.
- Modeling: Shallow, feature-based classifiers are used because the dataset is moderately sized and interpretability is important for public-health applications.The methodology also accounts for dataset-specific characteristics such as longitudinal users and cross-validation.
- Feature extraction: The pipeline resamples recordings and extracts handcrafted audio features at frame and segment levels.These features cover frequency-based, structural, statistical, and temporal attributes.
- Handcrafted features: 477 handcrafted features summarize duration, onset, tempo, period, frame-level descriptors, and MFCC statistics.Time-series features are represented using statistics including mean, median, extrema, quartiles, dispersion, skewness, and kurtosis.
- Transfer learning: VGGish converts 0.96-second waveform segments into 128-dimensional embeddings, aggregated into 256-dimensional segment features.The pretrained model operates at a 16 KHz sampling rate and uses segment-level means and standard deviations.
- Feature combination and reduction: Combined representations concatenate handcrafted and VGGish features, reaching up to 733 dimensions before PCA.PCA retains a selected proportion of explained variance to reduce the feature vectors.
5 EVALUATION
The evaluation tests three binary classification tasks using cough, breathing, or combined modalities with user-disjoint validation. Results show discriminatory signals across COVID-19 comparisons, while some findings remain preliminary because of data size and variability.
- Classification tasks: Three binary tasks compare COVID-positive users with non-COVID users, COVID-positive users with cough against non-COVID users with cough, and COVID-positive cough users against asthma cough users.The evaluation focuses on clinically meaningful distinctions defined from users’ reported test status, symptoms, and medical history.
- Evaluation protocol: User-based cross-validation keeps samples from the same user out of both training and test sets and balances the test set.Ten outer random user splits use an 80%/20% partition, with inner five-fold cross-validation for hyperparameter search.
- Sensitivity analysis: Adding age and sex did not substantially change results, with effects smaller than ±2 AUC.The reported demographic sensitivity was limited to onehot-encoded age and sex features.
- Task 1: 80% AUC for Task 1, with precision and recall around 70%, indicates discriminatory signals for COVID-positive versus non-COVID users.This task used a simple Logistic Regression classifier and had the lowest standard deviations across user splits.
- Feature ablation: Handcrafted features combined with VGGish features performed better than either representation alone.This result supports combining learned spectrogram-based features with handcrafted information.
- Task 2: 82% AUC and 80% precision for Task 2 show that cough sounds distinguish COVID-positive users with cough from non-COVID users with cough.Recall was 72%, so the model detected many, but not all, COVID-19 coughs.
- Task 3: 80% AUC for Task 3 distinguishes COVID-positive users with cough from users with asthma and cough, with breathing sounds serving as stronger signals in this task.Precision was high and recall was described as acceptable.
6 DISCUSSION AND CONCLUSIONS
The study presents encouraging but preliminary evidence that crowdsourced respiratory sounds may aid COVID-19 diagnosis, while emphasizing that the current results are not sufficient for standalone screening. Important limitations include limited data, absent health-status ground truth, uninvestigated disease progression, and an app that provides no medical advice.
- The results are encouraging but not yet solid enough to constitute a standalone screening tool.The authors describe the work as an ongoing effort and plan further data collection and controlled studies with COVID-19 test ground truth.
- The analysis used only a subset of the collected data because COVID-19-positive users were uncommon.
- The dataset lacks health-status ground truth, so likely healthy users were selected from countries where COVID-19 was not prevalent.This selection further limited the dataset.
- Although repeated sampling could support studying respiratory-sound progression, that analysis was not conducted in the current work.
- The current app collects data but does not provide medical advice.The authors note challenges involved in giving medical advice to users.
A.1 The impact of modalities and dimensionality
Different sound modalities contribute differently across classification tasks, and combining cough with breathing can improve Task 1 performance. The analysis also notes greater overfitting in Tasks 2 and 3 because of their sample sizes.
- Task 1: For Task 1, combining cough and breathing achieved the highest AUC and lowest standard deviation, while cough alone reached around 70% AUC.Feature dimensionality was not highly significant, although the combination appeared to improve with more features.
- Task 2: For Task 2, cough was more precise than breathing or their combination among users who actually had a cough.The combined feature set performed better at lower dimensionality, whereas cough improved with more features.
- Task 3: Task 3 followed similar modality trends, with overfitting becoming more pronounced because of sample sizes.
A.2 Data augmentation
The study augmented the negative class for Tasks 2 and 3 with standard audio transformations, keeping augmented samples out of testing. Performance improved across reported metrics, including nearly 90% AUC and higher recall.
- Negative-class samples for Tasks 2 and 3 were augmented by amplifying signals, adding white noise, and changing pitch and speed.Augmented samples were used only for training, while the test set remained intact.
- Nearly 90% AUC and almost a 10% increase in AUC and recall were observed after augmentation, with slightly lower standard deviation.These values were reported relative to the results in Table 1.
- The authors report that higher recall enabled the model to recognize almost all COVID coughs, while some healthy-user misclassifications could be handled by second-stage testing.